Power-System State Estimation based on PMUs:

Static and Dynamic Approaches - from Theory to Real

Implementation

THÈSE NO 7665 (2017)

PRÉSENTÉE LE 9 JUIN 2017
À LA FACULTÉ DES SCIENCES ET TECHNIQUES DE L'INGÉNIEUR
GROUPE SCI STI RC
PROGRAMME DOCTORAL EN ENERGIE

ÉCOLE POLYTECHNIQUE FÉDÉRALE DE LAUSANNE

POUR L'OBTENTION DU GRADE DE DOCTEUR ÈS SCIENCES

PAR

Lorenzo ZANNI

acceptée sur proposition du jury:

Dr J. Van Herle, président du jury

Dr S.-R. Cherkaoui, Prof. M. Paolone, directeurs de thèse
Dr W. Sattinger, rapporteur
Prof. A. Abur, rapporteur
Prof. J.-Y. Le Boudec, rapporteur

Suisse
2017
 Perfection is not attainable, but if
we relentlessly chase perfection,
we will catch excellence.

To my parents Paola and Andrea,

my grandparents Novella, Mirella, Franco and Nerio,
and my uncle Luca.
Acknowledgements
First of all, I would like to sincerely thank my two supervisors, Rachid and Mario, for the
guidance and support during the four years of PhD. Additionally, I express my gratitude
to Mario for his trust since we met for the first time. I truly thank Prof. Jean-Yves Le
Boudec that substantially contributed to my scientific works. A special thank goes
to Prof. Ali Abur for the opportunity to do an internship in his lab at Northeastern
University, in Boston.
Then, I want to thank all the people I worked and collaborated with: the amazing
colleagues, the secretaries as well as the system administrators, the technicians and,
last but not least, our industrial partners (National Instruments, Romande Energie,
Services industriels de Lausanne and Alliander).
A huge thank goes to my friends both in Lausanne and in Italy for the fun we had
together during aperitifs, dinners, nights and holidays. In Lausanne, I particularly
thank Paolo, Marco, Asja, Nadia, Chiara, Francesco and Alessio. In Bologna, special
thanks go to my cousins (Fede, Je, Nik), Depe&Co, Tox, Mauri, Riki&Co, Franzo&Co.
Football has been an important part of my life, so I thank all my football mates. A
unique thank goes to Mélanie for the special moments we spent together during the
last year.
Finally, I am extremely grateful to my family for all the love and support that were
fundamental in this successful journey.

Lorenzo

i
Abstract
An increasing number of phasor measurement units (PMUs) are being deployed in
power systems in order to enhance the situational awareness and, in the near future,
we expect that many networks will be extensively equipped with PMUs. These devices
provide accurate and synchronized voltage/current phasors (called synchrophasors) at
a reporting rate up to 60 measurements-per-second, which is a significantly different
type of information with respect to the commonly-used voltage/current magnitude
and power measurements of remote terminal units (RTUs). The large availability of
synchrophasor measurements might improve existing power-system functions or even
disrupt the status quo of several automation processes.
Power-system state estimation (SE) is a functionality that might largely benefit from
the use of synchrophasor measurements (the power-system state consists of the set of
nodal phase-to-ground voltage phasors at all network buses). Best current practice
consists in estimating the network state every few tens of seconds (or even minutes) by
using asynchronous measurements of RTUs. A measurement infrastructure exclusively
composed of PMUs allows SE to become a linear and not iterative process that uses a
set of phase-aligned synchrophasor measurements and is characterized by a refresh
rate of tens of estimates-per-second and sub-second time-latency. This is what we call
real-time SE.
PMUs are commonly associated with transmission systems, but are gaining consider-
ation also in the context of distribution networks in order to implement fast control
schemes due to the presence of highly-volatile distributed generation and for fault
location purposes. Therefore, SE may become a future functionality of distribution
management systems (DMSs). The estimated state can be exploited by several DMS
functions, such as voltage control, congestion management and even fault manage-
ment.
Even if SE is a well-established power-system function, it still deserves research in
view of the proliferation of PMUs. Indeed, a relevant change of the measurement
infrastructure leads to modifications of the SE algorithms. Moreover, improvements in
terms of accuracy, computational time and time-latency are required in order to make
SE suitable for a wide range of applications, from control to fault management.
In this dissertation, we first describe in detail the advantages of using exclusively
synchrophasor measurements for the most common SE algorithms, i.e., weighted least
squares (WLS), least absolute value (LAV) and Kalman filter (KF). Then, we propose

iii
Abstract

new methodologies that improve the accuracy and/or computational performance of

WLS and KF. The results of the three above-mentioned SE algorithms are compared via
numerical simulations and also by using measurements provided by PMUs installed in
two real power-systems. Finally, we illustrate a fault location method that is based on
the SE theory.
We developed two methods for the on-line estimation of the process-model uncer-
tainties used by the KF, because power-system operating conditions are continuously
varying. Our goal is to improve the estimation accuracy by effectively filtering the
measurement noise. We designed a heuristic method for quasi-static conditions and a
rigorous method that is also able to deal with step changes of the system state. The
former is fast and suitable for real-time implementation, whereas the latter has a sig-
nificantly higher computational burden. A performance assessment involving KF and
WLS is conducted for both transmission and distribution networks.
Zero-injections represent equality constraints in the SE problem. The information
associated with zero-injection buses is precious and can be exploited in a particularly
efficient manner for the specific case of linear WLS. We propose a method based on
LQ-decomposition that strictly satisfies the equality constraints while reducing the
state-vector dimension by the number of constraints. Therefore, the computational
time is significantly reduced for networks that include a large number of zero-injection
buses. Moreover, the proposed method preserves the structure of the linear WLS
equations and makes the problem less ill-conditioned.
An important contribution of this dissertation consists in the validation of the the-
oretical findings via real-scale experiments. We deployed PMUs at every bus in a
sub-transmission network and a distribution feeder, both located in Switzerland. First,
we demonstrate the practical feasibility of running SE at high refresh-rate (50 estimates-
per-second) and low time-latency (below 70 ms). Second, for the two case studies we
compare and discuss the results of WLS, LAV and KF by using real synchrophasor
measurements.
Applications related to fault management have stringent accuracy and timing require-
ments. We intend to prove that PMU-based real-time SE exhibits unique accuracy,
refresh rate and time-latency, which satisfy the requirements of fault location and,
potentially, protective relaying. We propose a fault detection and faulted-line identifi-
cation method based on WLS-SE. Although it requires a massive deployment of PMUs,
this method works for any network and fault type as well as in presence of large amount
of distributed generation. Provided that the PMU cost will drop in the coming years,
the proposed approach is particularly interesting for distribution networks where fault
management can be largely improved.

Keywords: smart grid, power system state estimation, synchrophasor, phasor mea-
surement unit (PMU), linear state estimation, dynamic estimation, weighted least
squares, least absolute value, adaptive Kalman filter, measurement noise covariance
matrix, process noise covariance matrix, covariance estimation, step processes, pos-

iv
 Abstract

itive definite, zero injection, equality constraints, normal equation, ill-conditioned

problem, phasor data concentrator, fault location, protective relaying.

v
Résumé
Un nombre croissant de “phasor measurement units (PMUs)” vont être déployés dans
les réseaux électriques, afin d’améliorer le monitoring et, dans un futur proche, nous
pensons que plusieurs réseaux seront en grand partie équipés avec des PMUs. Ces
dispositifs fournissent des mesures de phaseurs de tension et de courant, appelés “syn-
chrophaseurs”, qui sont précises, synchronisées, très fréquentes (jusqu’à 60 mesures
par seconde). Ce type de mesure est considérablement différent des mesures fournies
par “remote terminal units (RTUs)” consistant en des amplitudes de tension et de
courant ainsi que des puissances. Une vaste pénétration de PMUs peut améliorer les
fonctionnalités existantes ou même changer radicalement le principe de nombreux
systèmes d’automation.
L’estimation d’état des réseaux électriques peut bénéficier de l’utilisation des synchro-
phaseurs (l’état d’un réseau électrique est composé par les phaseurs de tension entre
phase et terre dans tous les nœuds). Actuellement, l’état du réseau est estimé chaque
dizaine de secondes (ou même chaque minute) en utilisant les mesures asynchrones
des RTUs. Au contraire, si nous utilisons une infrastructure de mesure composée ex-
clusivement par des PMUs, l’estimation d’état devient un algorithme qui possède
les caractéristiques suivantes : linéaire et non itératif ; taux de rafraîchissement de
quelques dizaines d’estimations par seconde ; temps de latence inférieur à une se-
conde. Nous appelons celui-ci un estimateur d’état en temps réel.
Les PMUs sont souvent associées aux réseaux de transmission, mais il y a un intérêt
croissant de les installer aussi dans les réseaux de distribution, afin de contrôler les
instabilités causées par la génération distribuée et de localiser les défauts. C’est pour-
quoi l’estimation d’état pourrait devenir une future fonctionnalité des “distribution
management systems (DMSs)” modernes. L’état estimé pourra être utilisé par diffé-
rentes applications, comme le contrôle de la tension, le contrôle des congestions des
lignes, et aussi pour la gestion des défauts.
Bien que l’estimation d’état soit une fonctionnalité consolidée dans le domaine des
réseaux électriques, l’avènement des PMUs introduit des nouveaux sujets de recherche.
Des modifications doivent être apportées aux algorithmes d’estimation d’état et une
considérable amélioration de précision et prestations computationnelles est néces-
saire, afin d’élargir la gamme des applications qui peuvent utiliser l’état estimé (du
contrôle à la gestion des défauts).
Dans cette thèse, premièrement nous décrivons les avantages à utiliser les synchropha-

vii
Résumé

seurs par les algorithmes d’estimation d’état les plus communs, comme la méthode
des moindres carrés pondérés (MMCP), la méthode des moindres valeurs absolues
(MMVA) et le filtre de Kalman (FK). Ensuite, nous proposons des nouvelles méthodes,
afin d’améliorer la précision et la prestation computationnelle du MMCP et du FK.
Nous comparons les résultats des trois estimateurs mentionnés ci-dessus, soit avec
des simulations numériques, soit en utilisant des mesures prises par des PMUs ins-
tallées dans deux réseaux électriques réels. Enfin, nous présentons une méthode de
localisation de défauts qui est basée sur la théorie de l’estimation d’état.
Nous avons développé deux méthodes pour l’estimation des incertitudes du modèle
de processus utilisé par le FK ; ils permettent de mettre à jour fréquemment les pa-
ramètres du modèle de processus, afin de suivre les conditions opérationnelles du
réseau qui changent en continu. L’objectif est d’atteindre une meilleure précision de
l’état estimé, grâce à un filtrage efficace du bruit de mesure. Une méthode heuristique
a été développée pour fonctionner en régime permanent ou légèrement variable et
une méthode rigoureuse a été développée pour suivre l’état du réseau aussi pendant
une variation soudaine. Le premier est plus rapide et prêt à une implémentation en
temps réel, tandis que le temps computationnel exigé par la deuxième méthode est
beaucoup plus élevé. Nous avons effectué une évaluation des performances du FK et
du MMCP avec des réseaux de transmission et de distribution.
Les injections-nulles ajoutent des contraintes d’égalité dans le problème d’estima-
tion d’état. Cette information est précieuse et peut être exploitée par le MMCP li-
néaire. Nous proposons une méthode basée sur la décomposition-LQ qui satisfait
les contraintes d’égalité et, en même temps, réduit la dimension du vecteur d’état
par le nombre des contraintes. C’est pourquoi, cette méthode baisse considérable-
ment le temps computationnel dans le cas d’un réseau qui a beaucoup de nœuds
à injection-nulles. De plus, la structure des équations du MMCP est préservée et le
conditionnement du problème est amélioré.
Une contribution très importante de cette thèse consiste à valider les résultats théo-
riques avec des expériences réelles. Nous avons déployé des PMUs dans chaque nœud
de deux réseaux en Suisse : un réseau de sub-transmission et un feeder de distribution.
Premièrement, nous visons à démontrer la faisabilité d’implémenter un estimateur
d’état qui donne une estimation chaque 20 millisecondes et qui a une latence en des-
sous de 70 millisecondes. Deuxièmement, nous comparons et discutons les résultats
de MMCP, MMVA et FK qui utilisent les mesures réelles des PMUs.
Les besoins opérationnels des applications pour la gestion des défauts sont très stricts
en termes de précision et de timing. Nous voulons démontrer que l’estimation d’état
basée sur les PMUs possède les caractéristiques nécessaires pour être utilisée pour
la localisation des défauts et, potentiellement, comme relais de protection. Nous
avons développé une méthode capable de détecter un défaut et d’identifier la ligne
correspondante. Cette méthode est basée sur la théorie de l’estimation d’état avec
le MMCP. Malgré qu’elle nécessite un déploiement massif des PMUs, cette méthode
fonctionne pour tous les types de réseau et de défaut, même en présence d’une grande

viii
 Résumé

quantité de génération distribuée. A condition que le coût des PMUs baisse dans les
prochaines années, cette approche est particulièrement intéressante pour les réseaux
de distribution dans lesquels il y a encore une marge d’amélioration significative dans
la gestion des défauts.

Mots clefs : réseau électrique intelligent, estimation d’état linéaire, synchrophaseur,

phasor measurement unit (PMU), méthode des moindres carrés pondérés, méthode
des moindres valeurs absolues, filtre de Kalman, matrice de covariance du bruit de
mesure, matrice de covariance du bruit du processus, estimation des matrices de
covariance, injections-nulles, contraintes d’égalité, localisation de défauts, relais de
protection.

ix
Contents
Acknowledgements i

Abstract (English/Français) iii

List of figures xv

List of tables xix

List of Symbols xxi

List of Acronyms xxiii

Nomenclature xxv

Introduction 1

1 Theory on PMU-based power-system state estimation 9

1.1 Linear Measurement Model . . . . . . . . . . . . . . . . . . . . . . . . . 9
1.1.1 The measurement matrix H . . . . . . . . . . . . . . . . . . . . . 11
1.1.2 The measurement noise covariance matrix R . . . . . . . . . . . 12
1.2 Linear weighted least squares . . . . . . . . . . . . . . . . . . . . . . . . 16
1.3 Least absolute value . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
1.4 Dynamic linear estimation via discrete Kalman filter . . . . . . . . . . . 20
1.4.1 General process-model and discrete Kalman filter . . . . . . . . 20
1.4.2 Historical notes on Kalman filter in power-system state estimation 22
1.4.3 PMU-based state estimation via discrete Kalman filter . . . . . . 25

2 Adaptive Kalman filtering for the on-line estimation of the process-model

uncertainties 27
2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
2.2 Literature review on the estimation of the Kalman filter covariances . . 29
2.3 Estimation accuracy: Kalman filter vs. weighted least squares . . . . . . 30
2.4 New heuristic method for the on-line assessment of Q . . . . . . . . . . 32
2.4.1 Analytical formulation . . . . . . . . . . . . . . . . . . . . . . . . 32
2.4.2 Simulations and results . . . . . . . . . . . . . . . . . . . . . . . . 33

xi
Contents

2.5 New prediction-error covariance estimation method (PECE) for step-

varying processes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
2.5.1 Analytical formulation . . . . . . . . . . . . . . . . . . . . . . . . 49
2.5.2 Simulations and results for ideal processes . . . . . . . . . . . . 51
2.5.3 Simulations and results for power-system state estimation . . . 60
2.6 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67

3 Linear state estimation with zero-injection equality constraints via LQ-decomposition 71

3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
3.2 Linear State Estimation with Linear Equality Constraints . . . . . . . . 73
3.2.1 Normal equation with virtual measurements (VM) . . . . . . . . 75
3.2.2 Normal equation with constraints (NE/C) . . . . . . . . . . . . . 75
3.2.3 Normal equation derived from LQ-decomposition (LQD) . . . . 75
3.2.4 QR-decomposition (QRD) . . . . . . . . . . . . . . . . . . . . . . 78
3.3 Simulations and results . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
3.3.1 Test conditions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
3.3.2 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
3.3.3 Impact of virtual-measurement variance on numerical stability 82
3.4 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82

4 Results of PMU-based state estimation of real networks 85

4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
4.2 Distribution feeder of the EPFL-campus network . . . . . . . . . . . . . 86
4.2.1 Network characteristics and measurement infrastructure . . . . 86
4.2.2 State estimation results . . . . . . . . . . . . . . . . . . . . . . . . 90
4.3 Sub-transmission network of the city of Lausanne . . . . . . . . . . . . 101
4.3.1 Network characteristics and measurement infrastructure . . . . 101
4.3.2 State estimation results . . . . . . . . . . . . . . . . . . . . . . . . 105
4.4 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112

5 A new method based on real-time state estimation for fault location and pro-
tective relaying 115
5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115
5.2 The proposed method for fault detection and faulted-line identification 117
5.3 Simulation set-up . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120
5.4 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124
5.4.1 Fault detection . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125
5.4.2 Faulted line identification . . . . . . . . . . . . . . . . . . . . . . 129
5.4.3 Faulted bus identification . . . . . . . . . . . . . . . . . . . . . . 129
5.4.4 Influence of operating conditions and distributed generation . 129
5.4.5 Computational time and latency . . . . . . . . . . . . . . . . . . 130
5.5 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 132

xii
 Contents

Conclusions 133

A Appendix 137
A.1 Derivation of the measurement matrix H . . . . . . . . . . . . . . . . . 137
A.2 Proof of Theorem 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141
A.3 Proof of Theorem 3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 142
A.4 Covariance estimation method of Myers and Tapley . . . . . . . . . . . 144

Bibliography 155

Curriculum Vitae

xiii
List of Figures
1 Typical structure of an EMS. . . . . . . . . . . . . . . . . . . . . . . . . . 2

1.1 Convention employed for current injections and flows with positive sign. 11

2.1 Network topology of the New England 39-bus system together with the
adopted measurement configuration. . . . . . . . . . . . . . . . . . . . 35
2.2 Time evolution of the active and reactive powers absorbed by the load
at bus #4. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
2.3 RMSE of the DKF as a function of the parameter N . . . . . . . . . . . . . 37
2.4 Normalized residuals of the estimates of the real part of the voltage at
bus #37. Left plot: time evolution of the residuals. Right plot: sample
ACFs of the residuals. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
2.5 RMSEs of the DKF and LWLS at each bus. . . . . . . . . . . . . . . . . . 38
2.6 Time evolution of the RMSEs of the DKF and LWLS. . . . . . . . . . . . 39
2.7 Time evolution of the voltage magnitude V and phase-angle δ at bus #37.
The true value and the values estimated by DKF and LWLS are displayed. 39
2.8 RMSE of the DKFQ sampling , DKFQ assessed and LWLS as a function of Qd . . 40
2.9 Numerical validation of Theorem 1. We show the LHS vs. RHS of (2.3).
The separate contribution of the two terms of the RHS is also shown. . 40
2.10 Network topology of the IEEE 123-bus test feeder together with the
adopted measurement configuration. This topology is obtained by set-
ting the switch positions specified by the benchmark and is composed
of 121 buses. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
2.11 Time evolution of the absorbed/injected powers of loads/generators. . 43
2.12 RMSE of the DKF as a function of the parameter N . . . . . . . . . . . . . 43
2.13 Normalized residuals of the estimates of the real part of the voltage at
bus #61. Left plot: time evolution of the residuals. Right plot: sample
ACFs of the residuals. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
2.14 RMSEs of the DKF and LWLS at each bus. The maximum RMSE among
the three phases is shown. . . . . . . . . . . . . . . . . . . . . . . . . . . 45
2.15 Time evolution of the RMSEs of the DKF and LWLS. . . . . . . . . . . . 45
2.16 Time evolution of the voltage magnitude V and phase-angle δ at bus #92
in phase a. The true value and the values estimated by DKF and LWLS
are displayed. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45

xv
List of Figures

2.17 Time evolution of the element of Q corresponding to the real part of the
voltage at bus #92 in phase a. . . . . . . . . . . . . . . . . . . . . . . . . 46
2.18 RMSE of the DKFQ sampling , DKFQ assessed and LWLS as a function of Qd . . 46
2.19 Numerical validation of Theorem 1 for buses #1–40. We show the LHS
vs. RHS of (2.3). The separate contribution of the two terms of the RHS
is also shown. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
2.20 Base case, q = 10−10 . Time evolution of the third diagonal element of
Pk|k−1 (the other diagonal elements exhibit an analogous behavior): true
and estimated values. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
2.21 Base case, q = 10−10 . Time evolution of the RMSEs. . . . . . . . . . . . . 55
2.22 Base case, q = 10−6 . Time evolution of the third diagonal element of
Pk|k−1 (the other diagonal elements exhibit an analogous behavior). . . 56
2.23 Base case plus steps, q = 10−10 (a step occurs at the 1 000th time-step).
Time evolution of the third diagonal element of Pk|k−1 (the other diago-
nal elements exhibit an analogous behavior). . . . . . . . . . . . . . . . 57
2.24 Base case plus steps, q = 10−10 (a step occurs at the 1 000th time-step).
Further zoom of Fig. 2.23 close to the step-variation that occurs at the
1 000th time-step. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
2.25 Base case plus steps, q = 10−10 (a step occurs at the 1 000th time-step).
Time evolution of the RMSEs. . . . . . . . . . . . . . . . . . . . . . . . . 57
2.26 Base case plus steps, q = 10−6 (a step occurs at the 1 000th time-step).
Time evolution of the third diagonal element of Pk|k−1 (the other diago-
nal elements exhibit an analogous behavior). . . . . . . . . . . . . . . . 58
2.27 Base case plus steps, q = 10−6 (a step occurs at the 1 000th time-step).
Time evolution of the RMSEs. . . . . . . . . . . . . . . . . . . . . . . . . 58
2.28 RMSEs as a function of N for the PECE and Myers methods. In the base
case, the RMSE is computed considering 2 000 time-steps. In the base
case plus steps, the RMSE is computed considering the 30 time-steps
after the state step-variation. . . . . . . . . . . . . . . . . . . . . . . . . 59
2.29 Base case, q = 10−10 , r = 10−7 , N = 100. Time evolution of the 2nd
diagonal element of Pk|k−1 and of the associated 2nd element of the state
vector: true state and state estimated by the DKF that uses the Myers
method. This figure shows the numerical stability problems of the Myers
method. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
2.30 Time profiles of the active and reactive power at bus #4. A drop of about
90 MVAR of reactive power occurs at 600 seconds. . . . . . . . . . . . . . 61
2.31 Time profile of the reactive power at bus #4: zoom close to the 90 MVAR
step-variation. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
2.32 Time evolution of the RMSEs. . . . . . . . . . . . . . . . . . . . . . . . . 64
2.33 Time evolution of the real part of the voltage at bus #4 that is the bus
where the reactive power step-variation occurs. . . . . . . . . . . . . . . 65

xvi
 List of Figures

2.34 Time evolution of the real part of the voltage at bus #31: true and esti-
mated values. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
2.35 RMSEs as a function of N for the PECE and Myers methods. For case (a)
where the network is in quasi-static conditions, the RMSE is computed
considering the 100 seconds prior to the step. For case (b) where the
network state has a step-variation, the RMSE is computed considering
the 2 seconds following the step. . . . . . . . . . . . . . . . . . . . . . . . 66
2.36 Computational time as a function of n. For the Myers and heuristic
methods, multiple curves refer to different values of N. . . . . . . . . . 68

3.1 Network topology of the New England 39-bus system together with the
adopted measurement configuration. . . . . . . . . . . . . . . . . . . . 79

4.1 Schematic of the monitored feeder of the EPFL-campus distribution

network. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
4.2 Limits of ratio error and phase displacement of the sensors (classes 0.1,
0.2, 0.5). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
4.3 Time evolution of voltages and powers at bus #1. . . . . . . . . . . . . . 93
4.4 Zoom of Fig. 4.3 from 30 s to 40 s in phase a. . . . . . . . . . . . . . . . . 94
4.5 Voltage magnitude and phase-angle at time-step 1. . . . . . . . . . . . . 94
4.6 Time evolution of the total grid losses and of the total reactive power
produced by the cable lines. . . . . . . . . . . . . . . . . . . . . . . . . . 95
4.7 Statistics of the distributions of the normalized measurement-residuals
of the LWLS. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96
4.8 Absolute value of the measurement residuals (not normalized) at time-
step 1. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
4.9 Load inrush at bus #3. Time evolution of active- and reactive-power
flows at bus #1. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98
4.10 Measurement residuals (not normalized) at time 3.94 seconds of the
time-window displayed in Fig. 4.9. This is the instant when a perturba-
tion in the power absorbed at bus #3 occurs. . . . . . . . . . . . . . . . . 98
4.11 Quasi-static conditions. Time evolution of the voltage magnitude and
phase-angle at bus #1 in phase a. . . . . . . . . . . . . . . . . . . . . . . 99
4.12 Step change. Time evolution of the voltage magnitude at bus #1 in the
three phases a, b, c. Note that we applied different time-scales in the
x-axis of the three graphs. . . . . . . . . . . . . . . . . . . . . . . . . . . 99
4.13 Cumulative distribution functions of time latencies. . . . . . . . . . . . 101
4.14 Schematic of the monitored portion of the sub-transmission network of
the city of Lausanne. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
4.15 Voltage magnitude and phase-angle at time-step 1. . . . . . . . . . . . . 106
4.16 Time evolution of the total grid losses and of the total reactive power
produced by the transmission lines. . . . . . . . . . . . . . . . . . . . . . 107

xvii
List of Figures

4.17 Statistics of the distributions of the normalized measurement-residuals

of the LWLS. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108
4.18 Statistics of the distributions of the residuals of the sending-end current-
flow measurements. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
4.19 Quasi-static conditions. Time evolution of the voltage magnitude and
phase-angle at bus #1 in phase a. . . . . . . . . . . . . . . . . . . . . . . 111
4.20 Step change. Time evolution of the voltage magnitude at bus #1 in phase
a. The left-graph shows the entire voltage dip; the right-graph shows
only the initial voltage drop. . . . . . . . . . . . . . . . . . . . . . . . . . 111
4.21 Step change. Time evolution of the element of Q used by the DKF-heur
corresponding to the real part of the voltage at bus #1 in phase a. . . . . 111

5.1 Flowchart of the proposed fault detection and location method. . . . . 119
5.2 Schematic of the 18-bus distribution feeder. . . . . . . . . . . . . . . . . 121
5.3 Comparison between the fault-current magnitude estimated by the
simulated PMU versus the idealized fault-current magnitude. . . . . . 123
5.4 Time evolution of the objective functions of the m = 17 state estimators
when a 3-phase 100Ω fault occurs at 1/4 of line L13,16 . The fault inception
is between 0.5 and 0.52 seconds. . . . . . . . . . . . . . . . . . . . . . . . 126
5.5 Time evolution of the objective functions of the m = 17 state estimators
when a 3-phase 100Ω fault occurs at 1/2 of line L13,16 . The fault inception
is between 0.5 and 0.52 seconds. . . . . . . . . . . . . . . . . . . . . . . . 126
5.6 Overall time-latency of the proposed method in identifying faults. . . . 131

A.1 Two-port π-model of a generic three-phase network branch. Note that

the model parameters are 3 × 3 complex matrices. . . . . . . . . . . . . 140

xviii
List of Tables
2.1 RMSEs of the DKF and LWLS for different measurement configurations. 40
2.2 Average computational time in milliseconds. . . . . . . . . . . . . . . . 45

3.1 Comparison of the considered methods. . . . . . . . . . . . . . . . . . . 81

3.2 Impact of the variance assigned to virtual measurements on numerical
stability. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82

4.1 Power-injection measurements at time-step 1. . . . . . . . . . . . . . . 92

4.2 Average computational time in milliseconds. . . . . . . . . . . . . . . . 100
4.3 Line parameters: length L in km, resistance R in Ω/km, reactance X in
Ω/km, and susceptance B in μS/km. The subscripts 0 and 1 stand for
zero and positive sequence, respectively. . . . . . . . . . . . . . . . . . . 103
4.4 Accuracy classes of PTs and CTs. The subscripts send and rec refer to the
sending- or receiving-end of the line, respectively. . . . . . . . . . . . . 104
4.5 Power-injection measurements at time-step 1. . . . . . . . . . . . . . . 104
4.6 Power-flow measurements at the sending-ends of the lines at time-step 1.105
4.7 Power-flow measurements at the receiving-ends of the lines at time-step 1.105
4.8 Average computational time in milliseconds. . . . . . . . . . . . . . . . 112

5.1 Line parameters: length L in km, resistance R in Ω/km, reactance X in

Ω/km, and susceptance B in μS/km. The subscripts 0 and 1 stand for
zero and positive sequence, respectively. . . . . . . . . . . . . . . . . . . 121
5.2 Rated power of the transformers at the secondary substations. . . . . . 122
5.3 Limits of ratio error and phase displacement for PTs and CTs. . . . . . . 124
5.4 Accuracy for 3-phase 1 Ω fault (neutral grounded) . . . . . . . . . . . . 127
5.5 Accuracy for 3-phase 100 Ω fault (neutral grounded) . . . . . . . . . . . 127
5.6 Accuracy for 2-phase 1 Ω fault (neutral grounded) . . . . . . . . . . . . 127
5.7 Accuracy for 2-phase 100 Ω fault (neutral grounded) . . . . . . . . . . . 127
5.8 Accuracy for 2-phase 1 Ω fault (neutral isolated) . . . . . . . . . . . . . . 127
5.9 Accuracy for 2-phase 100 Ω fault (neutral isolated) . . . . . . . . . . . . 127
5.10 Accuracy for 1-phase 1 Ω fault (neutral grounded) . . . . . . . . . . . . 128
5.11 Accuracy for 1-phase 100 Ω fault (neutral grounded) . . . . . . . . . . . 128
5.12 Accuracy for 1-phase 1 Ω fault (neutral isolated) . . . . . . . . . . . . . . 128

xix
List of Tables

5.13 Accuracy for 1-phase 100 Ω fault (neutral isolated) . . . . . . . . . . . . 128

xx
List of Symbols
A state-transition matrix
a, b, c phase indexes in a three-phase power system
d number of network branches
e measurement error vector
f power-system frequency
G gain matrix of the LWLS
H measurement matrix
I identity matrix
J objective function of the LWLS
k time-step index
K hat matrix of the LWLS
L Kalman gain
m number of measurements
n number of state variables
N number of previous time-steps used for the on-line estimation of
the DKF process model uncertainties
p number of equality constraints due to zero-injection buses
Pk|k−1 prediction-error covariance matrix of the DKF
Pk|k estimation-error covariance matrix of the DKF
Q process noise covariance matrix
R measurement noise covariance matrix
r measurement residual vector
r N normalized measurement residual vector
rNx normalized state-estimate residual vector
s number of network buses
S residual sensitivity matrix of the LWLS
T number of simulation time-steps
T covariance matrix of the innovations
v measurement noise vector
V voltage magnitude
x state vector
w number of zero-injection buses
w process noise vector

xxi
List of Symbols

y innovation vector
z measurement vector
δ voltage phase-angle
Ω covariance matrix of the LWLS measurement residuals

xxii
List of Acronyms
ACF auto correlation function
AKF adaptive Kalman filter
CT current transformer
DESL distributed electric system laboratory
DG distributed generation
DKF discrete Kalman filter
DMS distribution management systems
EKF extended Kalman filter
EMS energy management system
EPFL École Polytechnique Fédérale de Lausanne, which is the french name of the Swiss F
FPGA field-programmable gate array
GPS global positioning satellite
GUM Guide to the expression of uncertainty in measurement
KF Kalman filter
LAV least absolute value
LHS left-hand side
LNR largest normalized residual
LQD LQ-decomposition
LWLS linear weighted least-squares
MEP-ZIB maximum estimated power at the zero-injection buses
NE normal equation
PDC phasor data concentrator
PECE prediction-error covariance estimation
PMU phasor measurement unit
PT potential transformer
pu per unit
PV photovoltaic
QRD QR-decomposition
RHS right-hand side
RMSE root mean square error
RTS real-time simulator
RTSE real-time state estimation
RTU remote terminal unit

xxiii
List of Acronyms

SCADA supervisory control and data acquisition system

SE state estimation
TVE total vector error
WLS weighted least-squares

xxiv
Nomenclature
Static state: set of nodal voltage phasors at all network buses.
Dynamic state: dynamic state variables of the generator machines, such as phase-
angle and speed.
Measurement model: mathematical model that expresses the relationship between
the measurements and the state variables at a time-step k.
Process model: mathematical model of the time evolution of the system state at a
time-step k as a function of the system state at time-step k − 1 and of the controllable
inputs at a time-step k − 1.
Static estimator: state estimator that computes the system state by using only the
incoming set of measurements at a time-step k.
Dynamic estimator: state estimator that, at each time-step k, computes the probabil-
ity density function of the system state in two stages: first, it predicts the state from
time-step k − 1 to time-step k by exploiting a process model; then, it compute the
estimated state by using the incoming measurements at time-step k. Thus, unlike
static estimators, a dynamic estimator uses also the information contained in the
previous state-estimate.
Tracking estimator: static estimator that uses the incoming measurements at a time-
step k to update the previous estimate at time-step k − 1, instead of fully executing the
iterative estimation from a flat-start initialization at every time-step.
Robust estimator: state estimator with high breakdown-point. The breakdown-point
is the smallest number of bad data that can lead state estimation to an incorrect
solution. Note that the breakdown point is highly influenced by the measurement
redundancy; indeed, a poor redundancy limits the number of bad data that the estima-
tor can reject.
Real-time state estimator: state estimator characterized by high refresh-rate and low
latency.
Hybrid state estimator: state estimator that uses both SCADA measurements pro-
vided by RTUs and synchrophasor measurements provided by PMUs.
Conventional or SCADA measurements: measurements provided by RTUs, which

xxv
Nomenclature

typically consist of voltage/current magnitudes and power flows/injections.

Synchrophasor: accurate and time-synchronized phasor measured by PMUs.
Critical measurement: measurement whose loss makes the network unobservable.
Leverage measurement: a measurement characterized by a structure of the measure-
ment equation that is different from that of the other measurements. Specifically,
the row of the measurement matrix H corresponding to a leverage measurement has
very different (smaller or larger) entries compared to the other rows. This creates a
kind of insensitivity of the state estimator with respect to this measurement, i.e., the
measurement residual (defined as the difference between the raw and the estimated
measurement) corresponding to a leverage measurement remains small even in pres-
ence of a large error.
Pseudo-measurement: measurements with large uncertainties that represent predic-
tions typically based on historical data or forecasts.
Virtual measurement: measurements with very small uncertainties that are used to
approximate the exact information associated with zero-injections.

xxvi
Introduction
Context and Motivation
In general, the state of a power system refers to the minimum set of independent vari-
ables from which all the power-system quantities (voltages, current/power flows/in-
jections, etc.) can be computed. This is why the state fully defines the power-system
operating conditions. In this dissertation, the state always refers to the so-called static
state that is typically defined as the set of nodal phase-to-ground voltage phasors at all
network buses. We do not treat the estimation of the dynamic states that consist in the
state variables of the generator machines, such as rotor speed and angle.

The knowledge of the state is fundamental in the operation of the transmission system
as it is required by several functions, such as security control and economic dispatch.
Therefore, the procedure that infers the power-system state has to be accurate and
reliable.

Until the 1970s, the power-system state was computed by means of a load-flow cal-
culation that used the raw measurements of voltages and powers [1]. This procedure
had several weaknesses: for instance, it was vulnerable to missing measurements and
the presence of measurement errors affected dramatically the load-flow solution. To
overcome these limitations, Schweppe proposed to combine load-flow and statistical
estimation theory and introduced state estimation (SE) in power systems [2, 3, 4]. SE
consists in solving an optimization problem that processes the raw measurements
and a network model1 with the aim of determining the most likely estimate of the
power-system state, which is called estimated state [5, 6]. Intuitively, SE evaluates how
well the measurements fit the constraints given by the network model. Provided that
the network is observable2 and there is sufficient measurement redundancy 3 , the state

1
The network model consists of the combination of network topology and electrical parameters of the
network components.
2
A network is observable if all the state variables can be uniquely computed for a given set of measure-
ments and network topology [7].
3
Measurement redundancy is defined as the ratio between the number of measurements and the
number of state variables; a sufficiently high redundancy is fundamental to guarantee the benefits of SE,
such as the ability to identify measurement and network-model errors.

1
Introduction

Data Supervisory
Acquisition Control

SCADA

Energy Automatic State Security control

management generation Estimation
control
Network State
Resource Power / Freq. Contingency Optimal
Scheduling Control Analysis Load Flow

Energy Unit Economic Transmission

Transactions Commitment Dispatch Loss Factors

Load Short-circuit Protections

Forecast Analysis Tuning

Figure 1 – Typical structure of an EMS.

estimator is able to filter the measurement noise, tolerate measurement losses, identify
bad data 4 in the measurement set [8, 9, 10] and determine network-model errors
[11, 12, 13]. The foregoing properties of SE allow the system operator to obtain an
estimated state that is more reliable than the raw measurements as well as to compute
quantities that are not directly measured by a meter.

In the last decades, SE has become a core situation-awareness component for power-
system operators [14, 15, 16]. As depicted in Fig. 1, the SE solution constitutes the
essential input of several functions of modern energy management system (EMS). An
EMS is located at every power-system control center and helps the grid operator to
optimally and securely manage the electrical assets. The EMS includes a supervisory
control and data acquisition system (SCADA) that collects telemetered measurements
from various devices, such as remote terminal units (RTUs), and provides the measure-
ments to the state estimator.

Moreover, the increasing penetration of distributed energy resources at medium and

low voltage levels is calling for monitoring and control also in distribution systems.
Therefore, SE may become a standard functionality of distribution management sys-
tems (DMSs). The estimated state can be exploited by several DMS functions, such
as voltage control, congestion management and even fault management as we will
explain later in this dissertation.

Traditionally, SE is supplied with measurements provided by RTUs that consist of

voltage/current magnitudes and power flows/injections, which are commonly called
4
A bad datum or bad measurement is defined as a measurement that contains a gross error. This
can be due to several causes, such as malfunctioning telecommunication, meter failure, erroneous
wire-connection or software bug.

2
 Introduction

conventional measurements or SCADA measurements. The availability of this type

of measurements has lead to the development of state estimators with the following
characteristics:

1. The estimation algorithm is non-linear and iterative due to the non-linearity

of the function that links power and current-magnitude measurements to the
voltage state-variables;

2. The refresh-rate is in the order of tens of seconds or minutes due to the low
acquisition-rate of RTUs and the high-computation time of the SE algorithm;

3. RTU measurements are usually not time-stamped or are time-stamped with poor
accuracy (not sub-second); in the former case, they are time-stamped in the
SCADA by inferring the delay added by both RTU processing and telecommuni-
cation network. Therefore, the time-skew between measurements can introduce
significant and unknown errors in the estimated state;

4. One of the buses has to be chosen as the the reference bus for the voltage phase-
angle that is fixed to an arbitrarily selected value (usually equal to zero) at this
bus.

Concerning the SE methods, the most widely-used algorithm in power-system SE is the

weighted least-squares (WLS).The success of the WLS has been driven by its statistical
properties and low computational complexity that have eased the development of
efficient techniques (e.g., the fast-decoupled version [17]) and have enabled the use of
sparse matrix tools.

Nowadays, the increasing deployment of phasor measurement units (PMUs) enables

the incorporation of phasor measurements in power-system SE. PMUs provide ac-
curate and time-stamped phasors, called synchrophasors, at a typical reporting rate
of 30, 50 or 60 measurements per second [18]. Note that PMUs measure magnitude
and phase-angle of voltage/current phasors and, thus, they are capable of directly
measure the power-system state. The synchronization is usually provided by the global
positioning satellite (GPS) system, but can be also disseminated with dedicated com-
munication protocols characterized by jitter below 100 ns. The PMU technology is
experiencing a fast evolution triggered by the hardware-cost reduction and by the
increasing requirements of several power-system applications [19].

The voltage and current synchrophasors provided by PMUs can be included in the
measurement set to enhance the SE performance; a state estimator that uses both
conventional and PMU measurements is typically called hybrid state estimator [20, 21,
22, 23, 24, 25]. However, the improvement is still limited if SCADA measurements are
necessary to achieve the network observability or to ensure a sufficient measurement
redundancy.

3
Introduction

As PMUs start populating the network, in some portions of the grid the observability
and redundancy requirements are matched only with PMUs [26, 27, 28]. The fact that
the state of these portions of the grid can be estimated by using exclusively synchropha-
sor measurements leads to a remarkable improvement of the SE performance [29, 30].
The main benefits for SE relying only on this type of measurements are listed here
below:

1. When only voltage and current phasors are used, the equations linking mea-
surements and state variables, called measurement model, are linear. Therefore,
we can employ linear SE that is characterized by lower computational time, in-
creased numerical stability and no approximations in the measurement model.
For instance, it is well known that the linear WLS (LWLS) is not iterative as the
estimated state is given by the solution of an exact quadratic problem;

2. Each synchrophasor measurement is time-stamped and its phase-angle is aligned.

Hence, the measurements can be phase-aligned at the data collection point, even
if they are received at different time-instants. This ensures a superior accuracy of
the SE solution as the set of measured phasors is coherent with respect to time;

3. The very high PMU reporting-rate leads to a SE process characterized by high

refresh-rate and low latency, which is typically called real-time SE (RTSE);

4. As PMUs directly measure the phasor phase-angle, we do not need to choose

a reference bus where the voltage phase-angle is fixed. All the phase-angle
measurements are processed at the same time by the state estimator [31].

Besides, a PMU measures the phasors in each of the three phases, so that we can
use three-phase SE. The advantage consists in the possibility to account for network-
parameter asymmetries and for imbalances in the network operating-conditions.

Thanks to the linearity of the measurement model and the high refresh-rate, two
alternative SE techniques can be reconsidered for power-system SE: the least absolute
value (LAV) [32] and the Kalman filter (KF) [33]. WLS and LAV belong to the category
of static estimators, i.e., estimators that exploit only the information contained in the
incoming set of measurements. The KF is a dynamic estimator because, in addition to
the incoming measurements, it also makes use of the so-called process model that is a
mathematical representation of the time-evolution of the power-system state. To avoid
confusion throughout the dissertation, we clarify that static and dynamic estimators
will be used to estimate only the static state of the power-system (voltage phasors at all
network buses) and not the dynamic states (state variables of the generators).

The LAV estimator applied to linear SE with PMUs has been thoroughly investigated in
[34, 35, 36]. The results prove that the LAV is competitive with the LWLS in terms of
accuracy and computational time.

4
 Introduction

Regarding the KF, most of the contributions have investigated various versions of non-
linear KF, which are described in Section 1.4.2. The classic linear version of the KF is
typically called discrete KF (DKF). However, from the best of our knowledge, the topic
of DKF based on PMUs is tackled only in [37].

In this dissertation, we present and demonstrate the advantages of using a measure-

ment infrastructure exclusively composed of PMUs for linear power-system SE. Indeed,
in the near future we expect to have an increasing number of networks that are ob-
servable with synchrophasor measurements. Particular attention is dedicated to the
DKF that has been proposed in many different versions in the literature. We focus on
the validation of the persistent process-model and on the estimation of its uncertainty
(i.e., the error covariances); we also provide an extensive performance assessment
of the DKF with respect to the LWLS. Another important and original contribution
of this dissertation consists in the presentation of the SE results obtained in two real
power-systems equipped with PMUs, which allow us to validate the theoretical findings
in real-scale experiments. Finally, we prove that the outstanding characteristics of
PMU-based SE in terms of accuracy, refresh rate and latency can fulfill the stringent
requirements dictated by fault-location and even protections schemes. This peculiarity
might set the foundations for a new category of protection systems.

Dissertation outline
This dissertation is organized as follows.

In Chapter 1, we introduce the linear SE problem that uses only synchrophasor mea-
surements provided by PMUs. First, we describe the typical features of the linear
measurement model that links phasor measurements and state variables. In particular,
we illustrate the procedure to derive the measurement matrix and the measurement-
noise covariance matrix. Then, we illustrate the formulation and characteristics of
three algorithms that are commonly used for linear power-system SE: LWLS, LAV and
DKF.

Chapter 2 is entirely dedicated to the study of the DKF applied to power-system SE.
We propose two methods for the estimation of the time-varying uncertainties of the
process-model. The first is a heuristic method that continuously assesses the process-
noise variances in order to maintain an effective filtering action in changing network
operating-conditions, although it exhibits significant error during step changes. The
second method, called prediction-error covariance estimation (PECE), consists in a
convex optimization problem that is able to cope with step-processes. The adequacy
of the persistent process-model is statistically validated. We also theoretically prove
and numerically validate the better accuracy of the DKF with respect to the LWLS,
provided that a correct process-model is employed.

5
Introduction

In Chapter 3, we discuss the problem of LWLS-SE with linear equality constraints

associated with zero-injection buses. In particular, we propose a method based on LQ-
decomposition that strictly satisfies the equality constraints and reduces the dimension
of the state vector. We demonstrate the performance improvement of the proposed
method with respect to other well-known approaches to handle zero-injections via
numerical simulations.

In Chapter 4, we describe the implementation of RTSE in two real power systems: a

sub-transmission network and a distribution feeder equipped with PMUs. We briefly
describe the adopted measurement infrastructure, then we thoroughly analyze and
compare the SE results provided by the LWLS, LAV and DKF. These real-scale experi-
ments allow us to validate the findings of Chapters 1, 2 and 3 as well as to assess the
RTSE performance in real power networks.

In Chapter 5, we propose a new method based on LWLS-SE for fault detection and
faulted-line identification. We conducted extensive tests by considering different
network configurations and fault types as well as with a large amount of generation.
We evaluated both accuracy and time-latency performance of the proposed approach.

Finally, in Chapter 6 we provide the main outcomes of this dissertation and possible
directions for future research.

Original Contributions
The list of the original contributions of this dissertation is given in the following.

1. Included in Chapter 2 – We consider a persistent process-model (already used for

DKF-SE of power-systems) that is characterized by time-varying and unknown
uncertainties. Then, we propose a heuristic method that estimates the process-
noise variances used by the DKF from a set of previous state-estimates. This
method is conceived to maintain an effective filtering action in different network
operating-conditions that include quasi-static conditions but not step changes
in the system state. It is straightforward to implement and the user needs to set
a single parameter that is the number of previous state-estimates. Additionally,
it is designed to be computationally fast so that it is suitable for deployment in
real-time embedded hardware. It is worth pointing out that, although we test the
proposed heuristic method only in power-system SE, it can be also used in other
application fields as long as the assumptions hold.

2. Included in Chapter 2 – We propose a method, called prediction-error covari-

ance estimation (PECE), that, in addition to effectively filter the measurement
noise in quasi-static conditions, is able to follow the state during step changes.

6
 Introduction

The measurement model is assumed to be linear, known and time-invariant.

The PECE method consists in a convex optimization problem that infers the
prediction-error covariances from a set of previous innovations and is proved to
have a unique solution. Thanks to a set of constraints, the inferred prediction-
error covariance matrix is always positive semi-definite, which ensures the DKF
numerical stability. As for the heuristic method of point (1), the user needs to set
a single parameter that is the number of previous innovations. With respect to the
heuristic method, the PECE method is based on a rigorous theory and is able to
cope with step-varying processes, but it has a significantly higher computational
burden. To the best of our knowledge, no other method proposed in the literature
relying on the same assumptions is able to achieve the accuracy performance of
the PECE method in both quasi-static conditions and during step changes. The
PECE method can be applied also in other domains as long as the assumptions
hold.

3. Included in Chapter 2 – We present an extensive performance assessment of

the methods proposed in points (1) and (2) with respect to the classic LWLS via
numerical simulations. We assume that the measurements are affected only
by random noise and no systematic errors are present. We generated several
scenarios considering a transmission networks and a distribution feeder in quasi-
static conditions and when step changes in the state variables occur. We showed
that the DKF outperforms the LWLS in terms of accuracy when the network is
in quasi-static conditions. During a step change in the system state, the DKF
employing the PECE method of point (2) is able to accurately track the state,
whereas the DKF employing the heuristic method of point (1) experiences a
delay. The computational time of the DKF employing the heuristic method
is comparable with the LWLS one for the adopted case studies (medium size
networks), whereas the PECE method is computationally more expensive.

4. Included in Chapter 2 – Making reference to a DKF employing the heuristic

method of point (1), we statistically validate the adequacy of the persistent
process-model via numerical simulations. Moreover, we theoretically prove
and numerically validate the better accuracy of the DKF with respect to the LWLS,
provided that a correct process-model is employed.

5. Included in Chapter 3 – In the context of LWLS-SE with linear equality con-

straints associated with zero-injection buses, we propose a method based on
LQ-decomposition that strictly satisfies the equality constraints and reduces
the number of state variables, thus reducing the computational time. Other ad-
vantages of this method consist in making the problem less ill-conditioned and
in preserving both the structure of the LWLS normal equation and the positive
semi-definiteness of the coefficient matrix.

6. Included in Chapter 4 – We validate the theoretical findings via real-scale exper-

7
Introduction

iments. We installed PMUs at every bus in two real power-systems consisting

of a sub-transmission network and a distribution feeder. Network observability
and a high redundancy level are achieved by using exclusively synchrophasor
measurements. We describe the field implementation comprising sensor and
PMU installations, telecommunication network, phasor data concentrator and
RTSE. We demonstrate the feasibility of running RTSE at high refresh-rate (50
estimates-per-second) and low time-latency (below 70 ms). Then, the funda-
mental contribution consists in the presentation and detailed discussion of the
results of different linear SE algorithms (LWLS, LAV and DKF) for the two case
studies.

7. Included in Chapter 5 – We propose a fault detection and faulted-line identifi-

cation method based on LWLS-RTSE that uses synchrophasor measurements.
It does not need any model of loads/generators and has no prior knowledge of
the fault type and of the neutral connection, even though it requires that a PMU
is installed at every bus. We assume that every PMU measure the voltage and
current-injection phasors at the respective bus. We evaluated the faulted-line
identification accuracy by conducting extensive tests including different network
configurations and fault types as well as large amount of generation. We also
computed the time-latency and we demonstrated that PMU-based RTSE can
satisfy the requirements of fault location and, potentially, protective relaying.
Therefore, it may theoretically combine the execution of protection and fault-
location functions that are usually separated processes. The proposed approach
is particularly interesting for distribution networks where fault management can
be largely improved.

8
1 Theory on PMU-based power-
system state estimation

In this chapter we introduce the formulation of the PMU-based state estimation problem.
First, we describe the characteristics of the linear measurement model that links the
phasor measurements to the state variables. Second, we illustrate the formulation
and features of static and dynamic algorithms that we use in this dissertation: linear
weighted least squares (LWLS), least absolute value (LAV) and discrete Kalman filter
(DKF).

1.1 Linear Measurement Model

The main ingredient of a power-system state estimator is the measurement model
that is the function linking the state variables with the measurements. In this dis-
sertation, we assume to use only PMU synchrophasor measurements; therefore, the
measurement model becomes linear and exact provided that the phasors of both state
variables and measurements are expressed in rectangular coordinates. In most of the
literature on power-system SE and also in this dissertation, the state variables are the
bus-voltage phasors in the entire network. It is worth mentioning that current flows
have been used as state variables in distribution system SE [38]; in general, also bus
current-injections or branch voltages can be used, or even a mix of them, provided
that they are independent state variables.

Let us consider a three-phase power-system composed of a set of buses S with car-

dinality s = |S| and a set of branches D with cardinality d = |D|1 . The triplet of the
three phases is P = {a, b, c}. The network state is denoted by x ∈ Rn , where n = 3 · 2s
is the cardinality of the set of state variables N . The state is composed of the real and
imaginary parts of the bus phase-to-ground voltage phasors in every phase (a, b, c) at

1
The operator | | denotes the cardinality of a set.

9
Chapter 1. Theory on PMU-based power-system state estimation

every bus i ∈ S:
a,b,c a,b,c a,b,c a,b,c a,b,c a,b,c T
x = [V1,re , . . . , Vi,re , . . . , Vs,re , V1,im , . . . , Vi,im , . . . , Vs,im ] (1.1)

where

a,b,c a b c
Vi,re = [Vi,re , Vi,re , Vi,re ]
(1.2)
a,b,c a b c
Vi,im = [Vi,im , Vi,im , Vi,im ]

are, respectively, the real and imaginary parts of the voltage phasor at bus i ∈ S in
the three phases a, b and c. Note that we have not chosen any reference bus for the
voltage phase-angle as it is directly measured by PMUs [31]. All the voltage phase-
angle measurements are processed simultaneously by the state estimator together
with the other measurements; therefore, the state estimator can identify errors also in
phase-angle measurements, given a sufficient level of redundancy.

Let us define three sets of measurements: M1 is the set of buses where PMUs measure
phase-to-ground voltage phasors, M2 is the set of buses where PMUs measure current-
injection phasors, and M3 is the set of branches where PMUs measure current-flow
phasors. Then, the complete set of phasor measurements M with cardinality m = |M|
is assumed to be composed of:

• 3 · m1 bus phase-to-ground voltage phasors, where m1 = |M1 |;

• 3 · m2 bus current-injection phasors, where m2 = |M2 |;

• 3 · m3 current-flow phasors, where m3 = |M3 |;

Therefore, the measurement vector z ∈ Rm is equal to:

z = [ zV , zIinj , zIflow ]T (1.3)

where

a,b,c a,b,c a,b,c a,b,c a,b,c

zV = [ V1,re , . . . , Vi,re a,b,c
, . . . , Vm 1 ,re
, V1,im , . . . , Vi,im , . . . , Vm 1 ,im
] i ∈ M1
zIinj = [ Ia,b,c a,b,c a,b,c a,b,c a,b,c a,b,c
1,re , . . . , Ii,re , . . . , Im2 ,re , I1,im , . . . , Ii,im , . . . , Im2 ,im ] i ∈ M2 (1.4)
zIflow = [ Ia,b,c a,b,c a,b,c a,b,c a,b,c a,b,c
1,re , . . . , Ii,re , . . . , Im3 ,re , I1,im , . . . , Ii,im , . . . , Im3 ,im ] i ∈ M3 .

In the remainder of this dissertation, the bus phase-to-ground voltage is simply called
voltage and the bus current-injection is simply called current-injection, where not

10
 1.1. Linear Measurement Model

bus # i bus # h
branch
Iflow, sending Iflow, receiving
Iinjection

Load or
Generator
Figure 1.1 – Convention employed for current injections and flows with positive sign.

otherwise specified. Moreover, Fig. 1.1 shows the current directions that are assumed
to have positive sign. The convention employed for the current injections is as follows:
the positive sign denotes a current entering in a bus. Concerning the current flows,
the positive sign denotes a current exiting from a bus. On each branch there are two
current flows at both ends; if we consider a generic branch between buses # i and # h
where i < h, we call sending-end current flow the current flow departing from the
bus with the smallest index (bus # i) and receiving-end current flow the current flow
departing from the bus with the largest index (bus # h).

The linear measurement model that relates the measurements to the state variables is

z = Hx + v (1.5)

where H is a m × n matrix called measurement matrix and v is the measurement noise.

The measurement noise is assumed to be white and Gaussian:

p(v) ∼ N(0, R) (1.6)

where R is the measurement noise covariance matrix.

1.1.1 The measurement matrix H

The linearity of the measurement model makes H an exact matrix linking the measure-
ments with the state variables, whereas in conventional non-linear SE the matrix H is
a Jacobian matrix. Besides, H does not depend anymore on the state variables, so that
it remains constant in time with no need to update it at every time-step. It needs to
be recalculated only in case of changes in the network topology or in the composition
of the measurement set. Indeed, the elements of H are calculated from the network
topology and the electrical parameters of the network components, such as transmis-
sion lines and transformers. In Appendix A.1, we provide the derivation of the blocks
of H related to three type of measurements: voltage phasors, current-injection phasors
and current-flow phasors.

In addition to the aforementioned real-time phasor measurements from PMUs, in this

11
Chapter 1. Theory on PMU-based power-system state estimation

dissertation we will also consider pseudo-measurements and zero-injection buses.

Details are given in what follows.

Pseudo-measurements are, typically, nodal power-injection measurements consist-

ing in predictions based on historical data or forecasts; therefore, the uncertainty
associated with pseudo-measurements is large. To maintain the linearity of the mea-
surement model, power-injection pseudo-measurements at a bus can be transformed
in an equivalent current-phasors by assuming to know the voltage at that bus. Even if
the voltage is not directly measured, we can assume that it is close to 1 per unit (pu)
in normal operating conditions. This procedure has been already proposed in the
literature, e.g., [38].

Zero-injection buses are network buses characterized by an injected power strictly

equal to zero, which is, for instance, the case of interconnection substations. In linear
SE, zero-injections are treated as current phasors with real and imaginary parts equal
to zero. As this information contains no error, it constitutes a set of equality constraints
that can be added to power-system SE. In the literature there are several methods
to include equality constraints in SE. This subject is treated in Chapter 3 with focus
on linear SE. A simple yet approximated way to include equality constraints in SE is
to convert the constraining equations in additional measurements with very small
uncertainties, so-called virtual measurements [39, 40].

It is important to point out that in this dissertation we make a clear distinction between
virtual measurements and pseudo-measurements, whereas in the literature virtual
measurements are often considered as a particular type of pseudo-measurements.
Virtual measurements approximate the exact information linked to zero-injections
and pseudo-measurements are predictions characterized by large uncertainties that
are typically based on historical data or forecasts.

1.1.2 The measurement noise covariance matrix R

In this dissertation, we assume that the measurement noise covariance matrix R is

diagonal, which implies that there is no correlation between measurements (i.e., the
measurements are independent and identically distributed).

This is a common assumption in power-system SE, although there are cases where
this correlation is not negligible and accounting for that in the matrix R can lead to
an improvement of the SE accuracy. For example, this is the case of single-phase
measurements that are derived from three-phase multifunction meters [41], or when
we use pseudo-measurements that can be correlated between each other [42]. Whereas,
we assume that:

• the measurements are provided by different sensors installed in every phase;

12
 1.1. Linear Measurement Model

• the cross-talk interferences between nearby sensors are negligible;

• the measurements in every phase are treated separately and not merged in a
single-phase measurement.

Besides, it has been demonstrated that the correlation between the multiple channels
of the same PMU can be neglected without affecting the SE accuracy [42].

Each diagonal entry of R is the variance of a measurement, which includes the cumu-
lative uncertainty of both sensor and PMU. In this dissertation, a sensor refers to an
electrical device that transforms the input voltage or current signal taken from the elec-
trical network into signals that can be acquired by instruments or meters. Examples
of sensors are potential and current transformers (PTs and CTs), capacitor dividers or
Rogowsky coils. In the context of PMU-based SE, the sensor interfaces the network
voltage and current signals with a PMU.

The measurement uncertainty introduced by sensor and PMU is generally character-

ized by a component associated with a random effect (uncertainty of type A) and a
component associated with a systematic effect (uncertainty of type B). The definitions
of error and uncertainty are given in the Guide to the expression of uncertainty in
measurement (GUM) [43], and briefly recalled here below.

A measurement result is always affected by an error that has two components called
systematic and random. Note that the measurement error cannot be known exactly.
The GUM defines the two error components as follows [43]:

Random error presumably arises from unpredictable or stochastic temporal

and spatial variations of influence quantities. The effects of such variations, here-
after termed random effects, give rise to variations in repeated observations of
the measurand. Although it is not possible to compensate for the random error
of a measurement result, it can usually be reduced by increasing the number of
observations; its expectation or expected value is zero.
Systematic error, like random error, cannot be eliminated but it too can often
be reduced. If a systematic error arises from a recognized effect of an influence
quantity on a measurement result, hereafter termed a systematic effect, the effect
can be quantified and, if it is significant in size relative to the required accuracy of
the measurement, a correction or correction factor can be applied to compensate
for the effect. It is assumed that, after correction, the expectation or expected value
of the error arising from a systematic effect is zero.

Then, the uncertainty of a measurement result represents the lack of exact knowledge
of the measurand. The measurement result can be corrected from the recognized
systematic effects, but it is still only an estimate of the value of the measurand, because
of the uncertainty arising from random effects and from imperfect correction of the
systematic effects.

13
Chapter 1. Theory on PMU-based power-system state estimation

In the GUM, the uncertainty components are grouped into two categories based on
their method of evaluation, type A or type B evaluation, that is used to obtain the
uncertainty characterizing a random or systematic effect, respectively. Both types of
evaluation are based on probability distributions and the uncertainty components
resulting from either type are quantified by variances or standard deviations.

Therefore, the estimated variance characterizing an uncertainty component obtained

from a type A evaluation is calculated from series of repeated observations and is
the statistically estimated variance, and the estimated standard deviation is called a
type A standard uncertainty. The type A evaluation estimates the expected value and
the variance of a random quantity as the arithmetic mean and the square root of the
experimental standard deviation of the observations, respectively.

For an uncertainty component obtained from a type B evaluation, the estimated

variance is evaluated using available knowledge (not from repeated observations),
and the estimated standard deviation is called a type B standard uncertainty. The
latter is evaluated by scientific judgment based on all of the available information,
such as previous measurement data, experience, manufacturer’s specifications (e.g.,
instrument accuracy-class2 ), calibration certificates, etc.

If no information is available on the uncertainty distribution, one may assume that

a normal distribution was used to calculate the quoted uncertainty, and recover the
standard uncertainty by dividing the quoted uncertainty by a coverage factor for the
normal distribution, which is usually 3.

The type A uncertainty of a voltage or current measurement result accounts for the
random noise introduced by grid phenomena (e.g., thermal effects, converters), sensor
and A/D converter of the PMU. This noise, and thus the uncertainty, is highly depen-
dent upon the grid, the type of employed sensor and the PMU hardware/software.
Note also that, given multiple measurements of a certain quantity, it is also not trivial
to separate the noise from the variations of this quantity. To remove the quantity’s
variations, one should make the difference between two or more measurements of the
same quantity taken from distinct meters. From the analysis of actual PMU data, the
distribution of the type A uncertainty can be considered Gaussian [44].

The type B uncertainty of a sensor is specified by its class by means of the Standard
IEC 61869 that gives the maximum magnitude and phase-angle errors. For instance,
for CTs and PTs one should make reference to [45] and [46], respectively. Sometimes
the sensor’s calibration certificates are available so that the systematic effect can be
compensated. As stated in the GUM, in this specific case we can assume that the
expected value of the error arising from the sensor’s systematic effect is zero. The type

2
The accuracy class is a designation assigned to a measurement instruments the errors of which
remain within specified limits under prescribed conditions of use

14
 1.1. Linear Measurement Model

B uncertainty of a PMU is commonly derived by its total vector error (TVE) that is
defined in [47]. When separate values for the magnitude and phase-angle errors are
given, then these values are considered; otherwise, we need to make assumptions on
how the TVE is distributed between voltage magnitude and phase-angle (e.g., [40]). If
we assume a normal distribution, the type B uncertainty can be calculated by dividing
the uncertainty by a coverage factor, typically equal to 3.

The total uncertainty of a measurement is the sum of the type A and type B uncer-
tainties of sensor and PMU. This uncertainty, expressed as a variance, composes the
generic diagonal element of the measurement noise covariance matrix R.

The uncertainty of the instrumentation measuring voltage or current is commonly

given in terms of magnitude and phase-angle. In the linear measurement model
presented in 1.1, the phasor measurements are expressed in rectangular coordinates;
thus, we need to project the uncertainty from polar to rectangular coordinates. To this
end, we use the procedure reported in [48, Chapter 6.4]. The final formulas used to
perform the projection are recalled here below.

Let us assume to have a voltage-phasor measurement expressed in terms of magnitude

V and phase-angle δ, and the corresponding uncertainties σV and σδ , respectively. The
variances of the projected real and imaginary parts of the voltage phasor are

2
  
σV2re = V 2 e−2σδ cos2 δ cosh(2σδ2 ) − cosh(σδ2 )
 
+ sin2 δ sinh(2σδ2 ) − sinh(σδ2 )
2
  
+ σV2 e−2σδ cos2 δ 2 cosh(2σδ2 ) − cosh(σδ2 )
 
+ sin2 δ 2 sinh(2σδ2 ) − sinh(σδ2 ) (1.7)

2
  
σV2im = V 2 e−2σδ sin2 δ cosh(2σδ2 ) − cosh(σδ2 )
 
+ cos2 δ sinh(2σδ2 ) − sinh(σδ2 )
2
  
+ σV2 e−2σδ sin2 δ 2 cosh(2σδ2 ) − cosh(σδ2 )
 
+ cos2 δ 2 sinh(2σδ2 ) − sinh(σδ2 ) . (1.8)

In the literature, the measurement uncertainty is calculated in different ways: (1) as

a constant value [6, Chapter 7.2], [24], (2) as a percentage of the measured quantity

15
Chapter 1. Theory on PMU-based power-system state estimation

[49, 50], or (3) as a combination of the two [51, 23].

In the works presented in this dissertation, we will always specify the assumptions
on the measurement uncertainty. Based on these assumptions, R will be considered
time-variant or -invariant.

1.2 Linear weighted least squares

The WLS estimator is a static estimator as it does not exploit any time-variant de-
pendency of either the measurements or the state variables. It minimizes the sum of
the squared residuals weighted by the measurement variances, a residual being the
difference between the measurement and the fitted value obtained with the model.
The WLS relies on the following assumptions:

1. The measurement noise is Gaussian-distributed with mean value equal to zero.

2. The measurements are uncorrelated, so that R is diagonal.

3. Matrix H is of full rank, which means that the network is observable. Specifically:
(1) m ≥ n, which is a necessary condition for the network observability, and (2)
rank(H) = n.

, the
Given the linear measurement model in (1.5) and defining the estimated state as x
LWLS residual is defined as

r = z − H
x. (1.9)

We use the hat “” to denote estimated quantities throughout the dissertation. The
LWLS consists in solving an unconstrained optimization problem that minimizes the
following objective function:

1
min x) = rT R−1 r
J( (1.10)

x 2

that, as R is assumed to be diagonal, can be also written as

1  ri2
m
min J(
x) = . (1.11)

x 2 Rii
i=1

The solution of this optimization problem reduces to the well-known normal equation
(NE):

x = HT R−1 z
G (1.12)

16
 1.2. Linear weighted least squares

where G is the so-called gain matrix:

G = HT R−1 H . (1.13)

Equation (1.12) is a system of linear equations of the well-known form Ax = b, where
A is the coefficient matrix and corresponds to G, and b corresponds to HT R−1 z. The
NE is usually solved by Cholesky factorization as G is positive definite. Therefore, G
does not need to be inverted. The covariance matrix of the estimated state is given by

x) = G−1 .
cov( (1.14)

It is also interesting to give the relation between measurements z and estimated-

measurements  z, which is given by the following expression:

 x = HG−1 HT R−1 z = Kz
z = H (1.15)

The row entries of K (so-called hat matrix) indicate the local redundancy of every mea-
surement. For instance, a large diagonal element (relative to the off-diagonal elements)
implies that the estimated value of a measurement is almost entirely determined by its
measured value, i.e., the local redundancy is poor.

The residual vector is linked to the measurement-error vector e = z − Hx by the

so-called residual sensitivity matrix S, which represents the sensitivity of the residuals
to the measurement errors:

r = z−
z = (I − K)e = Se (1.16)

The matrices K and S can be checked to identify critical and leverage measurements,
which are both defined in what follows.

A critical measurement is defined as a measurement whose loss makes the network

unobservable, i.e., decreases the rank of H. A critical measurement appears in parts of
the network characterized by low local redundancy. Several topological or numerical
methods for the observability analysis and the identification of critical measurements
have been proposed in the literature, e.g., [52, 53].

A leverage measurement is defined as a measurement characterized by a structure

of the measurement equation that is different from that of the other measurements
[54]. Specifically, the row of H corresponding to a leverage measurement has very
different (smaller or larger) entries compared to the other rows. This creates a kind of
insensitivity of the state estimator with respect to this measurement, i.e., the residual
of a leverage measurement remains small even in presence of a large measurement
error. Leverage measurements can be identified with various methods proposed in the
literature, e.g., [55].

17
Chapter 1. Theory on PMU-based power-system state estimation

Both critical and leverage measurements have very small (identically zero for critical
measurements) residuals even when the measurements are affected by a large error,
which means that the corresponding diagonal element of S is very small (identically
zero for critical measurements). Therefore, the effects of critical and leverage measure-
ments are similar, the main difference lying in the fact that the elimination of leverage
measurement does not lead to a loss of network observability.

As it is known, the WLS is vulnerable to bad data as they lead to biased estimates.
Therefore, the WLS solution needs to be given to a bad-data processor that detects,
identifies and removes bad data. A widely-used method for bad-data processing is
the largest normalized residual test (LNR) [8, 6]. First, the residual vector is computed
using (1.9), then its covariance matrix is calculated as

cov(r) = Ω = R − HG−1 HT = RS (1.17)

and the normalized residual of the ith measurement is

|ri |
riN = √ . (1.18)
Ωii

The normalized residuals are Gaussian-distributed as ∼ N(0, 1). Therefore, the LNR
test detects the presence of bad data when the LNR exceeds a statistical threshold,
usually 3 or 4. In this case, the measurement with the LNR is flagged as bad data and
the SE is re-computed without using this measurement. We repeat the procedure until
no more bad data are detected. The LNR test can reliably identify only a single bad-data
or multiple non-interacting bad-data (i.e., simultaneous bad-data whose residuals
are not correlated). In presence of multiple interacting bad-data (i.e., simultaneous
bad-data whose residuals are correlated), the hypothesis testing identification method
is proven to have superior performance [9, 6].

1.3 Least absolute value

The LAV belongs to the category of robust estimators 3 , because it possesses an intrinsic
bad-data rejection capability and has no need to employ a separate bad-data processor
[32]. The LAV is the most employed robust estimator in power-system SE, as it can be
made computationally efficient by exploiting the power system’s properties. Despite
its desirable bad-data rejection property, the LAV has two main shortcomings: the
computational time is not competitive with that of the WLS, and the algorithm remains
vulnerable against leverage measurements (see Section 1.2 for the definition of leverage

3
A robust estimator is an estimator with high breakdown-point, i.e., the smallest number of bad data
that can lead to an incorrect SE solution. Note that the breakdown point is highly influenced by the
measurement redundancy; indeed, a poor redundancy limits the number of bad data that the estimator
can reject.

18
 1.3. Least absolute value

measurements). However, these weaknesses can be alleviated when the measurement

model is linear thanks to the use of PMUs [34]. The remarkable improvements of the
LAV performance are listed here below:

• the LAV computational time becomes comparable to that of the LWLS, especially
when bad data are present;

• the LAV optimization problem reduces to a linear-programming problem;

• the vulnerability to leverage measurements can be eliminated by strategic scaling

of the measurement matrix without affecting the SE solution.

The LAV is an optimization problem that minimizes the sum of the absolute values of
the residuals:

min cT |r| (1.19)

subject to: z − H
x=r

where c is a vector of ones. Note that the LAV does not require the measurement noise
covariance matrix R. The computation of the R entries is not trivial, as shown in
Section 1.1.2.

In [34], the optimization problem (1.19) is re-arranged as an equivalent linear pro-

gramming problem. The procedure is recalled in what follows. First, the problem is
formulated as

min c̄T |r̄| (1.20)

subject to: M̄r̄ = z
r̄i ≥ 0 i = 1, ..., 2n + 2m

where

c̄T = [02n 12m ] (1.21)

r̄ = [xa xb ra rb ] T
M̄ = [H − H Im − Im ] .

In (1.21), 02n is a 1 × 2n vector of zeros, 12m is a 1 × 2m vector of ones, Im is a m × m

identity matrix, xa and xb are 1 × n vectors, and ra and rb are 1 × m vectors. Finally, we
can compute the estimated state and the residual as

x = xTa − xTb (1.22)

r = rTa − rTb .

19
Chapter 1. Theory on PMU-based power-system state estimation

1.4 Dynamic linear estimation via discrete Kalman filter

The SE accuracy can be further enhanced by filtering the measurement noise. This
can be accomplished by using dynamic SE. We recall that dynamic SE models the
time-evolution of the physical process by means of a process model. Dynamic SE
consists in computing the probability density function of the system state in two steps:
the prediction step, also called time update, exploits a process model to predict the
state; the estimation step, also called measurement update, introduces the incoming
measurements to compute the estimated state. Hence, a sort of memory is introduced
in the state estimator, because it also employs prior information contained in the state
estimates of the previous time-steps.

1.4.1 General process-model and discrete Kalman filter

Let us consider a general process model consisting in the following discrete-time

time-variant linear equation [56]:

xk = Ak xk−1 + Bk uk + wk (1.23)

where

• k is the time-step index;

• x ∈ Rn is the system state;

• u ∈ Ruc is a set Uc of known controllable variables and uc = |Uc |;

• w ∈ Rn is the process noise;

• A is an n×n matrix, called state-transition matrix, that links the state at time-step
k − 1 with the state at time-step k;

• B is an n × uc matrix that links the state with the controllable variables;

The process noise wk is assumed to be a Gaussian-white sequence:

p(wk ) ∼ N(0, Qk ) (1.24)

where Qk is the process noise covariance matrix. Note that the matrices Ak , Bk and Qk
are, in general, time-variant.

In standard theory of dynamic state estimators, it is also assumed that the process
noise wk in (1.23) and the measurement noise vk in (1.5) are uncorrelated:

E[wk vkT ] = 0 . (1.25)

20
 1.4. Dynamic linear estimation via discrete Kalman filter

One of the most common dynamic estimators is the KF. The DKF is the linear KF
resulting from the linear process and measurement models defined in (1.23) and (1.5),
respectively.. The DKF relies on the following assumptions:

1. The process and measurement models are linear;

2. The process and measurement noises are Gaussian-distributed with mean value
equal to zero, as denoted by (1.6) and (1.24), respectively;

3. The process and measurement noises are uncorrelated, as denoted by (1.25);

4. Matrix H is of full rank, which means that the network is observable. Specifically:
(1) m ≥ n, which is a necessary condition for the network observability, and (2)
rank(H) = n.

The DKF recursive equations are given here below [56]:

Prediction step

k|k−1 = Ak x
x k−1|k−1 + Bk uk (1.26)

Pk|k−1 = Ak Pk−1|k−1 ATk + Qk (1.27)

Estimation step

Lk = Pk|k−1 HT (HPk|k−1 HT + Rk )−1 (1.28)

x k|k−1 + Lk (zk − H
k|k = x xk|k−1 ) (1.29)

Pk|k = (I − Lk H)Pk|k−1 (1.30)

where x k|k−1 is the predicted state given the knowledge of the process prior to time-step
k, Lk is the Kalman gain, x k|k is the estimated state given zk , and I is the identity matrix.
Finally, Pk|k−1 and Pk|k are the covariance matrices of the prediction and estimation
errors, respectively. Note that H lacks of the subscript k, because it is time-invariant
except for sporadic changes in the network topology.

From (1.26)–(1.30), it is clear that the KF optimality strongly depends on the selection
of the correct values of Qk and Rk . We will thoroughly discuss the assessment of the

21
Chapter 1. Theory on PMU-based power-system state estimation

process-model uncertainty in Chapter 2.

After the prediction step, we can also define the innovation yk and its covariance matrix
Tk as

yk = zk − H
xk|k−1 (1.31)

Tk = HPk|k−1 HT + Rk . (1.32)

The innovation plays an important role in Kalman filtering as it is used to check the
optimality of the filter as well as for bad-data detection/identification.

1.4.2 Historical notes on Kalman filter in power-system state estimation

In this section, we present in a chronological order some relevant contributions that

deal with power-system state estimation via KF. Other works that are inadvertently
missed constitute equally important contributions.

Since 1960 when R. Kalman presented the KF in his famous work [33], the research on
KF applied to power-system SE has been limited to few contributions. In principle,
the KF can provide an estimation accuracy better than the WLS, but this goes at the
expense of a higher algorithm complexity. Specifically, some of the problems related to
the KF implementation are listed here below:

1. the need of a process model that matches the power-system state dynamics that
we are interested to track, which is not trivial to be determined and may involve
the definition of several parameters;

2. the development of complementary applications becomes more challenging. For

instance, a bad-data processor has to take into account that a large residual can
be due not only to a measurement error, but also to a sudden state-change that
violates the process model;

3. the higher computational time does not facilitate real-time implementation.

In 1970, Schweppe mentioned the possibility to use a process model already in his first
work [2] and shortly after Debs and Larsonn dedicated an entire paper to power-system
SE via KF [57]. They propose a linear KF that employs a linear process model and a DC
load-flow approximation4 for the measurement model. The employed process model
4
The DC load-flow approximation consists in assuming loss-less lines (resistance equal to zero)
and voltages equal to nominal values of 1 pu at every bus. Therefore, losses and reactive-powers are

22
 1.4. Dynamic linear estimation via discrete Kalman filter

is the well-known random-walk process-model denoted by

xk = xk−1 + wk (1.33)

where the only parameter to be set is the process-noise covariance matrix Qk . This
matrix is assessed by an a-priori direct search that is known to be a neither robust nor
optimal approach as the simulated scenarios cannot take into account all the operating
conditions. The KF and WLS results are compared as a function of the process-noise
covariance matrix and of the time-resolution of the measurements. This contribution
remained a sporadic attempt and KF was rapidly abandoned in favor of the WLS.

Nishiya et al. [49] propose a method that is able to discriminate between three anoma-
lies, i.e. occurrence of bad data, changes in network configuration and sudden state-
variations. It consists in examining the innovation of a non-linear version of the KF,
so-called extended KF (EKF), that uses (1.33) as process model. In particular, the skew-
ness of the distribution of the innovation sequence is an indicator of the presence
of bad-data. Then, if the normalized innovations of the power-flow measurements
exceed a certain threshold, a change in the network configuration is suspected and
the network topology is re-estimated. Otherwise, the large values of the normalized
innovations are attributed to a sudden state-variation.

Da Silva et al. [51] propose an EKF that tracks the state of power systems operating
under quasi-static conditions. They use the general process-model of (1.23) where
A and B are assessed on-line by means of the Holt’s linear exponential smoothing
method presented in [58], whereas Q is kept constant. Nevertheless, the authors
suggest adaptive estimation of Q for improved performance. It is shown that the
KF looses track of the state trajectory when a sudden change occurs. In this case,
the advice of the authors is either to disregard the prediction or to use more frequent
measurements. In addition, a bad-data processing algorithm based on both innovation
and residual vectors is proposed.

In [59], the authors present an EKF that incorporates short-term load forecasting based
on artificial neural network to obtain a realistic process model. The estimation step is
iterative as they include second-order terms in the approximation of the non-linear
measurement function. The scheme for anomaly identification proposed in [49] is
used.

The EKF proposed in [60] incorporates a weight exponential-function in order to

increase the robustness of the filter. It includes second-order terms in the EKF as in
[59]. The performance is shown in presence of the three anomalies mentioned in [49].

The Authors of [61] explored a process model based on the load-flow equations. The
disregarded and the voltage phase-angles are linearly related to active powers. See [1, Chapters 9.7 and
14.5] for further details.

23
Chapter 1. Theory on PMU-based power-system state estimation

EKF performance are assessed by employing the DC load-flow approximation. In [62],

this approximation is not employed in order to include also the voltage magnitude
in the state vector by using active- and reactive-power measurements. They use an
iterated version of the EKF called iterated EKF.

In what follows, we present the recent works on power-system SE via KF and it is worth
pointing out that all of them assume to have PMU data in the measurement set.

In [63] and [64], Jain and Shivakumar evaluate the impact of using PMU phasor mea-
surements in combination with SCADA measurements on the estimation accuracy
of an EKF state estimator. They employ the process model proposed in [51]. Later in
2009, they also published a literature review on tracking and dynamic SE techniques
[65]. A tracking estimator is a static estimator that uses new measurements to update
the past estimate instead of fully executing the iterative estimation from a flat-start 5
initialization at every time-step [66, 58].

The Authors of [50] propose to use an unscented KF [67], which is a KF that is able to
deal with highly non-linear models and does not need the computation of the Jacobian
matrices. They use the scheme for anomaly identification proposed in [49] and they
employ the process model proposed in [51]. Due to the fact that Q is kept constant,
it is shown that the unscented KF estimation performance is degraded as soon as a
sudden state-change occurs, because the consequent state variation is not taken into
account in the process model. Once the anomaly is detected, the authors recommend
to increase Q or Pk|k−1 , as was already suggested in [51]. Another alternative is to
re-initialize the state vector with a static estimator.

The Authors of [37] propose a two-stage KF: a DKF uses only PMU phasor measure-
ments to estimate the static state, i.e., voltage magnitudes and phase-angles; the
estimated static-state is passed to an EKF that estimates the dynamic state, i.e., the gen-
erator speeds and rotor-angles. The DKF is a novel adaptive KF (AKF) with inflatable
noise variances. The AKF with inflatable noise variances inflates R and/or Q in case of
measurement errors and/or unexpected state-variations, respectively. It employs the
persistent model already proposed in [57]. The method is tested in fault condition and
in presence of bad-data.

In [68] a mixed-integer programming formulation of a distribution state estimator

based on PMUs is presented. The process model is the one used in [51] where the
parameters, including Q, are determined by trial-and-error. Then, they propose a new
formulation of the estimation step that is capable of simultaneously discarding pre-
dicted values whenever sudden state-changes are detected. The method is applicable
5
A flat-start means that all voltage phase-angles are set to zero and all voltage magnitudes are set
to 1 pu. This is a typical initial guess in iterative procedure (e.g., load flow or SE) as it is close to the
power-system state in normal operating conditions where the voltage magnitudes are close to the rated
values and the voltage phase-angle differences are small.

24
 1.4. Dynamic linear estimation via discrete Kalman filter

only if the state variables are the same as the measurement quantities. Therefore, indi-
rect voltage phasor measurements at buses without PMUs are calculated via voltage
and current measurements of adjacent buses by using the line parameters.

In [69], a real-time simulator (RTS) is used to test and compare three non-linear KFs:
EKF, unscented KF and cubature KF. Indeed, they use both SCADA and PMU mea-
surements. The cubature KF, invented by the authors of [70], is designed to handle
high-dimensional non-linear problems under the Gaussian assumption of the noise
distributions.

1.4.3 PMU-based state estimation via discrete Kalman filter

Many contributions cited in the previous Section 1.4.2 focused on how to improve the
approximation of model non-linearities and on the definition of new process models
aiming at accurately represent the power-system time-behavior. These issues are no
longer a concern when we consider state estimators based exclusively on PMUs.

The measurement model is linear (see Section 1.1) and the state is estimated every 20
ms with reference to a 50 Hz system. The first observation is that it makes no sense
to predict the state for such a short-time period. Thanks to the high-resolution PMU
measurements, the power-system can be considered to be in quasi-static conditions
during normal operation. We can assume that the state does not change significantly
from one time-step to the other, so that a good approximation of matrix A in (1.23) is
the identity matrix. In addition, the power-system inputs are usually not controllable
from the SE perspective, thus, we can remove the controllable inputs u in (1.23). There-
fore, the process model already proposed in [57, 37] is suitable for these conditions
and is recalled here below:

xk = xk−1 + wk . (1.34)

This model is the autoregressive integrated moving average – ARIMA (0,1,0), also called
persistent model or random-walk. A remarkable advantage of using this process model
is that it requires the setting of a single parameter, i.e., Q.

The linearity of both the process and measurement models enables the use of the
DKF. In the following, the DKF equations that were given in Section 1.4.1 for a general
process model are presented for the persistent process-model:

25
Chapter 1. Theory on PMU-based power-system state estimation

Prediction step

k|k−1 = x
x k−1|k−1 (1.35)

Pk|k−1 = Pk−1|k−1 + Qk (1.36)

Estimation step

Lk = Pk|k−1 HT (HPk|k−1 HT + Rk )−1 (1.37)

k|k = x
x k|k−1 + Lk (zk − H
xk|k−1 ) (1.38)

Pk|k = (I − Lk H)Pk|k−1 (1.39)

26
2 Adaptive Kalman filtering for the
on-line estimation of the process-
model uncertainties
In this chapter, we deal with the on-line estimation of time-varying uncertainties as-
sociated with the process-model used by a discrete Kalman filter (DKF). We present a
heuristic method that assesses the process-noise variances from the previous estimates
and we statistically validate the adequacy of the persistent process-model in quasi-static
conditions. Then, we define a convex optimization problem that is able to cope with step-
processes by inferring the prediction-error covariances from a set of previous innovations;
indeed, this method is called prediction-error covariance estimation (PECE). The two
above-mentioned methods are tested in the context of power-system state estimation,
but they can be applied also in other domains as long as the assumptions hold.

2.1 Introduction
In Chapter 1 we explained how the assumption of a measurement infrastructure fully
based on PMUs leads to the definition of two linear models:

• Persistent process-model (1.34)

xk = xk−1 + wk (2.1)

where p(w) ∼ N(0, Qk ) in which Qk is, in general, time-variant.

• Measurement model (1.5)

zk = Hxk + vk (2.2)

where H is time-invariant and p(v) ∼ N(0, Rk ), whereas Rk is, in general, time-

variant.

27
Chapter 2. Adaptive Kalman filtering for the on-line estimation of the
process-model uncertainties

Then, we gave the corresponding DKF equations in (1.35)–(1.39), which involve the
definition of the covariance matrices of the measurement noise R and of the process
noise Q. The KF optimality depends upon the proper assessment of both covariance
matrices. Indeed, a well-known challenge in the application of the KF to real systems
is the identification of the parameters of the stochastic-error distributions.

R represents the uncertainties associated with the measurements. As presented in

Section 1.1.2, these uncertainties consist of systematic and random errors. The random
errors may come from different sources, such as thermal noise and converters, whereas
the systematic errors are mainly introduced by the sensors, such as PTs and CTs. It
is known that, typically, the systematic errors of voltage and current measurements
in power systems are predominant. However, in this section we assume to have the
calibration certificates of the sensors so that we can neglect the systematic errors,
as explained in Section 1.1.2. Therefore, the diagonal elements of R represent the
variances of the random errors that are assumed to have a Gaussian-white distribution.
Besides, R is diagonal as we neglect the correlations between measurements for the
reasons mentioned in Section 1.1.2.

Q represents the process-model uncertainties associated with each state variable. The
assessment of Q is complicated as the power-system state is continuously moving. In
the literature dealing with KF-based power-systems SE (see Section 1.4.2), Q is usually
fixed to a certain value assessed from historical data, which is typically the largest value
that allows the process model to remain valid even in worst case scenarios. This can be
a very sub-optimal procedure, because it cannot adapt to the continuously-changing
operating conditions of the network. For instance, if we set a very large value of Q, the
process model is never violated and the KF never looses track of the system, but the
KF becomes essentially a LWLS; therefore, we do not exploit the filtering capability of
the KF as we use it in a very sub-optimal way. Especially in our case where we use the
persistent process-model, every variation of the power-system state is attributed to the
process noise and, thus, is reflected in the process-noise covariance matrix. Therefore,
we need to use an AKF that performs an on-line assessment of the process-model
uncertainties. This is a quite challenging task and, indeed, it is still a research subject
in many different domains, especially in the area of automatic control.

Another hidden hypothesis is that the process noise has to be smaller than the mea-
surement noise in order to allow the KF to effectively filter the measurement noise.
If it is not the case, it is logical to think that the measurements are already very close
to the optimal estimate achievable and, again, the KF cannot do much better than a
WLS. Indeed, when the process noise is larger than the measurement noise, the entries
of Q are larger than those of R and it is known that the KF solution approaches the
WLS one as Q → ∞ [57]. The aforementioned hypothesis is satisfied thanks to the
high time-resolution measurements of PMUs, which allows us to consider the power
system to be in quasi-static conditions, except for sporadic events, such as changes of

28
 2.2. Literature review on the estimation of the Kalman filter covariances

network configuration, load inrushes, faults or load/generator disconnections.

In this chapter, we present two methods that continuously-estimate the process-model

uncertainty in order to adapt it to different power-system operating conditions. Our
goal is to enhance the SE accuracy by filtering the measurement noise. The resiliency
against bad data and network-model errors is not treated in this chapter.

In Section 2.2, we provide the literature review on the assessment of the KF covariance
matrices.

In Section 2.3, we formally prove the better accuracy of the KF over the WLS if the
process model is correct.

In Section 2.4, we present a heuristic method for the assessment of Q that was firstly
proposed in [71] and further investigated in [72]. This method is effective in terms of
estimation accuracy when the network is in quasi-static conditions and is suitable for
real-time applications as the computation of Q takes less than a millisecond.

In Section 2.5, we present a more sophisticated method that is able to track sudden
changes of the system state. It consists in a constrained convex optimization problem
that has the computational time as main drawback.

2.2 Literature review on the estimation of the Kalman filter

covariances
In the early 1970’s, Mehra [73] classified into four categories the methods for the
estimation of the KF covariances: Bayesian [74, 75], maximum likelihood [76, 77, 78],
correlation [79, 80, 81, 82, 83, 84, 85, 86, 87, 88], and covariance matching [89].

The Bayesian and maximum-likelihood estimation methods are characterized by a high

level of complexity and are usually employed for the case of time-invariant systems. As
observed in [78], a relevant advantage of these approaches is that they can be applied
also when measurements are available at irregular intervals, which is quite common in
real applications.

Correlation techniques estimate the measurement and process-noise covariance ma-

trices by exploiting the sample autocorrelation functions of the innovations at different
lags. In [79, 80, 82, 83, 84, 85, 86, 87], the authors consider only time-invariant systems,
whereas [81] extends the problem formulation to time-variant systems. However, as
remarked in [73] and shown in [81], the correlation methods provide sound results
mainly for time-invariant systems. In [83], the authors propose a least-squares opti-
mization problem based on the correlations between the innovations; this method
is able to ensure the positive semi-definiteness of both R and Q, which is essential

29
Chapter 2. Adaptive Kalman filtering for the on-line estimation of the
process-model uncertainties

to guarantee the KF numerical stability and, thus, the accuracy of the solution (see
Section 2.5.2). Then, a follow-up of this paper is presented in [86], where the authors
present new conditions for the uniqueness of the covariance estimates as well as
an optimal weighting to be used in the least-squares objective to ensure minimum
variance in the estimates. The work presented in [88] deals with adaptive estimation
of both R and Q for time-variant models. The method is based on the correlation-
innovations approach and includes an approach to ensure the positive definiteness of
the covariance matrices.

Covariance-matching methods consist in adaptive algorithms that, at every sampling

step, assess the KF parameters directly from the past state estimates. They are par-
ticularly effective in case either R or Q is known, as stated in [73]. Myers and Tapley
[89] use a direct estimation of both R and Q from their sample covariances at every
sampling step.

Correlation and covariance-matching techniques need ad-hoc procedures in order to

ensure the positive semi-definiteness of the estimated covariance matrices, which is
not guaranteed by the method itself. The authors of [89] use an approximated coun-
termeasure to ensure positive diagonal elements of the covariance matrices, which
consists in replacing the estimated diagonal elements with their absolute values. Simi-
larly, in [88] the positive definiteness of the estimated R and Q is ensured with further
processing of the two matrices after their estimation; therefore, also this method is ap-
proximated. An effective way to ensure the positive semi-definiteness of the estimated
covariance matrices is to solve a constrained optimization problem, as in [83].

In the context of power systems, very few papers deal with the evaluation of the KF
process-model uncertainty. The value of Q is usually assumed to be constant, which
leads to a non-adaptive filter that could have very poor estimation accuracy with
respect to a filter whose stochastic parameters are frequently updated. In [37], the
authors formulate an optimization problem that is only able to inflate the process-
noise variances when a step-variation in the power-system state occurs. The authors
mention that the inflated Q could be decreased by employing an exponential decay
with time constant to be defined by the user.

2.3 Estimation accuracy: Kalman filter vs. weighted least squares

The content of this section is based on [72]. The accuracies of KF vs. WLS are theo-
retically compared as explained in what follows. To formally quantify this difference,
it is useful to recall that the KF process makes use of all the available measurements,
past and present, whereas the WLS algorithm uses only measurements of the current
time-step. The former should intuitively perform better, provided that the process
model hypotheses that underlie the KF are correct. The following theorem formalizes

30
 2.3. Estimation accuracy: Kalman filter vs. weighted least squares

this aspect. It states that the estimation error with the KF algorithm is always less than
the estimation error with the WLS algorithm, the difference being given equal to the
mean square difference between the two methods:

Theorem 1. Assume that the true (unknown) state xk satisfies the process model in
k,W LS and x
(2.1). Assume that the system parameters are known. Let x k,KF be the state
estimates obtained at time-step k with the WLS and KF algorithms, respectively. Then,
denoting E as the expected value operator,
     
k,W LS 2 = E xk − x
E xk − x k,KF 2 + E 
xk,W LS − xk,KF 2 (2.3)

Proof. First of all, by standard KF theory, e.g., [90], the estimation of the non-
observable state is equal to its conditional expectation, given the sequence of measure-
ments, i.e.,

k,KF = E[xk |Fk ]

x (2.4)

where Fk the σ-field generated by all measurements up to and including time-step k.

Second, consider the Hilbert space of random vectors (with values in Rn ) equipped with

the inner product V, W
H  E[ ni=1 Vi Wi ]. It is requested to show that the random
vector xk − x k,KF is orthogonal,1 in the sense of this Hilbert space, to all random
vectors W that are Fk -measurable (i.e., that are a function of the measurements up to
time k; in this context, the initial conditions of the estimation algorithms are assumed
to be known and non-random). Note that what needs to be shown is that:
   
E W, xk
= E W, x k,KF
(2.5)
n
with W, xk
= i=1 Wi xk,i .

To prove (2.5), observe that, for any (real-valued) random variable U that is measurable
with respect to Fk , it can be written that E[U |Fk ] = U and further [90]:

E[U xk,i |Fk ] = U E[xk,i |Fk ] = U x

k,KF,i

Note that such a U can be any non-linear real-valued function of (z1 , . . . , zk ). Take
expectations on both sides and use the fact that the expectation of the conditional
expectation is the same as the original expectation (law of total expectation [90]) and
obtain

E[U xk,i ] = E[U x

k,KF,i ] (2.6)

Consider now any Fk -measurable random vector W, apply (2.6) to U = Wi for all
1
We remind here that when two vectors are orthogonal, their inner product is equal to zero.

31
Chapter 2. Adaptive Kalman filtering for the on-line estimation of the
process-model uncertainties

coordinates i and sum over i; then it becomes

n
n

 
E Wi xk,i = E k,KF,i
Wi x (2.7)
i=1 i=1

which shows (2.5) as required and it means that

n


E k,i − x
Wi ( x k,KF
] = 0
k,KF,i ) = E[ W, xk − x (2.8)
i=1

k,KF are orthogonal.

namely W and xk − x

Now observe that both x k,W LS and xk,KF are Fk -measurable because they are de-
rived from the measurements. Therefore, the previous result can be applied to (2.3)
then follows from Pythagoras’s equality. (QED)

Final remark: The theorem applies as long as the process model in (2.1) holds. This
explains why it is important to verify the adequacy of the process model.

2.4 New heuristic method for the on-line assessment of Q

The content of this section is based on [71] and [72]. The original contributions are the
following:

1. The formulation of a new heuristic method for the on-line assessment of Q [71],
which is conceived to be deployable in real-time embedded hardware;

2. A comparative assessment of the performance (accuracy and computational

time) of the DKF and the LWLS estimators [72] via numerical simulations on a
transmission network (New England 39-bus system) and a distribution feeder
(IEEE 123-bus test feeder);

3. The numerical proof of Theorem 1 [72];

4. The numerical validation of the correctness of the persistent process-model

when the process-noise covariance matrix is assessed via the heuristic method of
point 1 [72].

2.4.1 Analytical formulation

The method described here below assesses the diagonal elements of Q by using a
moving window composed of the previous N estimated states. The estimated value of

Q is denoted by Q.

32
 2.4. New heuristic method for the on-line assessment of Q

The idea behind the formulation of this method is the following. In quasi-static con-
ditions the sample variance of the last N state estimates aims at representing the
process-model uncertainty, although the latter tends to be overestimated as we will
show in Section 2.5.2. Instead, when the state changes monotonically (increasing or
decreasing), the diagonal values of Q increase and the method is able to rapidly (but
 Hence, the method is able to react to quick variations
not instantaneously) increase Q.
of the system state.

Note that the proposed method is based on the fact that Q is not supposed to change
dramatically from one time-step to the other; therefore, Q should be updated more
frequently than the system-state dynamics that we want to track. Indeed, we make
reference to PMU-based state estimators characterized by a refresh rate of 50 estimates-
per-second in order to accurately track most of the grid-state changes (except for large
steps, as illustrated in Section 2.5).

The analytical formulation of the proposed method is given here below. At time-step k
and denoting the state estimate with x, the procedure to estimate Qk consists in the
following steps:

• The ith element of a vector g ∈ Rn is computed as the sample variance of the ith
elements of the last N estimated states:

gi = var[ k−N |k−N,i ]

xk−1|k−1,i , . . . , x (2.9)

• Then, the elements of g constitute the diagonal of the estimated Qk :

 k = diag(g).
Q (2.10)

Matrix Q is diagonal and, in general, the diagonal elements are different from each
 Parameter N can be determined by
other. The larger N , the larger the entries of Q.
pre-tuning as we do in Section 2.4.2.

This method is simple and straightforward to implement. In Section 2.4.2 we show

that it also provides a good estimation accuracy although, being heuristic, it remains
sub-optimal. Additionally, it is not specific for power-system SE and can be applied
also to other fields.

2.4.2 Simulations and results

In this section, we evaluate the SE accuracy via the procedure described here below:

1. The network true state and the true values of the measurements are computed
each 20 ms via a load-flow calculation;

33
Chapter 2. Adaptive Kalman filtering for the on-line estimation of the
process-model uncertainties

2. The measurements used in SE are generated by adding Gaussian-white random

noise to the true quantities computed in step #1 (we assume to have no systematic
errors as stated at the beginning of this chapter);

3. The estimated state is computed by using the network model, the measurements
and, in case of dynamic estimators, also a process model;

4. The SE accuracy is obtained by comparing the estimated state with the true state
computed in step #1.

Further details on the simulation procedure are given in the following sections that
present the testing results for a transmission network and a distribution feeder. It is
important to mention that the results shown in this section are obtained by removing
the first 500 time-steps from the original simulation. The reason is that the KF initial
state is usually set to an arbitrary value (e.g., a flat-start) that might be far away from
the true state. Therefore, the KF is characterized by an initial phase in which the state
estimate approaches the true state. In this initial stage, we set a constant and high
value of Q in order to accelerate the approach. After N time-steps, the value of Q used
by the DKF is assessed by the heuristic method described in Section 2.4.1. After 500
time-steps, the effects of this initial phase are negligible as we use values of N that are
much lower than 500.

Transmission network

We consider the New England 39-bus system; as the network is balanced, the direct
sequence is estimated. A schematic of the network and the measurement configuration
are shown in Fig. 2.1. The network data can be found in [91, Appendix A]. We assume
that bus #31 is the connection point of the system to an external network so that the
voltage at bus #31 is not fixed. The equivalent circuit of the external network is modeled
with a fixed voltage source in series with a short-circuit impedance, which are called
SB and Zsc in Fig. 2.1, respectively. The voltage source SB represents the slack-bus
in the load-flow calculation.The value of Zsc is computed by assuming a rated voltage
of 380 kV, a short-circuit power of 50 GVA. The base power is 100 MVA. The network
is assumed to be in quasi-static conditions, which means that the network state is
moving slowly and no sudden changes are present.The active powers and the voltage
magnitudes at the generator buses are fixed to the values specified in [91, Appendix A].
In order to reproduce realistic time-series, the load active/reactive powers come from
real power measurements provided at 50 measurements-per-second by PMUs installed
in the 125-kV sub-transmission network of Lausanne, Switzerland. The values of these
real power measurements are adapted to the values specified in [91, Appendix A]. As an
example, Fig. 2.2 shows the time-evolution of the active and reactive powers absorbed
by the load at bus #4. Note that the simulation time is 30 s that corresponds to T = 1500
time-steps, because the PMU reporting-rate is 50 measurements-per-second.

34
 2.4. New heuristic method for the on-line assessment of Q

Bus Zero-injection bus Transformer V Voltage meas.

Load Transmission line G Generator I Current meas.

G8
37
G10
30 26 28 29
25 V
I V I V V I
I
2 27 38
I V
G9
1 24
18 17 V
I
I V
3 16 G6
39 35
V V I
I I
15
G1 V
I 21 22
4 14 I V

I V

5 12 I V
19
6 23
11 13 V I
7
10 20
36
8 V I
V
I G7
V I 31 I 32 34 V I 33
9 V
Zsc
SB G3 G5 G4

Figure 2.1 – Network topology of the New England 39-bus system together with the
adopted measurement configuration.
Q absorbed at Bus 4 [MVAR]

500.2 229.2
P absorbed at Bus 4 [MW]

500 229.1
499.8 229
499.6 228.9
499.4 228.8
499.2 228.7
499 228.6
498.8 228.5
0 10 20 30 0 10 20 30
time [s] time [s]

Figure 2.2 – Time evolution of the active and reactive powers absorbed by the load at
bus #4.

35
Chapter 2. Adaptive Kalman filtering for the on-line estimation of the
process-model uncertainties

The PMU placement shown in Fig. 2.1 ensures the full-network observability, i.e.,
matrix H has full rank. There are 19 buses equipped with PMUs that measure the
nodal-voltage phasor and the current-injection phasor at the respective bus. We also
exploit the 12 zero-injection buses as virtual measurements that consist in null values
with small uncertainty. The total number of measurements is m = 100 and the state
dimension is n = 78 so that the redundancy level is about 1.3. This small redundancy
is chosen in order to have a quite large estimation uncertainty on the nodal voltages
that are not directly measured. As a consequence, the SE accuracy improvement due
to the DKF noise filtering is more pronounced and the DKF results are significantly
better than the LWLS ones. As the redundancy level increases, the difference between
the DKF and LWLS results becomes smaller. A comparison between the DKF and
LWLS accuracies for different measurement types and redundancy levels is provided
in Tab. 2.1.

The measurement noise is assumed to be a Gaussian-white sequence where the covari-

ance matrix R is diagonal, time-invariant and known. These hypotheses are justified
by the following considerations:

• the noise mean-value is considered to be zero as we assume to compensate

the systematic errors of the sensors by using their calibration certificates, as
explained in Section 1.1.2;

• R is considered to be diagonal, because the measurements are assumed to be

uncorrelated for the reasons presented in Section 1.1.2;

• R is approximated as a time-invariant matrix, because the random-noise charac-

teristics do not change significantly when the power system is in normal operat-
ing conditions2 .

The standard deviations of the real and imaginary parts of voltage and current measure-
ments are assumed to be 0.001 pu. We also performed tests where R was time-varying
as the noise standard-deviations were calculated as a percentage of the measured
values; however, the SE results were similar to those presented in what follows.

The SE numerical results for the transmission network case are presented and dis-
cussed in the following. First of all, we perform a tuning of parameter N that is used
to assess Q (see Section 2.4.1). We run the same simulation composed of T = 1500
time-steps with different values of N and we compute the root mean square error

2
This assumption may not be valid when some particular event occurs, such as a fault, where the
voltage and current magnitudes can vary dramatically. Further research is needed to examine these cases.

36
 2.4. New heuristic method for the on-line assessment of Q

-4
×10
2.6

2.4

RMSE [pu]
2.2

1.8

1.6
20 40 60 80 100
N

Figure 2.3 – RMSE of the DKF as a function of the parameter N .

(RMSE) of the DKF estimates as follows:



 1  T n

k,i − xk,i ) .
(xest
RMSE = true 2 (2.11)
T +n
k=1 i=1

Fig. 2.3 shows that N = 30 provides the best accuracy for this specific case; this value
is used for the simulations in this section.

Assuming to have the perfect knowledge of the measurement model (i.e., of H and
R), the KF optimality depends on how accurate the process model is. More precisely,
we need to verify that the covariance properties of the model do hold. To this end,
we define the ith element of the vector of the normalized state-estimate residuals at
time-step k as

k|k,i − x
x k−1|k−1,i
N
rx,i =  (2.12)
Qk−1,ii

and we check whether these residuals are uncorrelated. Fig. 2.4 shows the time
evolution of the residuals and the sample auto correlation functions (ACFs). For
brevity, we show only the plots related to the real part of the voltage at bus #37 where
the largest estimation error is observed, but similar results are obtained for the other
state variables. We observe the time evolution of the residuals to check whether the
stationarity assumption is satisfied, which is the case. A more formal verification is
√
to analyze the sample ACFs. The latter are computed by considering the first ∼ T
lags, where we recall that T = 1 500 is the number of simulation time-steps. If the
√
residuals are uncorrelated, the ACFs should be within the noise margins ±1, 96/ T
with 95% of probability [92]. The noise margins are the two straight lines in the ACF
plot of Fig. 2.4. It can be seen that the above-mentioned condition is fulfilled; therefore,
the normalized residuals of the state estimates are uncorrelated and the DKF process
model can be considered to be accurate. Note that the sample ACF is a statistically
distributed quantity. The fact that, in few cases, the ACFs are slightly beyond the

37
 Chapter 2. Adaptive Kalman filtering for the on-line estimation of the
process-model uncertainties

N
x
of Vre at Bus 37
Sample ACF of r
4 0.2
at Bus 37
of Vre

2
0 0
rN

-2
x

-4 -0.2
0 5 10 15 20 25 30 1 5 10 15 20 25 30 35 39
time [s] Lag

Figure 2.4 – Normalized residuals of the estimates of the real part of the voltage at bus
#37. Left plot: time evolution of the residuals. Right plot: sample ACFs of the residuals.

×10 -3
3
2.5 LWLS
DKF
RMSE - V [pu]

2
1.5
1
0.5
0
1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39
Bus
-3
×10
3
2.5 LWLS
DKF
RMSE - δ [rad]

2
1.5
1
0.5
0
1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39
Bus

Figure 2.5 – RMSEs of the DKF and LWLS at each bus.

noise margins does not violate the validity of the result and the numerical proof of the
statistical correctness of the process model.

Fig. 2.5 shows the RMSE of the voltage magnitude V and phase-angle δ at each bus;
it can be seen that the DKF errors are, on average, significantly smaller than those of
the LWLS. Fig. 2.6 shows the RMSE of the voltage magnitude and phase-angle at each
time-step. The DKF accuracy is better during the entire simulation. Fig. 2.7 shows
the time evolution of the voltage magnitude and phase-angle at bus #37. This figure
reflects the results presented in Figs. 2.5 and 2.6: the DKF is able to effectively filter
the measurement noise and to closely track the true state, whereas the LWLS state
estimates are characterized by a much larger variance.

We also make a comparison with a DKF that uses a matrix Q that is constant in time
and is diagonal with all the entries equal to a value Qd (called DKFQ sampling ). In Fig. 2.8,
we plot the RMSE of DKFQ sampling as a function of Qd = 1 ÷ 10−16 . We also display the

38
 2.4. New heuristic method for the on-line assessment of Q

×10 -3 ×10 -3
2.5 2.5
LWLS LWLS
2 2

RMSE - δ [rad]
DKF DKF
RMSE - V [pu]

1.5 1.5

1 1

0.5 0.5

0 0
0 10 20 30 0 10 20 30
time [s] time [s]

Figure 2.6 – Time evolution of the RMSEs of the DKF and LWLS.

1.04 0.248
LWLS
DKF 0.246
1.035 True 0.244

δ [rad] at Bus 37
V [pu] at Bus 37

0.242
1.03
0.24
0.238
1.025
0.236

1.02 0.234
0.232
1.015 0.23
0 5 10 15 20 25 30 0 5 10 15 20 25 30
time [s] time [s]

Figure 2.7 – Time evolution of the voltage magnitude V and phase-angle δ at bus #37.
The true value and the values estimated by DKF and LWLS are displayed.

RMSEs of the LWLS and of the DKF that uses a value of Q assessed continuously with
the heuristic method of Section 2.4.1 (called DKFQ assessed ), which are straight lines as
they are independent of Qd . The curve of the DKFQ sampling approaches the LWLS for
large values of Qd and has a minimum in Qd = 10−9 . The RMSE of the DKFQ assessed is
close to this minimum. Therefore, in this specific case, the DKFQ assessed is, on average,
as accurate as a DKF that uses the best constant value of Q obtained from an extensive
search. The advantage of the DKFQ assessed consists in being able to adapt its process
model uncertainty to changing operating conditions as it can dynamically change the
value of Q.

Finally, we verify whether the quantitative conclusion of Theorem 1 in Section 2.3

numerically holds. Intuitively, this has to be the case in view of the results of Fig. 2.8.
Fig. 2.9 shows the left- and right-hand sides (LHS and RHS, respectively) of (2.3). The
LHS and RHS are close to each other, which means that the equality is in expectation.
The expectations are estimated by empirical averages; therefore, a small discrepancy
is expected. In Fig. 2.9, the contributions of the two terms of the RHS of (2.3) are also
shown; the contribution of the second one is predominant, which proves that the DKF
is applied correctly.

It is worth clarifying that the above-presented results refer to the measurement con-

39
Chapter 2. Adaptive Kalman filtering for the on-line estimation of the
process-model uncertainties

×10 -3
1.2
1 DKF - Q sampling
DKF - Q assessed
RMSE [pu]

0.8 LWLS
0.6
0.4
0.2
0
10 -16 10 -14 10 -12 10 -10 10 -8 10 -6 10 -4 10 -2 10 0
Qd

Figure 2.8 – RMSE of the DKFQ sampling , DKFQ assessed and LWLS as a function of Qd .

×10 -6
15
LHS
RHS
10 DKF
)
RHS (x LWLS DKF
)
5

1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39
bus

Figure 2.9 – Numerical validation of Theorem 1. We show the LHS vs. RHS of (2.3). The
separate contribution of the two terms of the RHS is also shown.

Table 2.1 – RMSEs of the DKF and LWLS for different measurement configurations.

Network RMSELW LS RMSEDKF RMSELW LS / RMSEDKF

Conf. 1 9.4 · 10−4 1.8 · 10−4 5.2
Conf. 2 1.9 · 10−4 8.0 · 10−5 2.4
Conf. 3 1.6 · 10−4 7.6 · 10−5 2.1
Conf. 4 1.5 · 10−4 7.2 · 10−5 2.1

40
 2.4. New heuristic method for the on-line assessment of Q

figuration depicted in Fig. 2.1. It is interesting to how the SE accuracy of the DKF
and LWLS estimators changes as a function of the chosen measurements. To this end,
Tab. 2.1 presents the RMSEs of the two state estimators for different measurement
configurations that are listed here below:

• Configuration 1: the one depicted in Fig. 2.1;

• Configuration 2: voltage and the current-injection measurements at all the buses

but zero-injection buses;

• Configuration 3: voltage measurements at all the buses and current-flow mea-

surements at the sending side of every line;

• Configuration 4: voltage measurements at all the buses and current-flow mea-

surements at the sending and receiving sides of every line.

As the redundancy level augments, the DKF keeps being more accurate than the LWLS,
but the ratio between the LWLS and DKF errors decreases. The advantage of using a
DKF for reducing the measurement noise is stronger for low redundancy levels.

Distribution feeder

In this section, most of the simulation conditions and results are the same as for the
transmission network presented in the previous section. Therefore, the reader can
make reference to the section on the transmission network for the definitions and
explanations of repeated items.

We employ the IEEE 123-bus test feeder [93]. The power absorptions/injections are
assumed to be different in the three phases and mutual coupling between phases is
present; therefore, SE is performed in the three-phase domain. A schematic of the
network and the measurement configuration are shown in 2.10. We assume that the
network has a rated voltage equal to 15 kV and the line parameters correspond to the un-
balanced configuration #602 specified in [93]. The value of the short-circuit impedance
Zsc is computed by assuming a short-circuit power of 300 MVA and a resistance-to-
reactance ratio Rsc /Xsc = 1/10. The base power is 10 MVA. A photovoltaic (PV) plant
and a mini-hydro plant are placed at buses #92 and #112, respectively, and inject only
active power into the grid. The active/reactive powers absorbed by loads as well as the
active power injected by generators come from real power measurements provided
at 50 measurements-per-second by PMUs installed in the 20-kV distribution network
of the Swiss Federal Institute of Technology of Lausanne (the french name is École
Polytechnique Fédérale de Lausanne – EPFL), Switzerland. Fig. 2.11 shows the time
evolution of the aggregated active and reactive powers absorbed by all the loads and
the active power injected by a PV and a mini-hydro plants. Note that the PV injection

41
Chapter 2. Adaptive Kalman filtering for the on-line estimation of the
process-model uncertainties
M PMU
Bus with load/generator
Zero-injection bus
Mini-Hydro plant
32 M 29 M 116
250 350
117 118
30 51 111 110 112 113 114
33 251 M
50 151 300 M M M
M 28 49
31
M 109 107
25 47 M
48 46
26 M M 108 451
27 45 106 104M
64M
44 43 65 103 119
23 M 450
105 102
M 63 M 100
42 41 M
24 M 66 M 101
21 99
40 71
22 98 M
M 39M 62 M 197 M 70
M 38
36 97 69
19 135 35 M
20 18 68
M M 75
160 67 M
37 74
60
73
14 57 M
11 58
59 72 M 85
M M 61 610 79
121
M9 78
10 53 54 M 77
2M 52 56 M
152 M 55 76
Zsc 7 8 13
94
80 M
84
SB M M
96 M 76
149 120
115 1 M 34 M 88 M 81
150 12 92 M 90 M
1 17
M
15 87 86 83
M 95 91 89 82 M
3 93
5 6
M 16M 195 PV plant
4

Figure 2.10 – Network topology of the IEEE 123-bus test feeder together with the
adopted measurement configuration. This topology is obtained by setting the switch
positions specified by the benchmark and is composed of 121 buses.

is characterized by quick variations that generate fast changes in the state variables as
we will observe later in this section.

SE exploits the information related to the 35 zero-injection buses as virtual mea-

surements and uses the real-time phasor measurements provided by 51 PMUs. The
locations of zero-injection buses and PMUs are given in Fig. 2.10 and ensure the
full-network observability (i.e., H has full rank). The total number of measurements
is m = 822 and the state dimension is n = 726, which results in a redundancy level
of about 1.2. The noise characteristics are the same given in the transmission net-
work section, but we decrease the standard deviation of the current measurements
from 0.001 pu to 0.0001 pu, because the base value of the current is quite large for this
network.

For the above-described 123-bus test feeder, the result of a pre-tuning of parameter
N shown in Fig. 2.12 is that N = 30 provides the best DKF accuracy; this value will be
used in the simulations presented in the following.

Fig. 2.13 shows the time evolution of the residuals defined in (2.12) and the sample
ACFs in each phase of bus #61 where the largest estimation error is observed (similar
results are obtained for the other state variables). We observe that the residuals are
uncorrelated and the correctness of the DKF process model is verified.

42
 2.4. New heuristic method for the on-line assessment of Q

Phase a Phase b Phase c

6400 1200

Aggregated Q [kVAR]
Aggregated P [kW]
absorbed by loads

absorbed by loads
6300 1100
6200 1000
6100 900
6000 800
5900 700
0 10 20 30 0 10 20 30
time [s] time [s]
(a) Aggreagated active and reactive powers absorbed by all the loads.

Phase a Phase b Phase c

200 135
photovoltaic plant

mini-hydro plant
P injected [kW]

P injected [kW]
150 130
125
100
120
50 115
0 110
0 10 20 30 0 10 20 30
time [s] time [s]
(b) Active power injected by the PV and mini-hydro plants.

Figure 2.11 – Time evolution of the absorbed/injected powers of loads/generators.

-4
×10
4.5

4
RMSE [pu]

3.5

2.5

2
20 40 60 80 100
N

Figure 2.12 – RMSE of the DKF as a function of the parameter N .

43
Chapter 2. Adaptive Kalman filtering for the on-line estimation of the
process-model uncertainties
4 0.2
Phase a Phase a
2
0 0
-2

Sample ACF of r x of Vre at Bus 61

-4 -0.2
0 5 10 15 20 25 30 1 5 10 15 20 25 30 35 39
time [s] Lag
at Bus 61

4 0.2
Phase b Phase b
2
0 0

N
re

-2
r N of V

-4 -0.2
0 5 10 15 20 25 30 1 5 10 15 20 25 30 35 39
x

time [s] Lag

4 0.2
Phase c Phase c
2
0 0
-2
-4 -0.2
0 5 10 15 20 25 30 1 5 10 15 20 25 30 35 39
time [s] Lag

Figure 2.13 – Normalized residuals of the estimates of the real part of the voltage at bus
#61. Left plot: time evolution of the residuals. Right plot: sample ACFs of the residuals.

Fig. 2.14 shows that, on average, the DKF errors are much smaller than those of the
LWLS. In Fig. 2.15, it is interesting to observe that the DKF-RMSE time evolution
exhibits some peaks corresponding to the quick changes of the PV power injection (see
Fig. 2.11b), which result in fast changes of the system state as depicted in Fig. 2.16.
However, the DKF is able to recover quite rapidly thanks to the inflation of Q operated
by the heuristic method proposed in Section 2.4.1. Fig. 2.17 shows the time evolution
of the element of Q corresponding to the real part of the voltage at bus #92 in phase a.
Comparing Figs. 2.16 and 2.17, it is evident that this element of Q is inflated when fast
changes of the corresponding state variable occur; then, during periods when the state
variable varies smoothly, the corresponding element of Q decreases as expected.

As for the transmission network case, Fig. 2.18 shows that the DKFQ assessed is, on
average, as accurate as a DKF that uses the best constant value of Q obtained from
an extensive search. However, we recall that the advantage of assessing Q at each
time-step is that the DKF can closely track the system state in different operating
conditions, as shown in Figs. 2.16 and 2.17.

Finally, Fig. 2.19 verifies that (2.3) related to Theorem 1 in Section 2.3 numerically
holds. For lack of space, only the first 40 buses are shown, but the results are similar for
the remaining set of buses.

44
 2.4. New heuristic method for the on-line assessment of Q

-3
×10
4
LWLS
DKF
RMSE - V [pu]

0
1 6 11 16 21 26 31 36 41 46 51 56 61 66 71 76 81 86 91 96 101 106 111 116 121
Bus
-3
×10
4
LWLS
DKF
RMSE - δ [rad]

0
1 6 11 16 21 26 31 36 41 46 51 56 61 66 71 76 81 86 91 96 101 106 111 116 121
Bus

Figure 2.14 – RMSEs of the DKF and LWLS at each bus. The maximum RMSE among
the three phases is shown.

×10 -3 ×10 -3
2 2
LWLS LWLS
DKF DKF
RMSE - δ [rad]
RMSE - V [pu]

1.5 1.5

1 1

0.5 0.5

0 0
0 10 20 30 0 10 20 30
time [s] time [s]

Figure 2.15 – Time evolution of the RMSEs of the DKF and LWLS.
0.959 -0.086

0.958 -0.087
δ [rad] at Bus 92 in Phase a
V [pu] at Bus 92 in Phase a

0.957 -0.088
-0.089
0.956
-0.09
0.955
-0.091
0.954
-0.092
0.953 -0.093
LWLS LWLS
0.952 DKF -0.094 DKF
True True
0.951 -0.095
0 5 10 15 20 25 30 0 5 10 15 20 25 30
time [s] time [s]

Figure 2.16 – Time evolution of the voltage magnitude V and phase-angle δ at bus #92
in phase a. The true value and the values estimated by DKF and LWLS are displayed.

Table 2.2 – Average computational time in milliseconds.

Network n m LWLS DKF

39-bus network 78 100 0.29 ms 0.39 ms
123-bus feeder 726 822 42 ms 73 ms

45
Chapter 2. Adaptive Kalman filtering for the on-line estimation of the
process-model uncertainties

×10 -8
Q diagonal entry of Vre 8
at Bus 92 in Phase a
6

0
0 5 10 15 20 25 30
time [s]

Figure 2.17 – Time evolution of the element of Q corresponding to the real part of the
voltage at bus #92 in phase a.

×10 -3
1.2 DKF - Q sampling
DKF - Q assessed
1
RMSE [pu]

LWLS
0.8
0.6
0.4
0.2
0
10 -16 10 -14 10 -12 10 -10 10 -8 10 -6 10 -4 10 -2 10 0
Qd

Figure 2.18 – RMSE of the DKFQ sampling , DKFQ assessed and LWLS as a function of Qd .

×10 -6
20
LHS
RHS
15 )
DKF
RHS (xLWLS DKF
)
10

ab c ab c ab c ab c ab c ab c ab c ab c ab c ab c ab c ab c ab c ab c ab c ab c ab c ab c ab c ab c ab c ab c ab c ab c ab c ab c ab c ab c ab c ab c ab c ab c ab c ab c ab c ab c ab c ab c ab c ab c
-5
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40
bus, phase

Figure 2.19 – Numerical validation of Theorem 1 for buses #1–40. We show the LHS vs.
RHS of (2.3). The separate contribution of the two terms of the RHS is also shown.

46
 2.5. New prediction-error covariance estimation method (PECE) for step-varying
processes

Computational-time

The SE computational-times for the transmission network and the distribution feeder
are inferred by using MATLAB 2014b that runs in an Apple MacBook Pro with a 2.5-GHz
CPU and 16-GB RAM. In both case studies, we assumed R to be time-invariant so
that the gain matrix G of the LWLS estimator is also time-invariant. Therefore, G
could be inverted only once and, at every time-step, the computation of the LWLS
solution reduces to a simple matrix multiplication. However, for the computational-
time assessment we considered the general case in which R is time-variant and we
solved the LWLS NE at every time-step.

Tab. 2.2 shows the computational times in milliseconds of the LWLS and DKF estima-
tors for the New England 39-bus system and 123-bus test feeder. It can be seen that the
LWLS is faster than the DKF, but both computational times are below 100 ms, which
is compatible with real-time applications. In order to achieve a SE refresh-rate of 50
estimates-per-second, the state estimators can be pipelined. It is worth mentioning
that if, as mentioned before, the LWLS gain matrix were computed only once, the LWLS
computational time would be order of magnitudes lower. The time employed to esti-
mate Q via the heuristic method is always much smaller than the computational time
of the DKF equations; therefore, the computational burden added by the proposed Q
assessment method is negligible.

2.5 New prediction-error covariance estimation method (PECE)

for step-varying processes
The content of this section is based on [94].

The AKF presented in the previous Section 2.4 was able to provide satisfactory results
when the system is in quasi-static conditions. Similarly, the methods presented in the
literature review in Section 2.2 are conceived to estimate the stochastic parameters of
time-invariant or slow-varying processes. However, these methods have some delay
in tracking the state during step variations, which are common in power systems
as they are associated, for instance, with the connection/disconnection of different
components (mainly loads, generators and transmission lines). Still needed is a KF able
to filter the measurement noise and to track sudden step-variations with minimum
delay. In this section, we propose an AKF that addresses this problem by using a
new method for inferring Pk|k−1 ; this method is called prediction-error covariance
estimation (PECE). Nevertheless, our proposed AKF is not specific for power-system SE
only, actually it can be applied to every process where Assumptions #1–3 introduced in
Section 2.5.1 hold.

The analytical formulation of the PECE method is given in Section 2.5.1. In Sections

47
Chapter 2. Adaptive Kalman filtering for the on-line estimation of the
process-model uncertainties

2.5.2 and 2.5.3 we provide a performance assessment for the case of ideal processes
and for power-system SE, respectively. The PECE method is compared with the method
presented in [89], with our heuristic method presented in Section 2.4 and with the
LWLS.

Let us consider a system described by the persistent process-model in (2.1). The

process noise should be able to account for step variations. To this end, we enable the
covariance matrix Qk to be time-varying and unknown. In contrast, the measurement
model is assumed to be linear, known and time-invariant. Hence, it can be defined
as in (2.2) where H and R are time-invariant. We also assume that the system is fully
observable, i.e., H has full rank. In the following we provide the DKF equations given
in Section 1.4.3 where H and R lack of the subscript k as they are assumed to be
time-invariant:

Prediction
k|k−1 = x
x k−1|k−1 (2.13)
Pk|k−1 = Pk−1|k−1 + Qk (2.14)

Measurement update
Lk = Pk|k−1 HT (HPk|k−1 HT + R)−1 (2.15)
k|k = x
x k|k−1 + Lk (zk − H
xk|k−1 ) (2.16)
Pk|k = (I − Lk H)Pk|k−1 (2.17)

We also recall the innovation and its covariance matrix under the aforementioned
assumptions:

yk = zk − H
xk|k−1 (2.18)

Tk = HPk|k−1 HT + R . (2.19)

As it is known, the KF estimate is optimal if we use the optimal Kalman gain L that
depends on matrices Pk|k−1 , H and R. The matrix Pk|k−1 contains the uncertainty
of the process noise Q. Assuming the knowledge of H, the literature dealing with KF
parameters estimation proposes several methods to infer the correct R and Q (see
Section 2.2), but most of these methods are suitable for time-invariant or slow-varying
systems. However, some systems are characterized by sudden and unpredictable
state-steps. In these cases, given a certain process model, the Kalman gain has to be
updated.

Our goal is to develop an AKF that filters effectively the measurement noise during
quasi-static conditions and is also able to track step-variations of the system state. We
consider a process that can be modeled as (2.1) with constant or slow-varying Q and

48
 2.5. New prediction-error covariance estimation method (PECE) for step-varying
processes

that exhibits state step-variations that violate the process model. Assuming H and R
are known and time-invariant, the goal is to keep tracking the state in spite of these
violations. We exploit the knowledge of the sample covariance matrix of the past and
current innovations (1.31) in order to estimate the prediction-error covariance matrix
Pk|k−1 , from which follows the name PECE method. Hence, we automatically update
the Kalman gain before the measurement update. The PECE method consists in a
constrained convex optimization problem based on maximum-likelihood estimation
that ensures the positive semi-definiteness of Pk|k−1 .

2.5.1 Analytical formulation

The PECE method applies to cases where the following assumptions hold:

Assumption 1. The process model is the linear and time-variant persistent model
described by (2.1). Specifically, we consider the general case in which the process-noise
covariance matrix Qk is unknown and changes as a function of time. Furthermore, we
are interested in the case where the system state is characterized by step variations
that violate the process model.

Assumption 2. The measurement model is linear, known and time-invariant. We

assume to have the perfect knowledge of both the measurement matrix H and the
measurement-noise covariance matrix R.

Assumption 3. The system is assumed to be fully observable by using a number of

measurements equal to or higher than the number of states: m ≥ n and matrix H has
full rank.

If Qk were known, the prediction-error covariance matrix Pk|k−1 would be computed

iteratively from (2.14) and all the other quantities of interest would derive from (2.15)–
(2.17). However, this is not possible because of Assumption #1. The objective of PECE
is to provide an estimate of Pk|k−1 from the measurements, without direct estimation
of Qk . The procedure is given in the following.

Given Assumption #1, the KF prediction equation is given by (2.13). In order to quickly
react to step-variations of the system state, the PECE method takes advantage of the
innovation y that contains the new information brought by the measurements at each
time-step. In what follows, the PECE method’s algorithm is presented in four steps.

k|k−1 by
• Step 1. At time-step k, after the computation of the predicted state x
means of (2.13) and when the new measurement set zk is available, the innova-
tion yk is calculated by using (2.18).

• Step 2. In static conditions, the innovations represent a white Gaussian sequence

49
Chapter 2. Adaptive Kalman filtering for the on-line estimation of the
process-model uncertainties

with covariance matrix T∞ , defined as

T∞ = HP∞ HT + R. (2.20)

where P∞ is the value of Pk|k−1 as k → ∞. An approximation of T∞ is the

sample innovation-covariance matrix C k , which is calculated at time-step k by
considering a moving-window composed of N time-steps as

 k = cov(yk , yk−1 , ..., yk−N +1 )

C (2.21)

As it is known, for stationary processes the sample covariance tends to the true
one as N increases. In non-stationary conditions, the true innovation-covariance
matrix varies at each time-step and we denote it as Tk . Thus, given the knowledge
of the matrices H and R (Assumption #2), we can re-write (2.20) as (2.19), recalled
here below:

Tk = HPk|k−1 HT + R. (2.22)

Given the KF prediction equation (2.13), when the state has a sudden change, Tk
changes as well. The sample matrix C  k follows the variations of the true matrix
Tk , because the innovations incorporate the information of the measurements.
The PECE method exploits this feature in order to quickly react to the state
changes, as explained in the following steps.

• Step 3. Estimate Pk|k−1 from the innovation samples, using maximum likelihood
estimation, as described in Theorem 3 below. This provides the estimate P  k|k−1 ,
which is symmetric and positive semi-definite; using (2.15), this estimate is used
to calculate the Kalman gain L that is updated at every time-step. In quasi-static
conditions, L remains fairly constant. When a state step-variation occurs, the
KF prediction equation (2.13) is inaccurate, causing an increase of P  k|k−1 and
consequently of L. Therefore, the KF trusts the measurements more than the
predicted state, which is the right action to take in order to quickly react to the
state variation.

Estimation of Pk|k−1 . A natural method for estimating Pk|k−1 would consist in replac-
 k computed in Step
ing in (2.22) the matrix Tk by the sample innovation-covariance C
2 and solving for Pk|k−1 (note that matrices H and R are known, by Assumption #2).
However, this process, in general, will not produce a semi-definite matrix. A more
adequate and more generally applied method is a ML estimation that can guarantee
the positive semi-definiteness of Pk|k−1 .

The estimation procedure uses the following optimization problem:

   
min − log det(Σ) + trace(ΣE) (2.23)
Σ

50
 2.5. New prediction-error covariance estimation method (PECE) for step-varying
processes

subject to: Σ real symmetric and Σ  0

In − Σ  0.

In this optimization problem (i) the optimization variable is the n × n, real matrix Σ,
assumed to be symmetric, (ii) In is the identity matrix of size n, (iii) E is a known and
fixed matrix and (iv) the notation Σ  0 means that Σ is positive definite and the
notation In − Σ  0 means that In − Σ is positive semi-definite.
Theorem 2. The optimization problem (2.23) is a convex problem; it has one unique

optimal solution Σ.

The proof is given in Appendix A.2. Note that problem (2.23) is a MAXDET problem
for which there exists efficient software. Problem (2.23) is modeled and solved with
YALMIP employing the sdpt3 solver [95, 96].

We can now describe the estimation procedure. First perform a QR-decomposition of

1
the matrix R− 2 H and obtain
 
1 U
R− 2 H = V (2.24)
0m−n,n

where U is an upper triangular, real n × n matrix, 0m−n,n is the null rectangular matrix
of dimensions (m − n) × n and V is an orthogonal matrix of dimensions m × m. Note
that U is invertible because H and R are full rank.

 k be the square matrix of dimensions n × n made of the first n rows and

Second, let E
1
columns of VT R− 2 C  k R− 12 V:
 
 k = VT R− 12 C
E  k R− 12 V (1 : n, 1 : n) (2.25)

Theorem 3. The maximum likelihood estimate of Pk|k−1 based on the observation of

the innovations is equal to
 
 k|k−1 = U−1 Σ
P  −1 − In U−T (2.26)

 is the optimal solution of the problem (2.23) with E replaced by E

where Σ  k . The matrix

Pk|k−1 is symmetric and positive semi-definite.

The proof is given in Appendix A.3.

2.5.2 Simulations and results for ideal processes

In this section, we evaluate the accuracy of the PECE method by considering ideal
processes where the time evolution of the system state is controlled so that the true

51
Chapter 2. Adaptive Kalman filtering for the on-line estimation of the
process-model uncertainties

value of every parameter of the process is known. The purpose is to verify whether
the ability of the PECE method to track step-variations of the system state is not at the
expense of its accuracy when there are no steps3 . For this purpose, we first consider a
process without step-variations in the system state and then study a process where we
add a step to the state variables. For all the cases treated in this section, we assume R
to be diagonal and constant:

R = rIm (2.27)

where r is a scalar and Im ∈ Rm×m is the identity matrix.

The two case studies are described in detail here below:

• Base case: the process consists in a random walk as (2.1) with Q that is diagonal
and constant:

Q = qIn (2.28)

where q is a scalar and In ∈ Rn×n is the identity matrix. We have tested the two
possible conditions that can occur, i.e., q/r < 1 and q/r > 1.

• Base case plus steps: it consists in the same random walk processes of the base
case plus state step-variations of intentionally large amplitudes that violate the
process model.

It is worth observing that this section proves the effectiveness of the PECE method for
generic physical processes, as it can be applied to every process, as long as Assumptions
#1–3 of Section 2.5.1 hold.

The PECE method is compared with two other covariance-estimation methods:

• The method of Myers and Tapley [89], which will be called Myers henceforth in
this chapter. In [89], the assessment of both Q and R is discussed. Whereas, we
assume R is known, so that only Q has to be estimated. The formulation of the
Myers method is recalled in Appendix A.4;

• Our heuristic method proposed in Section 2.4, which will be called heuristic
henceforth in this chapter.

Three DKFs that use the three considered methods (PECE, Myers and heuristic) are run
3
Indeed, a simple way to track state step-variations would be to always overestimate the value of Q, as
the heuristic method proposed in Section 2.4 does (see Fig. 2.20). Obviously, the filtering effectiveness is
compromised when no steps are present. Another way would be to inflate Q only when the step occurs,
as it is proposed in [37]. The problem is what value of Q should be set before and after the step.

52
 2.5. New prediction-error covariance estimation method (PECE) for step-varying
processes

in parallel. We assume that the DKFs have the knowledge of R, but they do not know
Q.

The PECE, Myers and heuristic methods use a moving-window composed of N time-
steps in order to assess the respective covariance matrices. It is important to recall that
the PECE method uses N innovations to infer Pk|k−1 , whereas the Myers and heuristic
methods use N state estimates to infer Q. In Section 2.5.2, we assess the influence
of N on the SE accuracy. The PECE method uses a moving-window composed of
N = 5 000 time-steps, as it has been found to be an effective trade-off between filtering
performance and fast tracking of state step-variations. Also the moving-window length
employed by Myers method is assumed to be N = 5 000 time-steps. We choose N = 30
for heuristic method, because it does not need a large moving-window length to
estimate Q (see Section 2.4.2).

Base case

Let us consider a multi-dimensional state x ∈ Rn , with n = 6. The process and

measurement equations are

xk = xk−1 + wk−1
zk = x k + v k
w ∼ N(0, Q)
v ∼ N(0, R)

in which R and Q are defined as in (2.27) and (2.28), respectively. First, we consider
the case where q/r < 1, with r = 10−7 and q = 10−10 . As a consequence, the true
value of the KF prediction-error covariance matrix is P∞ = 3.21 · 10−9 In . Note that the
important parameter is the ratio q/r. Indeed, different values of r and q resulting in the
same ratio lead to similar results.

The estimated state x and the matrix Q, which are used in the three DKFs, are initialized
by using the initial conditions listed here below:

1. The state vector is initialized to x0 = 1 (where 1 is a vector of ones), which is

different from the true state.

2. For the first 15 000 time-steps, we use Q = q0 In . To help the convergence of the
initial state x0 towards the true one, we set a high value of q0 = 10−5 .

Fig. 2.20 shows the time evolution of the third diagonal element of Pk|k−1 estimated by
the three methods and its true value (the other diagonal elements exhibit an analogous
behavior). The time evolution of the RMSEs are shown in Fig. 2.21 for these three DKFs

53
Chapter 2. Adaptive Kalman filtering for the on-line estimation of the
process-model uncertainties
-6
10
True DKF - heur DKF - Myers DKF - PECE

-7

P k|k-1 (3,3)
10

10 -8

10 -9
1.5 2 2.5 3 3.5 4 4.5 5 5.5 6
time-step ×10 4

Figure 2.20 – Base case, q = 10−10 . Time evolution of the third diagonal element of
Pk|k−1 (the other diagonal elements exhibit an analogous behavior): true and estimated
values.

plus a DKF that uses the exact value of Q denoted as DKFQ exact . At time-step k, the
estimation error vector ek is defined as the difference between the estimated and the
true state:

est
ek = x k − xk
true
. (2.29)

Note that the RMSE accounts for the errors of all the six state variables. Until the
15 000th time-step, the three DKFs use a large value of Q equal to 10−5 In , so that
the estimation errors are the same for every DKF and remain large. Afterwards, the
process covariances start to be assessed and the errors of the three DKFs decrease,
thus reflecting the behavior of the Pk|k−1 estimates. As visible in Fig. 2.20, the Pk|k−1
inferred by the PECE method converges to the true value of Pk|k−1 in about 10 000 time-
steps, which corresponds to twice the moving-window length N = 5 000. Afterwards,
the Pk|k−1 estimates oscillate around the true value of Pk|k−1 (see the enlarged part of
Fig. 2.21) and the estimation errors become similar to those of the DKF that uses the
exact value of Q. The convergence of the Myers method is much slower and the Pk|k−1
estimates takes a longer time to converge to the true values, i.e., hundreds of thousands
of time-steps. However, the estimates are characterized by smaller variations compared
to the PECE method (see Fig. 2.20). The heuristic method’s convergence phase lasts
few time-steps, because it uses a small moving-window length N = 30. Then, it tends
always to overestimate the value of Pk|k−1 . As a consequence, the measurement noise
is not filtered effectively, which leads to significant estimation errors.

Let us now consider the case in which q/r > 1, with r = 10−7 and q = 10−6 . The
corresponding true value of P∞ is 1.09 · 10−6 In . Fig. 2.22 shows the time evolution of
the third diagonal element of Pk|k−1 for this case (the other diagonal elements exhibit
an analogous behavior). The Myers and PECE methods estimate its precise true value,
whereas the heuristic method provides again larger variances. It can be noted that the
PECE method estimates immediately the correct value of Pk|k−1 , whereas the Myers
method takes 5 000 time-steps to approach it. However, the estimation errors of these
three DKFs are almost equal to those of the DKF that uses the exact value of Q. We

54
 2.5. New prediction-error covariance estimation method (PECE) for step-varying
processes
×10 -4
8
DKF - Q exact DKF - heur DKF - Myers DKF - PECE

7 ×10 -4
3
2.5

6 2

RMSE
1.5
1
5 0.5
0
5.7 5.71 5.72 5.73 5.74 5.75 5.76 5.77 5.78 5.79 5.8
RMSE
time-step ×10 4
4

0
1 1.5 2 2.5 3 3.5 4 4.5 5 5.5 6
time-step ×10 4

Figure 2.21 – Base case, q = 10−10 . Time evolution of the RMSEs.

do not show the RMSEs for this case, because they are indistinguishable and oscillate
between 0.2 · 10−3 and 1.6 · 10−3 . These results are caused by the predominance of the
process noise, making simple its estimation from the samples. This case is not of high
interest, because the process model has a limited influence on the KF solution. Indeed,
the simple processing of the measurements leads to state estimates very close to the
optimal ones.

Base case plus steps

Let us consider the same process of the Base Case, with r = 10−7 , q = 10−10 , and x0 = 1.
For this case, we wait until the convergence phase is finished (time instant 0), and
1 000 time-steps later we simulate a state step-variation of amplitude 10−2 , which is
significantly larger than the process-noise standard deviation. By the time the state
change occurs, we expect Pk|k−1 to increase, because the assumed persistent model in
(2.1) is no longer accurate. Note that, as the step is applied to every state variable, they
become correlated. In order to show the effect of the estimated off-diagonal elements
of Pk|k−1 on the state estimates, we add a further DKF, called PECEdiag . The latter
uses the PECE method in which the variable Σ of the optimization problem (2.23) is
diagonal.

Fig. 2.23 shows the time evolution of the third diagonal element of Pk|k−1 estimated by
the four DKFs when the step occurs, and Fig. 2.24 shows a further zoom of Fig. 2.23
(the other diagonal elements exhibit an analogous behavior). The trend of the off-
diagonal elements of Pk|k−1 is highly dependent on the specific simulation parameters.
Their effect is evident by comparing the behavior of the DKF estimates given by the

55
Chapter 2. Adaptive Kalman filtering for the on-line estimation of the
process-model uncertainties
-4
10
True DKF - heur DKF - Myers DKF - PECE

P k|k-1 (3,3)

10 -5

10 -6

1.5 2 2.5 3 3.5 4 4.5 5 5.5 6

time-step ×10 4

Figure 2.22 – Base case, q = 10−6 . Time evolution of the third diagonal element of
Pk|k−1 (the other diagonal elements exhibit an analogous behavior).

PECE and PECEdiag methods. The DKF that uses the Myers method has a remarkable
delay in inflating Pk|k−1 and consequently in tracking the system state during and
after the step, as is visible from the RMSEs in Fig. 2.25. This behavior is typical of the
covariance-estimation methods that infer Q by exploiting only the knowledge of the
past state estimates. It is known that Q influences the state estimates, and vice versa,
as a closed-loop, therefore both take several time-steps to increase. At the time of the
step, the heuristic method had smaller errors because it overestimates the value of Q,
as we have seen for the base case. Shortly after the step, the heuristic method inflates
Q rapidly, so that the estimation errors decrease in only a few time-steps. Unlike the
other two methods, the PECE method uses the past innovations and the innovation at
the current time-step that already contains the information brought by the upcoming
set of measurements. Then, it updates the value of Pk|k−1 before the measurement
update. Indeed, the PECE method is the only one able to increase Pk|k−1 exactly by
the time the step occurs (i.e, with no delay) as shown in Fig. 2.24. This characteristic
enables the PECE method to immediately react to state step-variations. Fig. 2.25 shows
that the estimation errors of the DKF that uses the PECE method are the smallest, both
exactly when the step occurs and shortly afterwards. In Fig. 2.25, it is also evident
that the DKF that uses the PECE method reacts faster than the DKF that uses the
PECEdiag method. Therefore, the estimated off-diagonal entries of P  k|k−1 improve the
state-tracking capability.

Finally, we consider a case in which q/r > 1 (r = 10−7 and q = 10−6 ) and a state
step-variation is present. The step-variation is of amplitude 1 and occurs at the 1 000th
time-step. Note that the step amplitude is larger than the case where q/r < 1, because
a step amplitude of 10−2 would be comparable to the process noise. Fig. 2.26 shows

56
 2.5. New prediction-error covariance estimation method (PECE) for step-varying
processes
-4
10
DKF - heur
10 -5 DKF - Myers
DKF - PECE

(3,3)
-6 DKF - PECE diag
10

k|k-1
-7
10

P
-8
10

10 -9
1000 1020 1040 1060 1080 1100
time-step

Figure 2.23 – Base case plus steps, q = 10−10 (a step occurs at the 1 000th time-step).
Time evolution of the third diagonal element of Pk|k−1 (the other diagonal elements
exhibit an analogous behavior).

-4
10
-5
DKF - heur
10 DKF - Myers
DKF - PECE
(3,3)

-6 DKF - PECE diag

10
k|k-1

10 -7
P

-8
10
-9
10
995 1000 1005 1010
time-step

Figure 2.24 – Base case plus steps, q = 10−10 (a step occurs at the 1 000th time-step).
Further zoom of Fig. 2.23 close to the step-variation that occurs at the 1 000th time-step.

0.01
DKF - heur
0.008 DKF - Myers
DKF - PECE
DKF - PECE diag
RMSE

0.006

0.004

0.002

0
990 1000 1010 1020 1030 1040
time-step

Figure 2.25 – Base case plus steps, q = 10−10 (a step occurs at the 1 000th time-step).
Time evolution of the RMSEs.

57
Chapter 2. Adaptive Kalman filtering for the on-line estimation of the
process-model uncertainties
0
10
-1 DKF - heur
10 DKF - Myers
DKF - PECE
10 -2

(3,3)
DKF - PECE diag
10 -3

k|k-1
P 10 -4
10 -5
10 -6
995 1000 1005 1010
time-step

Figure 2.26 – Base case plus steps, q = 10−6 (a step occurs at the 1 000th time-step). Time
evolution of the third diagonal element of Pk|k−1 (the other diagonal elements exhibit
an analogous behavior).

10 -1
DKF - Myers
DKF - heur
DKF - PECE diag
10 -2 DKF - PECE
RMSE

10 -3

10 -4
995 1000 1005 1010
time-step

Figure 2.27 – Base case plus steps, q = 10−6 (a step occurs at the 1 000th time-step). Time
evolution of the RMSEs.

the time evolution of the third diagonal element of Pk|k−1 estimated by the heuristic,
Myers, PECE and PECEdiag methods (the other diagonal elements exhibit an analogous
behavior). As in the case where q = 10−10 , the PECE and PECEdiag methods increase
Pk|k−1 as soon as the step occurs, whereas the heuristic and Myers methods react with a
delay of one time-step. Indeed, only the PECE and PECEdiag methods are characterized
by limited errors during the step, as shown in Fig. 2.27. Note that, again, the accuracy
of the PECE method is better than that of PECEdiag , due to the effect of the estimated
off-diagonal elements of Pk|k−1 .

Influence of parameter N on the SE accuracy

The only parameter that has to be set in the PECE method is N , specifically the number
of previous innovations used to calculate the sample innovation-covariance matrix in
(2.21). In this section, we present the influence of N on the SE accuracy considering
both the base case and the base case plus steps in which q/r < 1. As already explained
for the base case, q/r < 1 is the case of interest where the noise filtering can be effective.
The investigation involves the PECE and Myers methods. Unlike in the simulations
related to the base case and base case plus steps, where the RMSEs was calculated at
each time-step, here we show the RMSE that takes into account all the errors of the
T considered time-steps. Besides, we show the median value of the RMSEs obtained
in 10 different simulations. It is worth mentioning that we wait until the convergence

58
 2.5. New prediction-error covariance estimation method (PECE) for step-varying
processes
0.002
-3 DKF - PECE DKF - Myers
10

RMSE
10 -4

10 -5 1 2 3 4
10 10 10 10
N
(a) Base case (q = 10−10 , r = 10−7 ).

10 -2 DKF - PECE DKF - Myers

RMSE

-3
10

10 -4 1
10 10 2 10 3 10 4
N
(b) Base case plus steps (q = 10−10 , r = 10−7 ).

Figure 2.28 – RMSEs as a function of N for the PECE and Myers methods. In the base
case, the RMSE is computed considering 2 000 time-steps. In the base case plus steps,
the RMSE is computed considering the 30 time-steps after the state step-variation.

×10 -8
8
DKF - Myers
6
P k|k-1 (2,2)

4
2
0
-2
2 2.5 3 3.5 4 4.5 5 5.5 6
time-step ×10 4

1
True DKF - Myers

0.998
x(2)

0.996

0.994
2 2.5 3 3.5 4 4.5 5 5.5 6
time-step ×10 4

Figure 2.29 – Base case, q = 10−10 , r = 10−7 , N = 100. Time evolution of the 2nd
diagonal element of Pk|k−1 and of the associated 2nd element of the state vector: true
state and state estimated by the DKF that uses the Myers method. This figure shows
the numerical stability problems of the Myers method.

59
Chapter 2. Adaptive Kalman filtering for the on-line estimation of the
process-model uncertainties

phase is finished before performing this assessment.

Fig. 2.28a shows the RMSEs as a function of N for the base case. The RMSE is computed
by considering T = 2 000 time-steps. The PECE method degrades the estimation accu-
racy as we consider a smaller number of innovations. The innovations are affected by
the measurement noise that is not effectively filtered if a small number of innovations
is used. The RMSEs of the Myers method remains quite stable and smaller than that of
the PECE method for N > 500. Note that the use of a smaller value of N also leads to
a faster convergence of the Myers method to the true value of Pk|k−1 compared with
Fig. 2.20. However, the accuracy of the Myers method drops for small values of N (i.e.,
N = 20 and N = 100) due to the fact that the positive semi-definiteness of Pk|k−1 is
not guaranteed by the Myers method. As a consequence, the KF numerical stability
might be compromised and the KF solution sometimes diverges from the true state.
This phenomenon is visible in Fig. 2.29 where N = 100: the state estimated by the
DKF that uses the Myers method loses track of the true state in many occasions, e.g.,
when negative diagonal elements of Pk|k−1 are estimated (i.e., at time-step 5.7 · 104 ).
For the ideal case considered in this section, significant estimation errors related to
this problem occur only for N < 300.

Fig. 2.28b shows the RMSEs as a function of N for the base case plus steps. The RMSE
is computed by considering the T = 30 time-steps after the state step-variation. As
expected, the PECE method outperforms the Myers method in the period following the
step. In contrast with the base case, the smaller N is, the smaller the errors are, because
the last innovations have a higher weight in the computation of the sample covariance
(2.21).

2.5.3 Simulations and results for power-system state estimation

The effectiveness of the PECE method is here proven in the context of power-system SE
via the procedure presented in Section 2.4.2. We use the New England 39-bus system
and we consider the network operating conditions described in Section 2.4.2, except
for the time evolution of the reactive power absorbed by the load at bus #4. Specifically,
in order to show the behavior of the PECE method when a state step-variation occurs,
we simulate a sudden drop of 90 MVAR of the reactive power at bus #4, which can be
caused, for instance, by the disconnection of a shunt reactor. This event is shown in Fig.
2.30. In Fig. 2.31, we highlight the fact that the drop affects two time-steps as a result
of the response to transients of the phasor-estimation algorithm of the PMU (see [97]
for further details). The reactive-power drop produces step-wise perturbations in the
state variables, which are similar to those of the base case plus steps in Section 2.5.24 . It

4
We would like to clarify that the results do not change if we simulate a sudden variation of the active
power only or of both active and reactive powers simultaneously, because the covariance assessment
methods do not depend on what state variables experience the step-variation.

60
 2.5. New prediction-error covariance estimation method (PECE) for step-varying
processes
501.5 240

Reactive Power [MVAR]

Active Power [MW]
501 220
500.5 200

Bus #4

Bus #4
500 180
499.5 160
499 140
498.5 120
0 200 400 600 0 200 400 600
time [s] time [s]

Figure 2.30 – Time profiles of the active and reactive power at bus #4. A drop of about
90 MVAR of reactive power occurs at 600 seconds.

240
Reactive Power [MVAR]

220
200
Bus #4

180
160
140
120
599.92 599.94 599.96 599.98 600 600.02 600.04 600.06 600.08
time [s]

Figure 2.31 – Time profile of the reactive power at bus #4: zoom close to the 90 MVAR
step-variation.

is well-known that in transmission networks the reactive power mainly influences the
voltage magnitude; as we express the state in rectangular coordinates, both real and
imaginary parts of the voltage are affected.

The network schematic and the measurement configuration are depicted in Fig. 2.1 in
Section 2.4.2. We recall that the measurement configuration ensures the full-network
observability, i.e., matrix H has full rank (Assumption #3 of Section 2.5.1 holds).

The measurement noise is assumed to be a Gaussian-white sequence where the co-

variance matrix R is known, time-invariant and diagonal. Further explanations can be
found in Section 2.4.2. The standard deviations of both the real and imaginary parts of
the measurements are assumed to be 0.33% of the full scales of the sensors.The voltage
sensors’ full scales are assumed to be equal to the rated voltage; the current sensors’
full scales are assumed to be equal to twice the largest value of the current-injection
magnitude at the respective bus in the considered simulation period. From the above-
mentioned hypotheses, the linear measurement model is known and time-invariant
(Assumption #2 of Section 2.5.1 holds).

Moreover, we use the process model given in (2.1) for the reasons already given in
Section 1.4.3 (Assumption #1 of Section 2.5.1 holds).

61
Chapter 2. Adaptive Kalman filtering for the on-line estimation of the
process-model uncertainties

Accuracy performance analysis

In the following we analyze the accuracy performance of five state estimators, i.e.,
four DKFs that use the Myers, heuristic, PECE and PECEdiag methods, and the LWLS.
As in Section 2.5.2, the heuristic method uses N = 30, whereas the PECE, Myers and
PECEdiag methods use N = 5 000. The influence of parameter N on the SE accuracy
for the specific application of power-system SE is presented later in this section.

The first part of the simulation consists in an initial phase of 300 seconds (correspond-
ing to 15 000 time-steps as for the base case in Section 2.5.2) in which we keep a constant
value of Q for the considered DKFs and their estimated states converge from a flat-start
initialization towards the true state. For brevity, the initial phase is not shown in this
section.

The time evolution of the RMSEs of the five considered state estimators is given in Fig.
2.32; it is composed of three sub-figures: Fig. 2.32a shows the entire simulation, Fig.
2.32b shows a portion of the simulation where the network is in quasi-static conditions,
and Fig. 2.32c shows the portion of the simulation close to the state step-variation. The
RMSEs of the Myers method exhibit high sporadic spikes that are clearly visible in Fig.
2.32a. This numerical stability issue of the Myers method has already been described
at the end of Section 2.5.2 (in particular, see Fig. 2.29), and it is due to the fact that the
Myers method does not ensure the positive semi-definiteness of Pk|k−1 . Note that the
smaller N , the worse the numerical stability issue.

Fig. 2.32b shows the RMSEs when the network is in quasi-static conditions, i.e., the
load powers are varying smoothly with no steps of significant amplitude. The errors of
the four DKFs are comparable except for the occasional spikes of the Myers method,
whereas the LWLS errors are about three times larger on average. We can observe that,
unlike in the base case in Section 2.5.2, the accuracy of the heuristic method is just
slightly worse than that of the other DKFs.

Fig. 2.32c shows the RMSEs during and after the step-variation. The LWLS keeps
the same estimation accuracy, because at every time-step it relies only on the mea-
surements taken at that time-step. The KF, instead, uses also the past information to
compute the system state. As soon as the step occurs, Pk|k−1 should be adapted in or-
der to attribute a lower weight to the predicted state with respect to the measurements.
In other words, the KF behavior should approach that of the LWLS. The Myers and
heuristic methods detect the occurrence of the step with a time-step delay, because
they use only the state estimates. Therefore, they are characterized by large estimation
errors when the step occurs. After the step, the heuristic method recovers the correct
state tracking faster than the Myers method. The PECE method is the only one that has
a peak of the errors comparable to the one of LWLS. Indeed, it estimates the new Pk|k−1
by using the innovations before the measurement update, so that the information

62
 2.5. New prediction-error covariance estimation method (PECE) for step-varying
processes

brought by the measurements at the time of the step is already taken into account in
the Kalman gain. The accuracy of the PECEdiag method is significantly worse than that
of the PECE method, proving the importance of using a full matrix in the optimization
problem (2.23). These results confirm those obtained for the base case in Section 2.5.2.

The reactive power step-variation affects mainly the real part of the voltage phasors
at bus #4 and at the neighbor buses. Fig. 2.33 shows the time evolution of the real
part of the voltage phasor (in pu of the rated voltage) at bus #4: the true value and the
value estimated by the considered state estimators. It reflects the estimation errors
shown in Fig. 2.32c. It is worth mentioning that some of the state variables estimated
by the Myers, heuristic and PECEdiag methods go even on the opposite direction of the
true state variation. For instance, this undesired behavior occurs with the Myers and
heuristic methods for the real part of the voltage phasor at bus #31, as shown in Fig.
2.34. This is due to the incorrect value of Pk|k−1 estimated by these methods when the
step occurs. These wrong estimates are other possible equilibrium points of the power
system far from the true ones. Note that this behavior was observed also in [37].

Influence of parameter N on the SE accuracy

This section shows the influence of parameter N on the SE accuracy. This study
is carried out separately for the network in quasi-static conditions and during the
step-variation. In the first case, we consider the 100 seconds prior to the step. In the
second case, we consider the 2 seconds following the step-variation. For both cases,
we compute the RMSE for all the time-steps contained in the considered time-window
and we show the median value of the RMSEs obtained in 10 different simulations.
The RMSEs as a function of N are shown in Figs. 2.35a and 2.35b for the system in
quasi-static conditions and during the step-variation, respectively. The accuracy of
both the PECE and Myers methods are examined.

Concerning the system in quasi-static conditions, the estimation accuracy of both

the PECE and Myers methods degrades as N diminishes. However, it can be seen that
the accuracy of the PECE method is much better than that of the Myers method for
all the considered values of N . This is mainly due to the numerical stability problem
of the Myers method that we have previously discussed in the accuracy performance
analysis.

As expected, also the study on the step-variation reveals that the PECE method out-
performs the Myers method for all the considered values of N . The accuracies of both
methods decrease as N decreases, which indicates an opposite trend with respect to
the results of the ideal case in Section 2.5.2. Regarding the Myers method, the reason
lies in more significant numerical stability problems. Instead, the worsening of the
PECE method’s accuracy as N decreases is due to the increasing effect of the measure-

63
Chapter 2. Adaptive Kalman filtering for the on-line estimation of the
process-model uncertainties
×10 -3
5
LWLS
4.5 DKF - heur
DKF - Myers
DKF - PECE
4 DKF - PECE diag

3.5

3
RMSE [pu]

2.5

1.5

0.5

0
300 350 400 450 500 550 600 650 700
time [s]

(a) Entire simulation.

-4
×10
6
LWLS
DKF - heur
DKF - Myers
5 DKF - PECE
DKF - PECE diag

4
RMSE [pu]

0
570 575 580 585 590 595 600
time [s]

(b) Zoom of Fig. 2.32a during quasi-static conditions.

×10 -3
3
LWLS
DKF - heur
DKF - Myers
2.5 DKF - PECE
DKF - PECE diag

2
RMSE [pu]

1.5

0.5

0
64 600 600.5 601 601.5 602 602.5 603
time [s]

(c) Zoom of Fig. 2.32a on the state step-variation.

Figure 2.32 – Time evolution of the RMSEs.

 2.5. New prediction-error covariance estimation method (PECE) for step-varying
processes
0.98

0.978

0.976

V re [pu] at Bus 4
0.974

0.972

0.97
True
LWLS
0.968 DKF - heur
DKF - Myers
DKF - PECE
DKF - PECE diag
0.966
599.8 600 600.2 600.4 600.6 600.8 601
time [s]

Figure 2.33 – Time evolution of the real part of the voltage at bus #4 that is the bus
where the reactive power step-variation occurs.

1.002

1.001

0.999
V re [pu] at Bus 31

0.998

0.997

0.996

0.995 True
LWLS
DKF - heur
0.994 DKF - Myers
DKF - PECE
DKF - PECE diag
0.993
599.8 600 600.2 600.4 600.6 600.8 601
time [s]

Figure 2.34 – Time evolution of the real part of the voltage at bus #31: true and estimated
values.

65
Chapter 2. Adaptive Kalman filtering for the on-line estimation of the
process-model uncertainties
10 -1
DKF - PECE DKF - Myers
10 -2

RMSE [pu]
10 -3

-4
10

-5
10 1 2 3 4
10 10 10 10
N
(a) Quasi-static conditions.

10 -1
DKF - PECE DKF - Myers
RMSE [pu]

-2
10

10 -3

10 -4 1 2 3 4
10 10 10 10
N
(b) State step-variation.

Figure 2.35 – RMSEs as a function of N for the PECE and Myers methods. For case (a)
where the network is in quasi-static conditions, the RMSE is computed considering the
100 seconds prior to the step. For case (b) where the network state has a step-variation,
the RMSE is computed considering the 2 seconds following the step.

ment noise on the sample covariance matrix (2.21). Indeed, we have observed that the
small peak of the errors of the PECE method close to the step-variation has the same
amplitude irrespectively of the value of N , whereas the errors after the step are highly
affected by the measurement noise. The latter is not effectively filtered if a small value
of N is set.

In conclusion, the PECE method guarantees a better accuracy than the Myers method
for all the considered values of N in both quasi-static conditions and during the state
step-variation.

Computational-time performance assessment

The remarkable estimation accuracy of the PECE method comes at the expense of
the computational time. The latter is almost entirely devoted to solving the convex
optimization problem of (2.23) that depends only on the state dimension n. We can
instead neglect the time used to compute the sample covariance matrix (2.21), so that
the computational time is not function of N . Fig. 2.36a shows the computational
time of the PECE method as a function of n. Each value is an average over 100 simu-
lations. These results are obtained by solving the optimization problem (2.23) with
YALMIP employing the sdpt3 solver implemented in MATLAB 2014b [95, 96]. The

66
 2.6. Conclusions

laptop is an Apple MacBook Pro with a 2.5-GHz CPU, 16-GB RAM. The increase of the
computational time shown in Fig. 2.36a is exponential as the y-axis is in logarithmic
scale.

On the contrary, the computational time of the Myers and heuristic methods is signif-
icantly affected by the parameter N . Figs. 2.36b and 2.36c show the computational
time of these methods, as a function of the state dimension n; the three curves refer
to three values of N , i.e., 100, 1 000 and 10 000. Both methods are considerably faster
than the PECE method. In particular, the Myers method is characterized by a computa-
tional time that increases exponentially as a function of n, and it remains below 100
milliseconds for the considered values of n. The heuristic method is the fastest one
with a computation time always below 10 milliseconds.

2.6 Conclusions
In this chapter we compare the performance of the LWLS and DKF algorithms in terms
of accuracy and computational time. This analysis is carried out by employing the
following assumptions: (1) there are no systematic-errors, only random Gaussian-
white noise is present; (2) the linear measurement-model is known; (3) we use only
high time-resolution measurements from PMUs; (4) there are no bad-data and network
model errors, thus, the SE accuracy is directly linked to the capability of reducing the
measurement noise.

The first part of this chapter is dedicated to the definition and testing of a new heuristic
method for the assessment of Q that is used in a DKF. The DKF employs the persistent
process-model defined in (2.1). This method estimates Q based on the knowledge
of the previous N state estimates. The method is straightforward to implement and
the only parameter that needs to be set is N that requires a tuning stage. The case
studies consist in the New England 39-bus system and the 123-bus test feeder that are
assumed to be in quasi-static operating conditions. The theory predicts that DKF is
more accurate than the LWLS as long as the process model is correct. We statistically
validated the adequacy of the persistent process-model used in combination with the
proposed heuristic method that infers Q. Then, we showed that the DKF outperforms
the LWLS in terms of accuracy, but the LWLS is faster in terms of computational time.
However, both of them are able to estimate the system state in less than 100 ms for these
case studies and are suitable for RTSE of networks of similar size. We observed that
the proposed heuristic method is fast and able to adapt Q to the changing operating
conditions of the network. We also noticed that the improvement of the DKF results
with respect to the LWLS results is particularly pronounced when the redundancy level
is low.

In the second part of this chapter, we proposed a new method, called prediction-error

67
Chapter 2. Adaptive Kalman filtering for the on-line estimation of the
process-model uncertainties
2
10

1
10

time [s]
10 0

10 -1

10 -2
0 20 40 60 80 100
n
(a) PECE method.

10 -1

10 -2
time [s]

-3
10

-4
10
N = 102
N = 103
N = 104
10 -5
0 20 40 60 80 100
n
(b) Myers method.

10 -2

10 -3
time [s]

10 -4

N = 102
N = 103
N = 104
10 -5
0 20 40 60 80 100
n
(c) Heuristic method.

Figure 2.36 – Computational time as a function of n. For the Myers and heuristic
methods, multiple curves refer to different values of N.

68
 2.6. Conclusions

covariance estimation (PECE), for the assessment of Pk|k−1 that is suitable for SE
of step-varying processes where the process-noise covariances are time-varying and
unknown. It relies on the knowledge of a linear and time-invariant measurement model
and was tested for a persistent process-model. The PECE method correctly estimates
the value of Pk|k−1 for a process characterized by constant noise-covariances, and
tracks the system state even during large step-variations. To the best of our knowledge,
no other method proposed in the literature relying on the same assumptions is able to
achieve such accuracy performance. The PECE method makes use of a constrained
convex optimization problem that computes Pk|k−1 from the innovations and ensures
the positive semi-definiteness of Pk|k−1 . Only parameter N has to be set, which is the
number of previous innovations that are used to compute Pk|k−1 . We tested the PECE
method for ideal cases and for DKF-SE of the New England 39-bus system where a
large step-variation of the system state is simulated. We also give a comparison with
the SE results of a LWLS, a DKF that use a value of Q assessed by the heuristic method
presented in Section 2.4.1 and a DKF that use a value of Q assessed by the method
presented in [89]. On one hand, the PECE method is faster than the other two methods
in tracking state step-variations. On the other hand, it is computationally expensive
for high-dimensional systems. It is also important to highlight that the PECE method
estimates the full matrix Pk|k−1 , and we showed that the correct assessment of the
off-diagonal entries of Pk|k−1 plays an important role in the proper tracking of state
step-variations. Future research will focus on a more general process model.

69
3 Linear state estimation with zero-
injection equality constraints via
LQ-decomposition
In this chapter, we discuss the problem of linear power-system state estimation with
linear equality constraints associated with zero-injection buses. Specifically, we focus on
the linear weighted-least-squares algorithm. We present various well-known approaches
to handle zero-injections; then, we propose a method based on LQ-decomposition that
strictly satisfies the equality constraints and reduces the dimension of the state vector.
We demonstrate the performance improvement of the proposed method in terms of
numerical stability and computational efficiency via numerical simulations conducted
on the 39-bus New England test system.

3.1 Introduction
The content of this chapter is based on [98]. The presence of zero-injection buses
is common in power systems. These buses are characterized by an injected power
strictly equal to zero, which is the case of interconnection substations. Power-system
SE can exploit zero-injections as an information that contains no error. Therefore, SE
becomes a constrained optimization problem where zero-injections are the equality
constraints.

Making reference to a WLS state estimator, a simple yet approximated way to include
equality constraints is to convert the constraining equations in virtual measurements
[39, 40]. The latter are defined in Section 1.1.1 as additional measurements with very
small uncertainties. This method is approximated as it cannot strictly satisfy the
zero-injection equalities. Besides, the main drawback of this approach is that the
coefficient matrix1 in the WLS NE may become poorly conditioned due to the disparity

1
The coefficient matrix refers to the matrix that contains the coefficients of the variables in a set of
linear equations.

71
Chapter 3. Linear state estimation with zero-injection equality constraints via
LQ-decomposition

of weights assigned to real-time measurements, virtual measurements (associated

with zero-injections) and pseudo-measurements (associated with historical data or
forecasts) [99, Chapter 22]. In Section 1.1.1 we clearly state the difference between
virtual measurements and pseudo-measurements.

Alternatively, zero-injection equality constraints can be strictly satisfied by employ-

ing the method of Lagrange multipliers. However, in this case the state dimension
increases and, in general, the modified coefficient matrix becomes not positive definite
[100, 101]. The positive definiteness of the coefficient matrix is important to enhance
the computational speed of the SE solution; with reference to the WLS, as G is positive
definite, it can be decomposed in Cholesky factors so that the NE can be solved in an
efficient way by forward/backward substitution.

Every approach that introduces the zero-injection information modifies the NE coeffi-
cient matrix and its condition. As the ill-conditioning of this matrix is a well-known and
important problem, in the following we summarize the relevant techniques that have
been proposed in the literature to alleviate this issue. In the context of non-linear WLS-
SE, several approaches are described and compared in [102, 54] [6, Chapter 3]. The use
of QR-decomposition (QRD) applied to the WLS NE considerably enhances the numeri-
cal stability at the expense of matrix sparsity and computation time [103, 104, 102, 105].
A comparative study focusing on numerical stability, computational efficiency, and
implementation complexity is given in [102]. In [6, Chapter 3] various approaches are
tested against different sources of ill-conditioning. The authors of [106] propose an
iterative method to satisfy the zero-injection equality constraints.

Note that all the above-mentioned methodologies were conceived for non-linear
iterative WLS-SE. When the measurements set contains only synchrophasor measure-
ments provided by PMUs, the SE problem becomes linear. Power-system linear SE
was firstly proposed in [29] and further investigated recently, e.g., [26, 72]. The mea-
surement matrix linking measurements and state variables is exact (not a Jacobian),
state-independent and constant in time. Consequently, the equality constraints associ-
ated with zero-injections are linear as well. As it is known, a constrained optimization
problem that has only linear equality constraints can be solved for a set of variables re-
duced by the number of constraints. The problem becomes an unconstrained problem
in a smaller number of variables.

In this chapter, we propose a method based on LQ-decomposition (LQD) to solve the

problem of LWLS-SE with linear equality constraints, which has the following features:

1. the state dimension is reduced by the number of linear equality-constraints

imposed by zero-injections;

2. the computation time is considerably decreased;

72
 3.2. Linear State Estimation with Linear Equality Constraints

3. the numerical stability is improved;

4. the equality constraints are strictly satisfied;

5. the structure of the WLS NE is preserved;

6. the coefficient matrix remains positive-definite.

It is worth pointing out that the LQD of a generic matrix W corresponds to the QRD
of WT . As the the NE structure is maintained, no extra modifications are required in
the implementation of already-developed techniques, such as the widely used LNR
method for bad-data processing [8]. The proposed method is compared with other
well-known techniques that were initially introduced for handling zero-injections in
non-linear WLS.

3.2 Linear State Estimation with Linear Equality Constraints

Let us consider a power-system with s buses and c zero-injection buses. We assume to
have a measurement infrastructure composed exclusively of PMUs that provide real-
time phasor measurements of nodal-voltages, nodal-current injections and current
flows.

First, we recall the LWLS problem that uses real-time measurements from PMUs and
pseudo-measurements, without considering zero-injections. We provide again the
LWLS unconstrained optimization problem defined in Section 1.2 by making explicit
the presence of x:

1
min J = (z − Hx)T R−1 (z − Hx) (3.1)
x 2
resulting in the well-known NE:

G x = HT R−1 z (3.2)

where these vectors and matrices are defined in Section 1.2. We recall that x ∈ Rn , z ∈
Rm includes only real-time measurements and pseudo-measurements, R is diagonal
and positive definite, and G = HT R−1 H. Given that the network is observable, H is full
rank and G is positive-definite. Therefore, G can be decomposed in Cholesky factors
so that the NE can be solved in an efficient way by forward/backward substitution.

73
Chapter 3. Linear state estimation with zero-injection equality constraints via
LQ-decomposition

The presence of w zero-injection buses adds p equality constraints2 to the problem in

(3.1):
1
min J = (z − Hx)T R−1 (z − Hx) (3.3)
x 2
subject to: Cx = 0

where C is the p×n measurement matrix related to the vector i0 of the current injections
at the zero-injection buses, which is obviously a null vector 0.

Assumption 1. The network is fully observable with the combination of real-time

measurements, pseudo-measurements and zero-injections.

This assumption implies that m + p > n, but not necessarily that m > n. Therefore, the
network can also be unobservable by considering only real-time measurements and
pseudo-measurements.

Note that H and C are exact matrices (i.e., not Jacobians). They are also constant
in time as they do not depend on the network state, but only on the network model
(i.e., network topology and electrical parameters). Therefore, H and C need to be
recomputed only if a change in the network model occurs.

On the contrary, we let R to be time-variant. However, if R were assumed to be time-

invariant, the formulations of the methods treated in this chapter would not change
and the computational performances would be enhanced. The case of a time-invariant
R is discussed in Section 3.3.2.

In what follows, we describe three methods that deal with zero-injection equality
constraints in the context of LWLS-SE:

1. NE including virtual measurements, called VM ;

2. NE with constraints, called NE/C;

3. NE derived from LQD, called LQD, which is the proposed method.

It is worth pointing out that the VM and NE/C methods have been widely discussed in
the literature and we simply reformulate them for linear SE in Sections 3.2.1 and 3.2.2.
Then, in Section 3.2.3 we present the proposed method based on LQD.

Finally, in Section 3.2.4 we recall the well-known QRD that improves the condition of
the coefficient matrix in the NE.
2
Specifically, p = 2w in case of single-phase SE and p = 6w in case of three-phase SE.

74
 3.2. Linear State Estimation with Linear Equality Constraints

3.2.1 Normal equation with virtual measurements (VM)

One way to treat equality constraints in (3.3) is to represent them as virtual measure-
ments. The augmented measurement vector including zero-injections is z0 ∈ Rm+p
and the corresponding measurement-noise covariance matrix is R0 . The measurement
matrix is

H
H0 = (3.4)
C

which has full rank thanks to Assumption 1 in Section 3.2. The NE becomes

HT0 R−1 T −1
0 H0 x = H0 R0 z0 . (3.5)

Compared to (3.2), the dimensions of the state and of the coefficient matrix do not
change, whereas the measurement vector is larger. The coefficient matrix is expected
to be more ill-conditioned as we added virtual measurements with small variances.

3.2.2 Normal equation with constraints (NE/C)

The optimization problem (3.3) can be solved by employing the Lagrange multipliers
method, which leads to the following Lagrangian function to be minimized as

1
min L = (z − Hx)R−1 (z − Hx) + λCT x (3.6)
x,λ 2

where λ is the vector of Lagrange multipliers. Deriving (3.6) with respect to x and λ
yields

αHT R−1 H CT x αHT R−1 z

= (3.7)
C 0 λ 0

where α = min(Rii ) is added as a scaling factor that improves the condition of the
coefficient matrix [6, Chapter 3]. In general, the coefficient matrix in (3.7) is no longer
positive definite.

3.2.3 Normal equation derived from LQ-decomposition (LQD)

The dimension of the optimization problem (3.3) can be reduced as it is linear with
only linear equality constraints. To this end, we propose a method based on LQD that

75
Chapter 3. Linear state estimation with zero-injection equality constraints via
LQ-decomposition

is described in what follows. We apply the LQD only to C:

  V1
C = UV = U1 0 = U1 V 1 (3.8)
V2

where U is a p × n lower triangular matrix and V is an n × n orthogonal matrix; then,

V1 is of size p × n, V2 is of size (n − p) × n, and U1 is a p × p lower triangular matrix.

Proposition 1. The full state x can be written as:

x = V2T u (3.9)

where u ∈ Rn−p is the new state vector that has a reduced number of variables.

Proof. First, the fact that V is orthogonal implies that

VVT = I (3.10)

VT V = I (3.11)

and from (3.10) we can also write

V1 V1T = I V2 V2T = I V1 V2T = 0 V2 V1T = 0 (3.12)

Second, using (3.8) and the fact that Cx = 0, we can write

Cx = U1 V1 x = 0 . (3.13)

and thus

V1 x = 0 . (3.14)

Then, we use (3.11) and (3.14) to write

x = VT Vx = V1T V1 x + V2T V2 x = V2T V2 x (3.15)

Therefore, we can define a vector u such that

u = V2 x (3.16)

and equation (3.15) becomes

x = V2T V2 x = V2T u . (3.17)

76
 3.2. Linear State Estimation with Linear Equality Constraints

(QED)

Considering Proposition 1, the constrained optimization problem (3.3) can be written

as an unconstrained optimization problem in a reduced number of variables as follows:

1
min J = (z − HV2T u)T R−1 (z − HV2T u) (3.18)
u 2
By defining a new m × (n − p) measurement matrix as

 = HVT
H (3.19)
2

the optimization problem (3.18) corresponds to the following NE:

H  u=H
 T R−1 H  T R−1 z (3.20)

 =H
where G  T R−1 H
 is the coefficient matrix of size (n − p) × (n − p).

Note that the LQD of C and the computation of H do not need to be performed at every
time-step, because C is constant in time, except for seldom changes in the network
model. Therefore, at every time-step we just need to solve the NE in (3.20).

 is positive definite.
Proposition 2. The coefficient matrix G

Proof. G is positive definite if both R and H

 have full rank. R has full rank by con-

struction, whereas we have to demonstrate it for H.

Since the number of rows of H  (m) is larger than the number of columns (n − p) by
 is reduced to {0}, i.e. we
Assumption 1, all we need to show is that the null space of H
have to prove that

 = HVT u = 0
Hu =⇒ u=0. (3.21)
2

Using (3.4), we can write

HV2T u
H0 V2T u = (3.22)
CV2T u

where HV2T u = 0 by assumption (3.21), whereas from (3.8) and using V1 V2T = 0 in
(3.12) we derive CV2T u = U1 V1 V2T u = 0.

Therefore, H0 V2T u = 0, which implies that V2T u = 0, because H0 has full rank accord-
ing to Assumption 1. Considering that V2 V2T = I in (3.12), it follows that u = 0.

(QED)

77
Chapter 3. Linear state estimation with zero-injection equality constraints via
LQ-decomposition

The proposed method consists in the following steps:


Pre-computation of H:
1. Compute the LQD of C in order to find V2 , as in (3.8);
 by using (3.19).
2. Compute H

State estimation at every time-step:

 in Cholesky factors and solve for u;
1. In the NE (3.20), decompose G

2. Compute the full state x by using (3.9).

3.2.4 QR-decomposition (QRD)

As it is known, the numerical stability of the NE can be improved by using QRD

[103, 104, 102]. This procedure can be applied to both equations (3.5) and (3.20)
and the corresponding methods are called VM+QRD and LQD+QRD, respectively. On
the contrary the QRD cannot be applied to the equation of the NE/C method, because
of the different structure of the coefficient matrix in (3.7).

To describe the method, we make reference to the general NE (3.2). First, we apply the
1
QRD to R− 2 H [103, 54]:

− 12
B1
R H = FB = [F1 F2 ] (3.23)
0

where F is an m × m orthogonal matrix, B is m × n, and B1 is an n × n upper triangular

matrix. The NE (3.2) can be re-written as follows:
1 1 1 1
HT R− 2 R− 2 H x = HT R− 2 R− 2 z (3.24)

and thus
1
BT FT FB x = BT FT R− 2 z . (3.25)

The orthogonal matrix property FT F = I yields

1
BT B x = BT FT R− 2 z . (3.26)

By using the matrix blocks in (3.23), we can further simplify (3.26) as follows:
1
B1 x = FT1 R− 2 z . (3.27)

This system of equations can be easily solved by backward substitution as B1 is upper

triangular.

78
 3.3. Simulations and results

Bus Zero-injection bus Transformer V Voltage meas.

Load Transmission line G Generator I Current meas.

G8
37
G10
30 26 28 I I I I 29
25 I V
I V I V V I I
I I I
2 I 27 38
I I V
G9
1 24
18 I 17 V
I I I
I I V
3 16 G6
I 39 35
I I I V V I
I 15
G1 V I
I I I 21 22
I I 4 14 I V

I I V

5 12 I V
19
6 I I
I 23 I

11 13 V I I
7
10 20 I
I I 36
8 V I I
V
I I G7
V I 31 I 32 34 V I 33
9 V
Zsc
SB G3 G5 G4

Figure 3.1 – Network topology of the New England 39-bus system together with the
adopted measurement configuration.

3.3 Simulations and results

3.3.1 Test conditions

In this section, we compare the methods described in Section 3.2 in terms of numerical
stability and computational efficiency. For sake of comparison, we also consider a
LWLS, which we call NZI, that uses no zero-injection information, i.e., it corresponds
to solve (3.2). We carry out single-phase SE for the 39-bus New England test system
depicted in Fig. 3.1, where we reproduced the network operating conditions described
in [91]. The network is assumed to be in static conditions.

The adopted measurement configuration is shown in Fig. 3.1. No pseudo-measurements

are used. PMUs are installed in 19 buses; each PMU measures the nodal voltage, the
nodal current-injection and all the current-flows in the lines departing from a bus. The
set of real-time phasor measurements is composed of 19 voltages, 19 current-injections
and 33 current flows (m = 142). This measurement placement aims at achieving the
full-network observability even without accounting for the zero-injection information,
so that we can also run the NZI estimator. Note that, in general, it is sufficient that
Assumption 1 in Section 3.2 holds. Given that the number of state variables is n = 78,
the redundancy level with only real-time measurements is m/n = 1.8. Considering

79
Chapter 3. Linear state estimation with zero-injection equality constraints via
LQ-decomposition

also the 12 zero-injection buses that add p = 24 equality constraints, the redundancy
level is (m + p)/n = 2.1.

The true values of voltages and currents are obtained from a load-flow calculation.
The phasor measurements are generated by adding Gaussian-white noise to the true
values and are expressed in rectangular coordinates. The noise standard-deviations
are expressed in polar coordinates: the magnitude standard-deviation is 0.17 % of
the measured value and the phase-angle standard-deviation is 3 mrad. However, the
diagonal entries of R need to be in rectangular coordinates; thus, these standard-
deviations are projected from polar to rectangular coordinates by using the formulas
described in [48, Chaper 6.4]. This procedure results in diagonal entries of R ranging
from 1.6 · 10−6 to 1.8 · 10−3 . With reference to the VM method, the augmented matrix
R0 contains also the variances of the virtual measurements, which are computed as

min(Rii ) 1.6 · 10−6

σV2 M = = (3.28)
γ γ

where γ is a scalar that we use to modify the virtual-measurement variances.

3.3.2 Results

Tab. 3.1 summarizes the features of the considered methods and reports the results
obtained for the adopted grid, which are discussed in what follows. We point out that
the results of Tab. 3.1 are obtained by using γ = 1 for the virtual measurements.

We observe that the LQD method uses the least number of measurements (m) and
estimates the least number of state variables (n − p); therefore, we expect it to be
the most computationally efficient, as proved later in this section. We recall that
the reduction of the state dimension via LQD is not repeated at every time-step as it
involves only matrix C that is constant in time.

As we consider R to be time-variant, each method has to solve a linear system of equa-

tion as Ax = b, where A is the coefficient matrix. Depending on the characteristics of
A, we solve the system with different methods. For instance, if A is full and positive
definite, Cholesky factorization is the most efficient. We used the backslash operator “\”
of MATLAB that automatically checks A and employs the fastest method, as reported
in Tab. 3.1.

The condition number of the coefficient matrix quantifies the matrix ill-conditioning;
thus, we use the condition number as numerical-stability indicator. The VM method
has the highest condition number, although we set γ = 1, which means that we do not
use lower variances for the virtual measurements. The QRD decreases significantly the
condition number from 107 to 103 . The LQD method also leads to a significant stability

80
 3.3. Simulations and results

Table 3.1 – Comparison of the considered methods.

Network NZI VM NE/C LQD VM+QRD LQD+QRD

z size m m+p m m m+p m
x size n n n+p n−p n n−p
Coeff. matrix: Type Full Full Full Full Triangular Triangular
Coeff. matrix: Positive def.? YES YES NO YES YES YES
Method used to solve Ax=b Cholesky Cholesky LU fact. Cholesky Back. subst. Back. subst.
Coeff. matrix: Cond. number 2.5 · 105 1.4 · 107 1.8 · 105 6.4 · 104 3.7 · 103 2.5 · 102
RMSE [pu] 1.7 · 10−3 7.5 · 10−4 7.5 · 10−4 7.5 · 10−4 7.5 · 10−4 7.5 · 10−4
MEP-ZIB [kW] 36 · 103 70 10−2 10−2 70 10−2
CPU time - mean value [μs] 380 420 480 260 510 380

improvement of three order of magnitudes, but we get the smallest condition number
of 102 with the joint application of LQD and QRD.

The estimation accuracy is quantified by the RMSE of the state estimates in a simulation
of T = 1 000 time-steps; it is computed as


 1  T  n
RMSE = k,i − xk,i ) .
(xest true 2 (3.29)
T +n
k=1 i=1

In this specific case, every method provides the same RMSE, because the coefficient
matrices are not highly ill-conditioned. Only the NZI method has a larger RMSE,
because it does not exploit the zero-injection information.

We check whether the zero-injection equality constraints are satisfied by computing

the maximum estimated power at the zero-injection buses (MEP-ZIB), which should
be zero. The VM method does not strictly satisfy the constraints with γ = 1; the MEP-
ZIB is 70 kW. Whereas, the NE/C and LQD methods are able to match the constraints;
the MEP-ZIB is few tens of watts due to the numerical-accuracy limit. The CPU time
averaged on T = 1 000 time-steps indicates the computational efficiency. Simulations
are carried out in MATLAB 2014b running on an Apple MacBook Pro with 2.5-GHz CPU
and 16-GB RAM. It is also worth pointing out that the sparsity property of the matrices
is not exploited by means of dedicated tools. First, let’s examine the methods that do
not employ the QRD. The NE/C method is the slowest. The LQD method improves the
CPU time of 40% with respect to the VM method. Second, it can be seen that adding
the QRD worsens the CPU time for this specific case. However, for smaller networks
(approximately less than 20 buses), the QRD decreases the CPU time: applying the
QRD and then solving by backward substitution is faster than solving directly the NE by
Cholesky factorization. As we consider larger networks, the QRD becomes dominant
in the computation time.

As a final observation, in LWLS-SE, if H and R are time-invariant, the coefficient matrix

81
Chapter 3. Linear state estimation with zero-injection equality constraints via
LQ-decomposition
Table 3.2 – Impact of the variance assigned to virtual measurements on numerical
stability.

VM VM VM VM+QRD VM+QRD VM+QRD

γ=1 γ = 103 γ = 107 γ=1 γ = 103 γ = 107
Coeff. matrix: Condition number 1.4 · 107 1.4 · 1010 1.4 · 1014 3.7 · 103 1.2 · 105 1.2 · 107
RMSE [pu] 7.5 · 10−4 7.5 · 10−4 8.4 · 10−4 7.5 · 10−4 7.5 · 10−4 7.5 · 10−4
MEP-ZIB [kW] 70 10−2 10−2 70 10−2 10−2

of all the methods is constant in time and can be computed only once. At each time-
step, the computation just consists in a matrix multiplication and the CPU time is
dramatically decreased. Therefore, if R is assumed to be time-invariant, the CPU times
of all the methods are in the order of few tens of μs for the considered example.

3.3.3 Impact of virtual-measurement variance on numerical stability

In Tab. 3.2, we illustrate the impact of virtual-measurement variance on numerical

stability. We use the parameter γ to change the variance: larger γ corresponds to lower
variance, as denoted by (3.28).

The case of γ = 1 is the same of Tab. 3.1. Setting γ = 103 we obtain a MEP-ZIB
that is similar to the NE/C and LQD methods, but we worsen the condition number.
When γ = 107 , the ill-conditioning becomes severe and starts affecting the estimation
accuracy. The QRD maintains the numerical stability even for high ill-conditioning
levels.

3.4 Conclusions
In the context of linear SE with only linear equality constraints, the proposed method
based on LQD strictly satisfies the equality constraints derived from zero-injections
and reduces the number of state variables by the number of constraints. The structure
of the NE remains the same, which enables the use of existing techniques linked to
the WLS algorithm, such as the QR-decomposition and the LNR test. The positive
semi-definiteness of the coefficient matrix is preserved so that the NE can be solved
by employing efficient procedures. Tests were performed on the New England 39-bus
system (that has 12 zero-injection buses) in order to assess the performance in terms
of numerical stability and computational time with respect to other methods that
handle equality constraints: the use of virtual measurements (VM) and the method
of Lagrange multipliers (NE/C). The simulation results showed that the LQD method
improves significantly both numerical stability and computational time compared to
the other methods. Additionally, the combination of LQD and QRD further enhances

82
 3.4. Conclusions

the numerical stability at the expense of the computational time.

83
4 Results of PMU-based state estima-
tion of real networks

In this chapter, we present the results of real-time state estimation implemented in

two real power systems: a distribution feeder and a sub-transmission network. State
estimation is exclusively based on synchrophasor measurements provided by PMUs. We
describe the network characteristics as well as the measurement infrastructure. Then, we
provide a detailed analysis of the accuracy and computational performance of different
state-estimation algorithms: linear weighted least squares (LWLS), least absolute value
(LAV) and discrete Kalman filter (DKF).

4.1 Introduction
From the best of our knowledge, the implementation of PMU-based RTSE in a real
power system is described only in one paper, i.e., [26], and further detailed in [27].
It consists in a three-phase LWLS installed in the 500-kV system of Virginia Electric
and Power Company with the purpose of investigating phase-unbalance issues and of
improving the tuning of protective relays. The state estimator provides 30 estimates-
per-second by using voltage and current-flow synchrophasor measurements and has a
redundancy level of 2.2. The implementation and testing are widely discussed, whereas
the RTSE results are briefly presented.

In this chapter, we present the implementation and results of PMU-based RTSE for two
real power-networks: a 20-kV feeder of the distribution network of the EPFL campus,
and a portion of the 125-kV sub-transmission network of the city of Lausanne, which
is operated by Service Industriels de Lausanne (SiL). The originality of this work lies in
the use of synchrophasor measurements produced by PMUs installed in real networks.
We compare the accuracy and computational performance of the LWLS, LAV and DKF
estimators described in Chapters 1, 2 and 3.

85
Chapter 4. Results of PMU-based state estimation of real networks

Bus Zero-injection bus V Voltage meas.

Load Transmission line I Current meas.
PV PV panels Power transformer

1 6
311 m

2 3 4 5
V V V V V
460 m 73 m 72 m 35 m
I
I I I I

1260 kVA 2260 kVA 1260 kVA 1260 kVA

PV PV PV PV

Figure 4.1 – Schematic of the monitored feeder of the EPFL-campus distribution

network.

4.2 Distribution feeder of the EPFL-campus network

In the framework of the NanoTera S3 -Grids Swiss project [28, 107], we installed sen-
sors and PMUs in a feeder of the EPFL-campus distribution network. The design
and implementation of the measurement infrastructure comprising sensors, PMUs,
telecommunication network and data processing are the result of a joint work of two
EPFL laboratories: the distributed electric system laboratory (DESL) and the computer
communications and applications laboratory 2 (LCA2). The purpose of the project
is to evaluate the accuracy and computational performance of RTSE in a real active
distribution network. We particularly focus on assessing the accuracy of the SE solu-
tion and we compare the results of various algorithms. In Section 4.2.1, we provide the
characteristics of the monitored network as well as of the measurement infrastructure.
Further device characteristics and implementation details can be found in [28]. The
SE results are illustrated in Section 4.2.2.

4.2.1 Network characteristics and measurement infrastructure

The network schematic is depicted in Fig. 4.1. The feeder has a rated voltage equal to
20 kV line-to-line and is composed of 6 buses and 5 three-phase lines. Buses #2 to #5
represent secondary substations. Each substation supplies buildings where there are

86
 4.2. Distribution feeder of the EPFL-campus network

offices, laboratories and a considerable amount of PV panels installed on the building

roofs; thus, this is a typical example of active distribution network. Recently, a 750 kW
- 560 kWh battery has been connected to bus #2 in the context of the SCCER-Furies
Swiss project [108]. The transmission lines are underground cables with cross section
equal to 150 mm2 . The line parameters are R = 0.159 ohm/km, X = 0.113 ohm/km,
B = 84.8 μS/km, and the line lengths are given in Fig. 4.1. We do not have any
information about the mutual coupling between phases; therefore, we assume that
the phases are totally independent. In this case, three-phase SE can be conducted
separately for each phase. The feeder is operated in open-ring configuration where
bus #6 is a zero-injection bus.

We installed combined voltage/current sensors at the root of the feeder (bus #1) and
at the medium-voltage terminals of the transformers in every secondary substation
(buses #2 to #5), as shown in Fig. 4.1. A sensor is installed in each of the three phases.
The voltage sensors consist of 0.1-class capacity dividers and the current sensors
consist of 0.5-class Rogowsky coils. The rated voltage of voltage sensors is the line-to-
line rated voltage of the grid, i.e., 20 kV. The rated current of the sensors at bus #1 is
200 A and the one of the sensors at buses #2 to #5 is 40 A.

The signals that exit from the sensors at a given substation are provided to a PMU
through shielded cables. The PMU synchrophasor-estimation algorithm is the en-
hanced interpolated discrete-Fourier-transform proposed in [97]. This algorithm
adopts an acquisition time-window containing 3 periods of a signal at the nominal
power-system frequency (e.g., 60 ms at 50 Hz) and is characterized by an average mea-
surement reporting-latency of 44 ms. The PMU is characterized by a maximum TVE
of 0.02 % (defined in [47]) that is maintained both in steady-state and in most dy-
namic conditions, irrespectively of the harmonic-distortion level. The synchrophasor-
estimation algorithm is deployed on the field-programmable gate array (FPGA) of a
National Instruments CompactRIO 9068. The CPU is used to perform other tasks, such
as the streaming of the estimated quantities. The PMU synchronization is achieved
by means of a GPS signal. The phasor estimates are generated every 20 ms and user-
datagram-protocol (UDP) datagrams are encapsulated according to IEEE C37.118.2-
2011 [109].

Single-pair high-speed digital subscriber line (SHDSL) modems transmit these data-
grams to a server room over a secured and dedicated communication network com-
posed of twisted pair cables (originally installed for telephony). Then, the datagrams
are sent through optical fibers to a phasor data concentrator (PDC) located in another
EPFL building.

The PDC and state estimator are hosted on a dedicated workstation and are coded
in LabVIEW (the LWLS and DKF-heur algorithms are implemented in the real-time
environment). This workstation is GPS-synchronized. The PDC is fully developed by

87
Chapter 4. Results of PMU-based state estimation of real networks

the DESL of EPFL and has two main functionalities: data aggregation that enables
to gather data coming from multiple PMUs into phase-aligned datasets, and data
pushing that enables to forward these datasets to further applications, such as SE. A
thorough description of the architecture and performance of this PDC is given in [110].
We implement the so-called absolute-time data-pushing logic, i.e., the time-aligned
datasets are reported at a constant reporting rate that is equal to the PMU reporting
rate (50 times-per-second). By using such a logic, the latency variations introduced by
the telecommunication media are completely nullified.

Once a complete set of measurements is available, it is pushed to the SE algorithm that

runs at 50 estimates-per-second. The measurements and the SE solution are saved
locally and then transferred on a database. We made the data publicly available on
[107].

The full set of real-time measurements is composed of 5 three-phase voltage pha-

sors and 5 three-phase current-injection phasors, i.e., 60 measurements. If the zero-
injections at bus #6 are treated as virtual measurements, the state estimator has to
process 66 measurements and provides 36 state variables, resulting in a redundancy
level of 1.8. For the case of a LWLS that employs the method proposed in Chapter 3,
there are 60 measurements and 30 state variables, resulting in a redundancy level of 2.

Derivation of the measurement-noise covariance matrix

The procedure to derive the diagonal elements of the measurement-noise covariance

matrix R (used by LWLS and DKF) is based on the concepts introduced in Section 1.1.2
and is explained in the following. We disregard the contribution of random noise as we
observed that the noises on the measured magnitudes and phase-angles are about a
order of magnitude smaller than the sensor uncertainties. The latter are predominant
also compared to the TVE value of PMUs; therefore, we neglect the PMU uncertainty
in the computation of R. However, when particularly fast and large voltage/current
variations occur, such as during faults, the PMU error may be not negligible, as we
will observe in Fig. 4.9. Under these assumptions, the measurement variances are
evaluated from the limits of ratio error and phase displacement defined by the sensors’
accuracy classes, which belong to type B standard uncertainties. We assume a normal
distribution and we divide these limits by a coverage factor equal to 3 in order to
obtain the standard deviations. The final step consists in the projection from polar to
rectangular coordinates by using (1.7) and (1.8), because linear SE uses measurements
in rectangular coordinates (see Section 1.1). More details are given here below.

The limits of ratio error and phase displacement for voltage and current sensors are
plotted as a function of the measurand in Fig. 4.2. For the voltage sensors, the standard
IEC-61869-3 [46] specifies the limits between 80 % to 120 % of rated voltage, as shown

88
 4.2. Distribution feeder of the EPFL-campus network

2 3
10 10
Limit of ratio error [% of measurand]

Limit of phase displacement [mrad]

0.1-class 0.1-class
0.2-class 0.2-class
0.5-class 0.5-class
IEC 61869-3 IEC 61869-3
10 1 10 2

0 1
10 10

-1 0
10 10
0.2 10 20 30 40 50 60 70 80 90 100 110 120 0.2 10 20 30 40 50 60 70 80 90 100 110 120
Measurand [% of rated voltage] Measurand [% of rated voltage]

(a) Voltage sensors.

10 2 10 3
Limit of ratio error [% of measurand]

Limit of phase displacement [mrad]

0.1-class 0.1-class
0.2-class 0.2-class
0.5-class 0.5-class
IEC 61869-2 IEC 61869-2
10 1 10 2

0 1
10 10

10 -1 10 0
0.2 10 20 30 40 50 60 70 80 90 100 110 120 0.2 10 20 30 40 50 60 70 80 90 100 110 120
Measurand [% of rated current] Measurand [% of rated current]

(b) Current sensors.

Figure 4.2 – Limits of ratio error and phase displacement of the sensors (classes 0.1,
0.2, 0.5).

89
Chapter 4. Results of PMU-based state estimation of real networks

in Fig. 4.2a. We define also the limits below 80 % of rated voltage in order to run
SE even in abnormal conditions, such as in case of deep voltage sags; in particular,
we define the limits at 0.2 % and 5 % of rated voltage and we linearly interpolate the
values in between. If the voltage measurement is between 0 and 0.2 % of rated voltage,
we assume the limit to remain equal, in absolute value, to the one at 0.2 % of rated
voltage. For the current sensors, the standard IEC-61869-2 [45] specifies the limits at
5 %, 20 %, 100 % and 120 % of rated current and we added the limits at 0.2 % and 1 %, as
shown in Fig. 4.2b. Then, we linearly interpolate the values in between. If the current
measurement is between 0 and 0.2 % of rated current, we assume the limit to remain
equal, in absolute value, to the one at 0.2 % of rated current.

For example, let us consider a 0.1-class current sensor with rated current equal to 100 A
and a current-magnitude measurement equal to 20 A (i.e., 20 % of the rated current).
Considering Fig. 4.2b, the corresponding limits of ratio error and phase displacement
are 0.2 % of 20 A (i.e., 0.04 A) and 2.4 mrad, respectively. The standard-deviations are
obtained by dividing these values by a coverage factor equal to 3. Then, we project the
uncertainties from polar to rectangular coordinates by using (1.7) and (1.8). Finally,
the measurement variances are the squared standard deviations.

The zero-injections at bus #6 are treated in different ways depending on the SE algo-
rithm: (1) for the LWLS, we use the method proposed in Chapter 3; (2) for the DKF,
zero-injections are virtual measurements (see Section 1.1.1) with variances equal to the
smallest variance of the other real-time measurements; (3) for the LAV, zero-injections
are additional constraints in the optimization problem.

4.2.2 State estimation results

In this section, we use the case study described in Section 4.2.1 in order to compare
the results of the following state estimators:

• LWLS (see Section 1.2);

• LWLS that uses the LNR test to identify and remove bad data, called LWLS - LNR
(see Section 1.2);

• LAV (see Section 1.3);

• DKF that uses the heuristic method proposed in Section 2.4 to estimate Q, called
DKF - heur (see Section 2.4);

• DKF that uses the PECE method proposed in Section 2.5 to estimate Pk|k−1 ,
called DKF - PECE (see Section 2.5).

90
 4.2. Distribution feeder of the EPFL-campus network

We recall that we do not need a phase-angle reference for SE as PMUs directly measure
the phase-angle [31]. However, in the figures where we plot the phase-angle as a func-
tion of time, we adopted a common phase-angle reference just for display purposes.
Indeed, in real networks the power-system frequency is always different from the
PMU reporting-rate (exactly 50 Hz); the changes of the phase-angle measurements are
dominated by this frequency difference making not visible the changes due to network
phenomena (note that for each mHz of difference, the phase-angle measurements
rotate of 6.3 mrad per second). Therefore, we generate a phase-angle reference by
rotating a phasor at each time-step k of a rotation angle θk computed as

(fk − f s )
θk = 2π (4.1)
fs

where fk is the power-system frequency measured by the PMUs (we use the mean
value of the frequencies measured by all the PMUs) and f s is the PMU reporting-rate
(50 Hz). Then, we subtract this phase-angle reference to every phase-angle that we plot
as a function of time (e.g., right-graphs of Fig. 4.3a).

For the above-mentioned reasons, the DKF persistent process-model that we defined
in Section 1.4.3 has to be adapted for the application in real networks. We need to
add a state-transition matrix Ak that rotates the state of an angle θk at every time-step,
which is defined as

cos(θk ) · In/2 −sin(θk ) · In/2

Ak = (4.2)
sin(θk ) · In/2 cos(θk ) · In/2

where In/2 is the identity matrix of dimension n/2 and n is the state dimension. The
DKF process-model becomes

xk = Ak xk−1 + wk . (4.3)

The number of past state estimates used by the DKF-heur to infer Q is N = 50; this
parameter was assessed by means of off-line simulations in a pre-tuning stage, as
explained in Section 2.4.2. For DKF-PECE, the number of past innovations used to
infer Pk|k−1 is N = 5 000 that is the same number employed in Section 2.5. The
LWLS-LNR uses a threshold equal to 4 in the LNR test.

We consider a 100 s time window of measurements taken on November 17th 2014

at 11:12 a.m. (Swiss local time). In this time interval the network is in quasi-static
conditions as there are no particular events taking place.

In Tab. 4.1, we report all the power-injection measurements at time-step 1 of the

considered time window (power measurements are derived from voltage and current
phasors). The convention for positive and negative signs of the power is reported in Fig.

91
Chapter 4. Results of PMU-based state estimation of real networks

Table 4.1 – Power-injection measurements at time-step 1.

Bus # #1 #2 #3 #4 #5
Pinj,a [kW] 94.1 -4.7 -20.3 -41.8 -27.1
Pinj,b [kW] 104.2 -4.6 -19.3 -51.3 -28.0
Pinj,c [kW] 93.4 -3.8 -20.0 -43.8 -25.3
Qinj,a [kVAR] -12.1 2.0 -0.8 -1.0 -0.5
Qinj,b [kVAR] -3.8 0.7 -0.2 -4.2 -3.5
Qinj,c [kVAR] 1.2 1.4 -0.5 -10.4 -3.4

1.1. All the loads absorb active power and the feeder imports about 100 kW per phase
from the upstream grid. The reactive powers are considerably unbalanced among the
three phases. In our case study, the reactive power generated by the cable capacitances
is often equal or greater than the overall reactive power absorbed by the loads; this
occurs in phases a and b, where the feeder behaves as a capacitor exporting reactive
power to the upstream grid. On the contrary, in phase c there is an important reactive
power absorption at bus #4 that forces the feeder to absorb reactive power.

Fig. 4.3 shows the time evolution of the voltage (magnitude and phase-angle) and
powers (active and reactive) at bus #1 in the three phases a, b, c. In Fig. 4.4 we display a
shorter time window (from 30 s to 40 s) in order to highlight fast changes that are not
visible in Fig. 4.3. The high-resolution measurements of PMUs allow the tracking of
very quick dynamics, such as the periodic steps in the active power starting at 36 s,
which are similar to a square wave. It can be seen that the LWLS estimates follow closely
the measured values. Small discrepancies are present due to the ratio error and phase
displacement of the sensors as well as due to line-parameter errors. In Fig. 4.3a we
can observe that the random noise on the measurements is quite small; the standard-
deviation of this noise is below 10−4 pu and 0.1 mrad for the magnitude and phase-
angle, respectively. This is the reason why we decided to neglect the random-noise
contribution in the computation of the measurement-noise variances (see Section
4.2.1).

In Fig. 4.3 we also observe that there is no correlation between the time evolutions of
voltages and power injections, because the power changes are too small to influence
the voltage. This is due to stiffness of this grid that has short lines and high rated
voltage. Essentially, the voltage fluctuations in Fig. 4.3a are caused by phenomena that
are external to this feeder and the relative voltage drop between buses is extremely
small. These observations are supported by Fig. 4.5 that compares the measured and
estimated voltages at all the buses at time-step 1. The measurements are affected by
the sensors’ errors, whereas the voltage estimates look equal at all the buses. Indeed,
the estimated voltage drop and phase-angle separation between buses are in the order

92
 4.2. Distribution feeder of the EPFL-campus network

δ a [rad] at Bus 1
V a [pu] at Bus 1
1.04 0.982
Meas
1.038 LWLS 0.981
1.036 0.98
1.034 0.979
0 20 40 60 80 100 0 20 40 60 80 100
time [s] time [s]

δ [rad] at Bus 1
V b [pu] at Bus 1

1.045 -1.11
1.04
-1.112
1.035
1.03 -1.114

b
0 20 40 60 80 100 0 20 40 60 80 100
time [s] time [s]
V c [pu] at Bus 1

δ c [rad] at Bus 1
1.042 3.082
1.04 3.08
1.038 3.078
1.036 3.076
0 20 40 60 80 100 0 20 40 60 80 100
time [s] time [s]

(a) Voltage magnitude and phase-angle.

105 -5

Qinja [kVAR]
Pinj a [kW]
at Bus 1

at Bus 1
100 -10
95 LWLS -15
Meas
90 -20
0 20 40 60 80 100 0 20 40 60 80 100
time [s] time [s]
115 5
Qinjb [kVAR]
Pinj b [kW]
at Bus 1

at Bus 1

110 0
105 -5
100 -10
0 20 40 60 80 100 0 20 40 60 80 100
time [s] time [s]
100 5
Qinjc [kVAR]
Pinj c [kW]
at Bus 1

at Bus 1

95 0

90 -5
0 20 40 60 80 100 0 20 40 60 80 100
time [s] time [s]

(b) Active- and reactive-power flows.

Figure 4.3 – Time evolution of voltages and powers at bus #1.

of 10−5 pu and 10−5 mrad, respectively, which are not visible. It is worth pointing out
that the differences between measurements and estimates in Fig. 4.5 remain almost
unchanged if we consider another time-step as they are mainly due to systematic
errors of the sensors and marginally to random noise. For example, from Fig. 4.5 we
observe that the difference between the measured and estimated value of V1,b is about
0.003 pu and this difference is kept constant in the entire time window shown in Fig.
4.3a.

As we measure the power injections at every bus, we can evaluate the correctness of the
measurements by computing the sum of the active-power injections at all the buses,
which represents the grid losses. In Fig. 4.6 we show the measured and estimated grid
losses as well as the total reactive power produced by the cable lines. Measured and
estimated values can be considered to be in agreement even if the difference in the
grid losses appears to be significant. Indeed, the measured losses are computed as

93
Chapter 4. Results of PMU-based state estimation of real networks

δ a [rad] at Bus 1
V [pu] at Bus 1

1.0375 0.981
1.037 0.9805
1.0365 LWLS 0.98
Meas
a

1.036 0.9795
30 32 34 36 38 40 30 32 34 36 38 40
time [s] time [s]

(a) Voltage magnitude and phase-angle.

-10

Qinja [kVAR]
104 Meas
Pinj [kW]
at Bus 1

at Bus 1
LWLS -12
102 -14
a

100 -16
30 32 34 36 38 40 30 32 34 36 38 40
time [s] time [s]

(b) Active- and reactive-power flows.

Figure 4.4 – Zoom of Fig. 4.3 from 30 s to 40 s in phase a.

1.045 0.364
Meas
δ [rad]
V a [pu]

LWLS 0.362
1.04
0.36
a

1.035 0.358
1 2 3 4 5 6 1 2 3 4 5 6
Bus Bus
1.04 -1.729
δ b [rad]
V b [pu]

1.038 -1.73
1.036 -1.731
1.034 -1.732
1 2 3 4 5 6 1 2 3 4 5 6
Bus Bus
1.042 2.461
δ [rad]
V c [pu]

1.041 2.46
1.04 2.459
c

1.039 2.458
1 2 3 4 5 6 1 2 3 4 5 6
Bus Bus

Figure 4.5 – Voltage magnitude and phase-angle at time-step 1.

94
 4.2. Distribution feeder of the EPFL-campus network

Grid reactive power [kVAR]

3
Meas -30
Grid losses [kW]

2 LWLS
-35
1
-40
0
-45
-1
0 20 40 60 80 100 0 20 40 60 80 100
time [s] time [s]

Figure 4.6 – Time evolution of the total grid losses and of the total reactive power
produced by the cable lines.

the sum of all the measured power, thus, they contain the sum of the errors of all the
measurements; whereas, the estimated losses are much more accurate, because SE
exploits also the line parameters. Therefore, a difference of few kilowatts is expected.
The estimated grid losses are in the order of a few tens of watts so that this network
can be considered loss-free. On the contrary, it produces a considerable amount of
reactive power (relative to the loads) that is often exported to the upstream grid, as can
be seen also in Fig. 4.3b.

The SE accuracy is easy to be assessed in a simulation environment and it is derived

from the difference between the true state (e.g., obtained from a load-flow compu-
tation) and the estimated one. Obviously, the true state of real networks is unknown
and the measurement residuals are the accuracy indicators that quantify the matching
between measurements and network model. In Section 1.2 we explained how the
magnitudes of the LWLS normalized residuals are used to detect, identify and remove
bad data in the measurement set. The normalized residuals should remain below a
threshold (usually 3 or 4) when there are no bad data and network-model errors.

Fig. 4.7 displays the distributions of the LWLS normalized residuals for the 100 s
time window of Fig. 4.3. The distribution statistics are visualized as box plots1 . The
mean value of the normalized residuals related to some voltage measurements are
largely above 3. An on-site investigation allowed us to figure out that this result was
caused by an improper installation of the sensors. The normalized residuals of the
current measurements are below 3 except for an outlier in the real part of all the
current measurements in phase b. This outlier corresponds to an event occurring at
bus #3 at time-instant 70.66 s (see Figs. 4.3b and 4.6), which is probably due to the
startup of a three-phase load. The power absorbed at bus #3 has a step change and the
PMU synchrophasor-estimation algorithm is affected by error during step changes,
as illustrated in [97]. The current residuals in phase b are the most affected by the
1
Box plots provide a visualization of summary statistics for sample data and contain the following
features. On each box, the central mark is the median, the edges of the box are the 25 % and 75 %
percentiles, the two vertical lines extend to the most extreme data points not considered outliers (that
corresponds to ±2.7 standard deviations and 99.3 % coverage if the data are normally distributed), and
outliers are plotted individually as circles.

95
Chapter 4. Results of PMU-based state estimation of real networks

10 10

r of Eim
re
rN of E

5 5

N
0 0
1_a
1_b

2_a
2_b

3_a
3_b

4_a
4_b

5_a
5_b

6_a
6_b

1_a
1_b

2_a
2_b

3_a
3_b

4_a
4_b

5_a
5_b

6_a
6_b
1_c

2_c

3_c

4_c

5_c

6_c

1_c

2_c

3_c

4_c

5_c

6_c
bus_phase bus_phase

(a) Real and imaginary parts of the voltage measurements.

1.5
3

im
r of Iinj re

rN of Iinj
2
0.5
N

0 0
1_a
1_b

2_a
2_b

3_a
3_b

4_a
4_b

5_a
5_b

6_a
6_b

1_a
1_b

2_a
2_b

3_a
3_b

4_a
4_b

5_a
5_b

6_a
6_b
1_c

2_c

3_c

4_c

5_c

6_c

1_c

2_c

3_c

4_c

5_c

6_c
bus_phase bus_phase

(b) Real and imaginary parts of the current-injection measurements.

Figure 4.7 – Statistics of the distributions of the normalized measurement-residuals of

the LWLS.

measurement errors and we observe large current residuals at every bus although
the step is present only at buses #1 and #3. This behavior is due to the fact that the
current-injection measurements are highly interacting as can be seen by looking at the
elements of the covariance matrix of the residuals, which is defined in (1.17). In the
rows of current measurements, the off-diagonal entries related to the currents in the
other buses are large compared to the diagonal entry. An error in one of the current
measurements provokes large and almost equal residuals for all the current measure-
ments in the same phase. Hence, bad data in the current-injection measurements can
be detected but not identified in this case study.

In Fig. 4.8, we compare the absolute values of the residuals (not normalized) related
to voltage and current-injection measurements obtained at time-step 1 with different
state estimators (LWLS, LWLS-LNR, LAV and DKF-heur). Note that the LWLS-LNR
The LWLS-LNR removes the measurements V3a , V1b , V5b , because their normalized
residuals exceed the selected threshold equal to 4. Note that, for sake of comparison
with the other estimators, we display the LWLS-LNR residuals of V3a , V1b , V5b even
if these measurements are not used. The voltage residuals are similar for every SE
algorithm and the maximum difference is in the order of 0.001 pu. Note that these
residuals are in agreement with Fig. 4.7a (e.g., see V3a,re ). The current residuals of
LWLS, LWLS-LNR and DKF-heur are almost identical, whereas the LAV gives different
results at some buses, but the difference remains small as all the residuals are below
0.1 A.

96
 4.2. Distribution feeder of the EPFL-campus network

×10 -3 ×10 -3
6
5 LWLS 3
LWLS - LNR

res. of E im [pu]
[pu]

4 LAV
DKF - heur. 2
3
re
res. of E

2 1
1
0 0
-1 a b c a b c a b c a b c a b c a b c a b c a b c a b c a b c a b c a b c
1 2 3 4 5 6 1 2 3 4 5 6
bus, phase bus, phase

(a) Real and imaginary parts of the voltage measurements.

0.08
0.06 LWLS
LWLS - LNR 0.06

[A]
res. of Iinjre [A]

LAV
0.04

im
DKF - heur. 0.04

res. of Iinj
0.02 0.02

0 0
a b c a b c a b c a b c a b c a b c a b c a b c a b c a b c a b c a b c
1 2 3 4 5 6 1 2 3 4 5 6
bus, phase bus, phase

(b) Real and imaginary parts of the current-injection measurements.

Figure 4.8 – Absolute value of the measurement residuals (not normalized) at time-
step 1.

LWLS vs. LAV

The accuracy and computational performance of LWLS and LAV estimators have al-
ready been compared in the literature for a coherent measurement set and in presence
of bad data, e.g., in [34]. In this section, we perform this comparison by using real
synchrophasor measurements.

Figs. 4.7a and 4.8a show that both approaches are able to identify the small anomalies
that are present in the voltage measurements. Sometimes, current measurements are
flagged as bad data during load inrushes that generate step changes in the waveforms
acquired by PMUs. It is known that the PMU synchrophasor-estimation algorithm is
affected by error during step changes (as illustrated in [97]), which can generate large
current residuals. Such a case is depicted in Fig. 4.9 that shows the powers at bus #1
during a load inrush occurring at bus #3 in all the three phases. A particularly relevant
mismatch between the measured and estimated values is visible at time 3.94 s. The
normalized residual of all the currents raise up to 6; additionally, they are all very close
due to the high interaction (already discussed in the description of Fig. 4.7b), which
causes the LWLS-LNR to remove some correct measurements. At 3.94 s, the LWLS-LNR
rejects in turn V3a , I1b , I5b , I1a , I1c , V4b . The removal of two constraints related to the
current injections in phase b causes the estimates of these two currents to become
extremely inaccurate (as demonstrated also in [111]). Indeed, the LWLS-LNR current
estimates at buses #1 and #5 are in the order of 400 A, as shown in Fig. 4.10. This is
due to the stiffness of this network that makes practically impossible to accurately

97
Chapter 4. Results of PMU-based state estimation of real networks

120 20

Qinja [kVAR]
Meas
Pinj a [kW]
at Bus 1

at Bus 1
110 LWLS 0
100 -20
90 -40
0 1 2 3 4 5 0 1 2 3 4 5
time [s] time [s]
120 20

Qinjb [kVAR]
Pinj b [kW]
at Bus 1

at Bus 1
110 0

100 -20
0 1 2 3 4 5 0 1 2 3 4 5
time [s] time [s]
110 40

Qinjc [kVAR]
Pinj c [kW]
at Bus 1

at Bus 1
20
100
0
90 -20
0 1 2 3 4 5 0 1 2 3 4 5
time [s] time [s]

Figure 4.9 – Load inrush at bus #3. Time evolution of active- and reactive-power flows
at bus #1.

500 120
LWLS
400 LWLS - LNR 100
[A]
[A]

LAV 80
300
im
re

60
res. of Iinj

res. of Iinj

200
40
100 20
0 0
a b c a b c a b c a b c a b c a b c -20 a b c a b c a b c a b c a b c a b c
-100
1 2 3 4 5 6 1 2 3 4 5 6
bus, phase bus, phase

Figure 4.10 – Measurement residuals (not normalized) at time 3.94 seconds of the
time-window displayed in Fig. 4.9. This is the instant when a perturbation in the power
absorbed at bus #3 occurs.

estimate currents without using their measurements. On the contrary, the residuals of
the LAV increase compared to normal conditions but remain below 1 A, because the
LAV exploits all the available measurements.

The importance of correctly estimate the system state in such particular events is
questionable; indeed, inrushes or similar events causing step perturbations cannot
be tracked by using PMUs, because the PMU synchrophasor estimate is a kind of
“average value” over the acquisition time-window. PMUs are able to accurately track
only events characterized by time-constants that are larger than the acquisition time-
window. However, the above-mentioned LWLS-LNR behavior is undesired and should
be avoided; if small measurement errors can be neglected, a solution consists in
augmenting the threshold of the LNR test.

98
 4.2. Distribution feeder of the EPFL-campus network

1.0405 -0.153

1.04
-0.1535
V a [pu] at Bus 1

δa [rad] at Bus 1
1.0395
-0.154
1.039
Meas -0.1545
1.0385 LWLS
DKF - heur.
DKF - PECE
1.038 -0.155
80 82 84 86 88 90 80 82 84 86 88 90
time [s] time [s]

Figure 4.11 – Quasi-static conditions. Time evolution of the voltage magnitude and
phase-angle at bus #1 in phase a.
V [pu] at Bus 1

1.05
Meas
1.04 LWLS
DKF - heur.
1.03 DKF - PECE
a

1.02
100 105 110 115 120 125 130 135 140
time [s]
V [pu] at Bus 1

1.05
1.04
1.03
b

1.02
117.4 117.5 117.6 117.7 117.8 117.9 118 118.1 118.2 118.3 118.4
time [s]
V c [pu] at Bus 1

1.045

1.04

1.035
131.8 131.9 132 132.1 132.2 132.3 132.4 132.5 132.6
time [s]

Figure 4.12 – Step change. Time evolution of the voltage magnitude at bus #1 in the
three phases a, b, c. Note that we applied different time-scales in the x-axis of the three
graphs.

99
Chapter 4. Results of PMU-based state estimation of real networks

Table 4.2 – Average computational time in milliseconds.

LWLS LWLS-LNR LAV DKF-heur DKF-PECE

0.16 ms 1.8 ms 3.5 ms 0.60 ms 690 ms

LWLS vs. DKF

Fig. 4.11 shows the time evolution of the voltage magnitude and phase-angle at bus #1
in phase a during a time window in which the network is in quasi-static conditions.
The DKF-heur and the DKF-PECE provide similar estimates that follow the same path
of the LWLS estimates, but a filtering effect is clearly visible.

We also consider the most challenging event for the DKF, which is a step change. In
Fig. 4.12, a voltage dip causes multiple step changes in the voltage magnitude in the
three phases. In the plot related to phase a, we show the entire voltage dip, in the plot
related to phase b we show the step change occurring at 117.8 s and in the plot related
to phase c we show a short voltage drop occurring at 132.2 s. It can be seen that the
DKF-PECE is essentially as fast as the LWLS to follow the step change, whereas the
DKF-heur has an evident delay.

These observations are in agreement with the results presented in Section 2.5.3.

Computational performance and latency

The computational times of the considered state estimators averaged over 5 000 time-
steps are reported in Tab. 4.2. As expected, the LWLS is the fastest algorithm, followed
by the DKF-heur. LWLS-LNR and LAV have comparable computational times in the
order of few milliseconds, because the LWLS-LNR, at each time-step, has to compute
the covariance matrix of the residuals and re-estimate the state as many times as the
number of identified bad-data. For this case study, some voltage measurements are
identified as bad data and removed at every time-step (see Fig. 4.7); on average, the
LWLS-LNR rejects 2.8 measurements per time-step. Finally, DKF-PECE is much slower
than the other algorithms; its accuracy in tracking step variations (see Fig. 4.12) is at
the expense of the computational time, as already illustrated in Section 2.5.

The total time-latency of the process is the time interval between the central instant
of the PMU acquisition time-window and the instant the state estimator provides the
estimated state. Therefore, it comprises the contributions of PMU, telecommunica-
tion network, PDC and SE. As every PMU and the data-concentration workstation
(where PDC and SE are implemented) are time-synchronized through the GPS, the
latency of every element of the chain as well as the total latency can be assessed. Fig.

100
 4.3. Sub-transmission network of the city of Lausanne

Figure 4.13 – Cumulative distribution functions of time latencies.

4.13 shows the total latency expressed as cumulative distribution functions of time
differences between the various elements of the process. The PMU data sampling and
synchrophasor estimation (t1 and t2 ) are deterministic as they are performed on an
FPGA (note that t1 is exactly half of the PMU acquisition time-window). The PMU
data encapsulation (t3 ) is performed on a CPU and is the most non-deterministic
contribution to latency. The telecommunication network (t4 ) adds a delay of approxi-
mately 1.5 ms. The pre-processing of the data performed by the PDC introduces an
additional delay (t5 ). Finally, the SE computational time is lower than 1 ms (we recall
that LWLS and DKF-heur are implemented in real-time environment). The measured
total-latency has a mean value of 61 ms and a standard deviation of 1.8 ms. Recently,
the total latency was decreased to about 42 ms by improving t3 and t5 .

4.3 Sub-transmission network of the city of Lausanne

In the context of a joint project between the network operator of Lausanne (Service
Industriels de Lausanne – SiL) and the DESL of EPFL, a portion of the sub-transmission
network of the city of Lausanne has been equipped with PMUs in order to enhance
the real-time situation-awareness as well as to test new fault-location schemes. We
provide the characteristics of the network and of the measurement infrastructure in
Section 4.3.1; then, we present and discuss the results of different SE algorithms in
Section 4.3.2.

4.3.1 Network characteristics and measurement infrastructure

A schematic of the monitored portion of the 125 kV sub-transmission network of the

city of Lausanne is depicted in Fig. 4.14, which is composed of 7 buses and 10 three-

101
Chapter 4. Results of PMU-based state estimation of real networks

Bus Underground cable Line or transformer outside

of the considered network
Load Overhead line
G Generator Power transformer Voltage and current
measurements
s Sending end r Receiving end

220 kV 220 kV
125 kV 125 kV
Bus 2
Bus 7 Bus 1
r r r s s s s s r r s
Line 8 (3.800 km) Line 1 (4.682 km)
G
Line 10 (3.800 km) Line 9 (4.682 km)
Line 7
(2.841 km)
Line 6 r Bus 6 Line 2
(4.291 km) (1.625 km)
r
Line 5 (4.249 km)

Line 4 (1.849 km) Line 3 (1.916 km)

s s r s r s r
Bus 5 Bus 4 Bus 3

Figure 4.14 – Schematic of the monitored portion of the sub-transmission network of

the city of Lausanne.

102
 4.3. Sub-transmission network of the city of Lausanne

Table 4.3 – Line parameters: length L in km, resistance R in Ω/km, reactance X in

Ω/km, and susceptance B in μS/km. The subscripts 0 and 1 stand for zero and positive
sequence, respectively.

L R0 X0 B0 R1 X1 B1
Line #1 4.682 0.217 0.756 14.3 0.112 0.372 15.1
Line #2 1.625 0.168 0.093 63.2 0.051 0.205 63.2
Line #3 1.916 0.168 0.093 63.2 0.051 0.205 63.2
Line #4 1.849 0.185 0.102 60.4 0.051 0.210 60.4
Line #5 4.249 0.177 0.498 57.2 0.061 0.201 57.2
Line #6 4.291 0.168 0.093 63.2 0.051 0.205 63.2
Line #7 2.841 0.226 0.611 57.4 0.064 0.210 57.4
Line #8 3.800 0.420 1.272 1.8 0.159 0.410 2.8
Line #9 4.682 0.217 0.756 14.3 0.112 0.372 15.1
Line #10 3.800 0.420 1.272 1.8 0.159 0.410 2.8

phase lines. The sending- and receiving-ends of the lines (marked as “s” and “r”,
respectively) are chosen based on the convention of Fig. 1.1. The lines mainly consist
of underground cables with two exceptions: (1) lines #8 and #10 are two parallel
overhead-lines, and (2) lines #1 and #9 are two parallel lines split in two sections, i.e.,
they depart from bus #1 as overhead lines and after 3.682 km they go underground
until bus #2. The positive- and zero-sequence line parameters are reported in Tab. 4.3.

Buses #1 and #7 are connected to a higher voltage grid through step-up transformers.
At buses #2 to #6 there are step-down transformers that supply distribution networks.
Moreover, at buses #1, #2 and #7 there are departing lines that link this portion of the
network to the remainder of the 125 kV network. The current-flows in these lines are
measured by PMUs, but these lines are not considered in this chapter. We use the
current-flow measurements in these lines only to derive the power-injection measure-
ments at buses #1, #2 and #7 that are reported in Tab. 4.5. No zero-injection buses are
present in this network.

The monitoring infrastructure is composed of 15 PMUs that receive voltage and current
signals from existing PTs and CTs installed at both ends of each line. The accuracy
classes are specified in Tab. 4.4; the rated voltage of PTs is 125 kV and the rated cur-
rent of CTs is 600 A. The PMUs consist of National Instruments Grid and Automation
Systems [112] that implements the synchrophasor-estimation algorithm presented
in [97] and already illustrated in Section 4.2.1. A line breaker is present at both ends
of each line and the breaker statuses are included in the PMU datagrams. Thus, we
have the real-time knowledge of the network topology that is updated every 20 ms.
The full set of measurements is composed of 20 three-phase voltage phasors and 20

103
Chapter 4. Results of PMU-based state estimation of real networks

Table 4.4 – Accuracy classes of PTs and CTs. The subscripts send and rec refer to the
sending- or receiving-end of the line, respectively.

PTsend CTsend PTrec CTrec

Line #1 0.2 0.2 0.2 0.2
Line #2 0.2 0.5 0.5 0.5
Line #3 0.5 0.5 0.2 0.5
Line #4 0.2 0.5 0.2 0.5
Line #5 0.2 0.2 0.5 0.5
Line #6 0.2 0.5 0.2 0.5
Line #7 0.2 0.2 0.5 0.5
Line #8 0.2 0.2 0.2 0.2
Line #9 0.2 0.2 0.2 0.2
Line #10 0.2 0.2 0.2 0.2

Table 4.5 – Power-injection measurements at time-step 1.

Bus # #1 #2 #3 #4 #5 #6 #7
Pinj,a [MW] 32.214 -18.307 -6.400 -9.108 -8.033 -5.576 15.262
Pinj,b [MW] 32.276 -18.814 -6.221 -8.764 -7.703 -5.274 14.592
Pinj,c [MW] 32.837 -19.287 -6.335 -8.895 -7.913 -5.359 15.047
Qinj,a [MVAR] -3.337 0.343 -1.406 -2.376 0.005 -1.081 1.845
Qinj,b [MVAR] -3.943 0.981 -1.510 -2.477 -0.029 -1.224 2.236
Qinj,c [MVAR] -3.462 -0.079 -1.209 -2.104 0.328 -0.955 1.467

three-phase current-flow phasors, which leads to a redundancy level of 5.7. Such a

high redundancy level was chosen in order to enable many different research studies
and applications. Note that multiple voltage measurements are available at each bus.
Overall, we estimate 42 state variables by using 240 measurements.

Each PMU is connected to a switch through an Ethernet cable and PMU data are
streamed through optical-fiber to the PDC that is integrated in a workstation of the
control center of Service industriels de Lausanne. We adopted the same PDC used
to monitor the EPFL-network feeder, which is described in Section 4.2.1. The state
estimator is implemented in the PDC workstation where a LWLS estimator runs in real-
time. The same considerations on the derivation of the measurement-noise covariance
matrix R for the distribution feeder (see Section 4.2.1) apply also to this case study.

104
 4.3. Sub-transmission network of the city of Lausanne

Table 4.6 – Power-flow measurements at the sending-ends of the lines at time-step 1.

Line # #1 #2 #3 #4 #5 #6 #7 #8 #9 #10
Ps,a [MW] 10.739 3.008 -3.406 -12.526 -6.043 -14.513 -11.596 -0.358 10.611 -0.373
Ps,b [MW] 10.532 2.131 -4.099 -12.862 -6.126 -14.437 -11.411 -0.064 10.450 -0.053
Ps,c [MW] 10.670 2.281 -4.060 -12.977 -6.370 -14.518 -11.729 -0.250 10.937 -0.249
Qs,a [MVAR] -0.105 0.456 -0.450 -2.184 -0.067 -1.569 -0.951 -0.881 -0.519 -0.880
Qs,b [MVAR] -0.770 0.563 -0.421 -2.249 0.058 -1.815 -0.920 -0.944 -0.382 -0.927
Qs,c [MVAR] -0.705 -0.490 -1.182 -2.653 -0.128 -1.669 -1.022 -0.637 -0.446 -0.652

Table 4.7 – Power-flow measurements at the receiving-ends of the lines at time-step 1.

Line # #1 #2 #3 #4 #5 #6 #7 #8 #9 #10
Pr,a [MW] -10.717 -2.994 3.418 12.523 6.042 14.532 -11.618 0.358 -10.598 0.372
Pr,b [MW] -10.513 -2.122 4.097 12.860 6.141 14.464 -11.415 0.058 -10.432 0.069
Pr,c [MW] -10.652 -2.275 4.082 12.974 6.376 14.550 -11.736 0.249 -10.916 0.248
Qr,a [MVAR] -0.266 -0.956 -0.192 1.642 -1.160 0.196 0.079 0.813 0.153 0.836
Qr,b [MVAR] 0.398 -1.088 -0.228 1.728 -1.276 0.456 0.052 0.884 0.020 0.896
Qr,c [MVAR] 0.334 -0.027 0.549 2.124 -1.107 0.299 0.152 0.575 0.076 0.593

4.3.2 State estimation results

In the following, the results of the state estimators listed in Section 4.2.2 are compared
and discussed for the case study described in Section 4.3.1.

The DKF algorithm uses the process model defined in (4.3). The number of past
state estimates used by the DKF-heur to infer Q is N = 100 and the number of past
innovations used by the DKF-PECE to infer Pk|k−1 is N = 5 000. The LWLS-LNR uses a
threshold equal to 4 in the LNR test.

We consider a 100 s time window of measurements taken on August 16th 2014 at 10

a.m. (Swiss local time) where the network is in quasi-static conditions as there are no
particular events taking place.

A snapshot of the grid operating-conditions is presented in Tabs. 4.5, 4.6 and 4.7, which
includes the measurements taken at time-step 1 of power-injections, sending-end
power-flows and receiving-end power-flows, respectively (power measurements are
derived from voltage and current phasors). The sending- and receiving-ends of the
lines are specified in Fig. 4.14 and the conventions for positive and negative signs of
power injections and flows are reported in Fig. 1.1. Note also that the power injections
at buses #1, #2 and #7 are the aggregated powers of both the transformers and the
departing lines that do not belong to the considered network (see Fig. 4.14). In Tab. 4.6,
we can observe the presence of not negligible active-power flow unbalances among the
phases (e.g., almost 1 MW difference between phases a and b of line #2) and between
the two parallel lines #1 and #9 (about 300 kW in phase c). The same unbalances are

105
Chapter 4. Results of PMU-based state estimation of real networks

1.008 -2.842
Meas

δ a [rad]
V a [pu]

1.006 LWLS -2.844

1.004 -2.846
1.002 -2.848
1 2 3 4 5 6 7 1 2 3 4 5 6 7
Bus Bus
1.006 1.348

δ [rad]
V b [pu]

1.004 1.346
1.002 1.344

b
1 1.342
1 2 3 4 5 6 7 1 2 3 4 5 6 7
Bus Bus
1.008 -0.746

δ c [rad]
V c [pu]

1.006 -0.748
1.004 -0.75
1.002 -0.752
1 2 3 4 5 6 7 1 2 3 4 5 6 7
Bus Bus

Figure 4.15 – Voltage magnitude and phase-angle at time-step 1.

present also at the receiving-ends, which proves that they are not due to measurement
errors. This phenomenon is even more evident in the reactive-power flows (e.g., almost
1 MVAR difference between phases a and c of line #2). These considerations underline
the importance of a three-phase state estimator. Comparing Tabs. 4.6 and 4.7, it can be
seen that sending- and receiving-end active-power flows are consistent, which allows
us to qualitatively presume that no gross measurement-errors are present.

Fig. 4.15 shows the voltage measurements and LWLS estimates at all the buses at
time-step 1. The relative voltage drop between buses is quite small as all the voltages
lie in a range of 0.001 pu; this is justified by the low reactive power flows (see Tabs.
4.6 and 4.7). On the contrary, higher active-power flows result in larger phase-angle
separations in the order of a few milliradians. However, both voltage magnitude and
phase-angle variations are small due to the stiffness of this grid. As for the distribution
feeder case in Section 4.2, the mismatch between measured and estimated values is
constant in time, because it is due to the systematic errors of the PTs. The random
measurement-noise is again very small and can be disregarded.

From Fig. 4.14, we can notice that power-flow measurements are taken in all the lines
departing from each bus (even in the interconnection lines departing from buses #1,
#2, #7, which are external to the considered network ). Therefore, we can compute the
sum of all the active-power flows, which represents the grid losses. In Fig. 4.16 we show
the measured and estimated grid losses as well as the total reactive power produced by
the transmission lines. Considering that the measured values contains the sum of the
errors of all the measurements, the estimated and measured values can be considered
to be in good agreement. Therefore, this result quantifies the qualitative consideration
regarding the absence of gross measurement-errors that we drew by comparing Tabs.
4.6 and 4.7. The total grid losses (about 100 kW) are very small compared to active-
power flows, whereas the reactive power produced by the transmission lines is much
larger than the reactive-power flows as they mainly consist of underground cables.

106
 4.3. Sub-transmission network of the city of Lausanne

4
×10

Grid reactive power [kVAR]

350 -1.79
Meas
Grid losses [kW]

300 LWLS -1.795

250
-1.8
200
150 -1.805

100 -1.81
0 2 4 6 8 10 0 2 4 6 8 10
time [s] time [s]

Figure 4.16 – Time evolution of the total grid losses and of the total reactive power
produced by the transmission lines.

Fig. 4.17 displays the distributions of the LWLS normalized residuals for the considered
100 s time window. All the voltage measurements can be considered to be correct as
their normalized residuals are below the selected threshold for bad-data identification,
which is equal to 4. However, the majority of the normalized residuals related to
current-flow measurements (both at the sending- and receiving-ends) are much larger
than 4 and reach values up to 40. We can further comment that the high normalized
residuals are not outliers that occurs once in a while, because they have median values
up to 25. The classic cause of normalized residuals that are systematically large is
network-model errors. As we are sure about the correctness of the topology, the main
suspects are the line parameters. We exclude measurement errors as possible cause,
because sending- and receiving-end current-flow measurements are coherent (see
Tabs. 4.6 and 4.7) and the total power balance derived from the measurements is in
agreement with the estimated one. Investigations on this theme are ongoing.

However, although these normalized residuals may appear quite high, we should
remind that current-flow measurements are very sensitive to line-parameter errors.
Indeed, if we carry out the same simulation by using only voltage and current-injection
measurements (the current injections can be computed from the current-flows by
means of the Kirchhoff’s law, because we measure all the current-flows departing from
each bus), all the normalized residuals fall below 2. In addition, the difference between
measured and estimated currents expressed as a percentage of the current magnitudes
is always below 6 %, which is acceptable.

In Fig. 4.18, we compare the distributions of the absolute values of the residuals
related to the real part of the sending-end current-flow measurements (not normalized)
obtained with different state estimators (LWLS, LWLS-LNR, LAV). We do not show the
residuals related to the imaginary parts and to the receiving-end current-flows as they
are almost identical. We do not show the DKF residuals neither as they are very similar
to the LWLS ones. The LWLS-LNR always removes some current measurements so
that the corresponding residuals should not exist, but, for sake of comparison with the
other estimators, in Fig. 4.18 we show the difference between the estimated values and
the measurements.

107
 rN of Ir rN of Is rN of E
re re re

108
0
0.5
1
1.5

0
10
20
30
40

0
10
20
30
1_a 1_a 1_a
1_b 1_b 1_b
1_c 1_c
2_a 2_a 1_c
2_b 2_b 2_a
2_c 2_c
2_b

of the LWLS.
3_a 3_a
3_b 3_b 2_c
3_c 3_c
3_a
4_a 4_a
4_b 4_b 3_b
4_c 4_c 3_c
5_a 5_a
5_b 5_b 4_a
5_c 5_c 4_b
6_a 6_a
6_b 6_b 4_c
bus_phase

line_phase
line_phase
6_c 6_c 5_a
7_a 7_a
5_b
7_b 7_b
7_c 7_c 5_c
8_a 8_a
6_a
8_b 8_b
8_c 8_c 6_b
9_a 9_a 6_c
9_b 9_b
9_c 9_c 7_a
10_a 10_a 7_b
10_b 10_b
10_c 10_c 7_c

N
r of E
N im
N
r of Is im
0
0.5
1
1.5

r of Ir im
1_a
0
10
20
30
40

0
10
20
1_a 30 1_a 1_b
1_b 1_b
1_c
1_c 1_c
2_a 2_a 2_a
2_b 2_b
2_c 2_b
2_c
3_a 3_a 2_c
3_b 3_b
3_c 3_c 3_a
4_a 4_a 3_b
4_b 4_b
4_c 4_c 3_c
5_a 5_a 4_a
(a) Real and imaginary parts of the voltage measurements.

5_b 5_b
5_c 5_c 4_b
Chapter 4. Results of PMU-based state estimation of real networks

6_a 6_a 4_c

bus_phase

6_b 6_b
5_a
line_phase

6_c
line_phase

6_c
7_a 7_a 5_b
7_b 7_b
(b) Real and imaginary parts of the sending-end current-flow measurements.

5_c
(c) Real and imaginary parts of the receiving-end current-flow measurements.

7_c 7_c
8_a 8_a 6_a
8_b 8_b
8_c 8_c 6_b
9_a 9_a
6_c
9_b 9_b
9_c 9_c 7_a
10_a 10_a
7_b
10_b 10_b
10_c 10_c 7_c
Figure 4.17 – Statistics of the distributions of the normalized measurement-residuals
 res. of Isre [A] res. of Isre [A] res. of Isre [A]

0
10
20
30
40
50
60
0
10
20
30
40
50
60
0
10
20
30
40
50
60

1_a 1_a 1_a

1_b 1_b 1_b

1_c 1_c 1_c

flow measurements.
2_a 2_a 2_a

2_b 2_b 2_b

2_c 2_c 2_c

3_a 3_a 3_a

3_b 3_b 3_b

3_c 3_c 3_c

4_a 4_a 4_a

4_b 4_b 4_b

4_c 4_c 4_c

5_a 5_a 5_a

5_b 5_b 5_b

5_c 5_c 5_c

(c) LAV.
6_a 6_a 6_a
(a) LWLS.

line_phase
line_phase
line_phase

6_b 6_b 6_b

(b) LWLS-LNR.
6_c 6_c 6_c

7_a 7_a 7_a

7_b 7_b 7_b

7_c 7_c 7_c

8_a 8_a 8_a

8_b 8_b 8_b

8_c 8_c 8_c

9_a 9_a 9_a

9_b 9_b 9_b

9_c 9_c 9_c

10_a 10_a 10_a

10_b 10_b 10_b

10_c 10_c 10_c

109
4.3. Sub-transmission network of the city of Lausanne

Figure 4.18 – Statistics of the distributions of the residuals of the sending-end current-
Chapter 4. Results of PMU-based state estimation of real networks

The residuals of the LWLS and LAV are always small: below 17 A and 11 A, respectively.
Concerning the LWLS-LNR, the residuals with large mean value are related to the
measurements that are most frequently rejected (e.g., Is6a ). The residuals of some
measurements, e.g., Is4c , have a small median value but many outliers that raise up to
60 A; it means that these measurements are rejected only in few time-steps, but when
they are rejected the difference between measured and estimated value becomes large.
On average, the LWLS-LNR rejects 17.2 current measurements per time-steps. However,
these measurements should not be rejected, because, most likely, the reason of the
high residuals is line-parameter errors. This situation leads the LAV and LWLS-LNR to
provide significantly different results, as shown already in Fig. 4.10 for the distribution
feeder case. As already mentioned, investigations on this subject are still ongoing.

LWLS vs. DKF

Here below we provide the comparison among LWLS, DKF-heur and DKF-PECE in
quasi-static conditions and when a step change in the system state occurs.

Fig. 4.19 shows the time evolution of the voltage magnitude and phase-angle at bus #1
in phase a during a time window in which the network is in quasi-static conditions.
We can notice the filtering capability of the two DKF algorithms with respect to the
LWLS. The DKF-heur follow very closely the LWLS estimates, whereas the DKF-PECE
moves away from the LWLS estimates for short time intervals.

In Fig. 4.20, a three-phase fault (due to a failure of a breaker placed externally to our
network) causes a significant voltage dip of about 30 %. The left-graph shows the
entire voltage dip and the right-graph shows only the initial voltage drop. Note that
the voltage magnitude measurement provided by the PMU takes 5 time-steps to go
from 1 pu to 0.8 pu, whereas it takes only 3 time-steps in Fig. 4.12. The reason is that
in Fig. 4.12 the PMUs use the synchrophasor-estimation algorithm presented in [97]
that has a 60 ms acquisition time-window, whereas the PMUs used another algorithm
with an acquisition time-window of 100 ms when this three-phase fault occurred. Only
recently the PMUs in the network of the city of Lausanne have been upgraded with
the new algorithm that has a shorter acquisition time-window. As a consequence, the
smoother PMU response to the step facilitates the DKF-heur that has enough time to
increase Q.

The estimates of the DKF-PECE and LWLS are overlapping, whereas the DKF-heur has
a small delay, although its response can be considered quite fast as well. The DKF-heur
is able to inflate the value of Q when this large step occurs, as shown in Fig. 4.21 for
the diagonal element related to the real part of the voltage at bus #1 in phase a. After
the step, the estimated Q comes back to its previous values.

110
 4.3. Sub-transmission network of the city of Lausanne

1.0065 0.122
0.1218
1.006
V [pu] at Bus 1

δ [rad] at Bus 1
0.1216
0.1214
1.0055
0.1212
a

Meas 0.121

a
1.005
LWLS
DKF - heur. 0.1208
DKF - PECE
1.0045 0.1206
110 112 114 116 118 120 110 112 114 116 118 120
time [s] time [s]

Figure 4.19 – Quasi-static conditions. Time evolution of the voltage magnitude and
phase-angle at bus #1 in phase a.

1.05 1.05

1 1

0.95
V a [pu] at Bus 1

V [pu] at Bus 1

0.95
0.9
0.9
0.85
a

0.85
0.8
Meas Meas
LWLS 0.8 LWLS
0.75 DKF - heur. DKF - heur.
DKF - PECE DKF - PECE
0.7 0.75
126.5 127 127.5 128 128.5 129 129.5 127.78 127.8 127.82 127.84 127.86 127.88 127.9 127.92 127.94 127.96
time [s] time [s]

Figure 4.20 – Step change. Time evolution of the voltage magnitude at bus #1 in phase
a. The left-graph shows the entire voltage dip; the right-graph shows only the initial
voltage drop.

10 0
re
Q diagonal entry of V
at Bus 1 in Phase a

10 -5

10 -10
120 125 130 135 140 145 150
time [s]

Figure 4.21 – Step change. Time evolution of the element of Q used by the DKF-heur
corresponding to the real part of the voltage at bus #1 in phase a.

111
Chapter 4. Results of PMU-based state estimation of real networks

Table 4.8 – Average computational time in milliseconds.

LWLS LWLS-LNR LAV DKF-heur DKF-PECE

0.52 ms 46 ms 16 ms 2.0 ms 1330 ms

Computational performance and latency

The computational times of the considered state estimators averaged over 5 000 time-
steps are reported in Tab. 4.8. LWLS and DKF-heur are the fastest algorithms, whereas
DKF-PECE is largely the slowest. Unlike for the distribution feeder, the LWLS-LNR
computational time becomes higher than the LAV one, because the former rejects a
large number of current measurements per time-steps (17.2 on average), thus, it needs
to re-estimate the state many times.

The time-latency of the process is almost the same of the distribution feeder presented
in Section 4.2. Indeed, the elements of the SE process are the same except for the
telecommunication network that in this case is totally composed of optic fiber. How-
ever, the latency of the telecommunication network is small (a few milliseconds) and
does not influence significantly the total latency. The measured total latency has a
mean value of about 41 ms and a standard deviation of 1 ms (we recall that the LWLS is
implemented in real-time environment).

4.4 Conclusions
We presented the results of RTSE implemented in two real power systems consisting of
a distribution feeder and a sub-transmission network. Network observability and a
high redundancy level are achieved by using exclusively synchrophasor measurements.

First, we proved the feasibility of running RTSE at 50 estimates-per-second with a

total time-latency below 70 ms. Note that this is valid for networks of similar size
and for measurement infrastructures with the same latency performance. For in-
stance, the latency can significantly change for networks of much larger size (as the SE
computational-time increases) and if less deterministic telecommunication networks
are employed (e.g., not dedicated wireless networks). In case of large-size networks,
parallel state estimators could be computed for smaller sub-networks.

Second, we validated the theoretical findings of Chapters 1, 2, 3 via real-scale experi-

ments. We compared the state estimates and the computational times of LWLS, LAV
and DKF.

The filtering feature of the DKF that uses the Q assessment method proposed in Section

112
 4.4. Conclusions

2.4.1 (called DKF-heur) is clearly visible by comparing its estimates with the LWLS
ones. However, we noticed that the measurement noise contained in synchrophasor
measurements is much lower than the systematic errors of the sensors. As the sensors’
calibration sheets were not available, we could not compensate these systematic errors;
consequently, the assumption made in Chapter 2 according to which systematic errors
in the measurements are negligible is violated. Therefore, we cannot claim that the
DKF-heur leads to an accuracy improvement in these case studies.

We also verified that the DKF based on the PECE method (called DKF-PECE) is able to
immediately track the system state during steps, whereas the DKF-heur has a delay
of a few time-steps. However, the DKF-PECE is computationally expensive, whereas
DKF-heur is suitable for real-time implementation.

For the distribution-feeder case, SE enabled the identification of an incorrect installa-

tions of the sensors. For the sub-transmission network case, it was interesting to ob-
serve that many LWLS normalized residuals systematically exceed the selected thresh-
old for bad-data identification when current-flow measurements are used, whereas
they remain below the threshold when only current-injections are used. The residuals
of current-flow measurements are sensitive to line-parameter errors that are the most
probable cause of the high normalized residuals. However, investigations are still
ongoing.

113
5 A new method based on real-time
state estimation for fault location
and protective relaying
In this chapter, we intend to prove that PMU-based RTSE exhibits unique accuracy,
refresh rate and time determinism, which satisfy the requirements of fault location and,
potentially, protective relaying. In this respect, we propose a new method based on
LWLS-SE for fault detection and faulted-line identification. The proposed technique is
validated by using a RTS where we model a distribution feeder and the PMU devices.
Tests are conducted in order to validate the proposed method for many different scenar-
ios: grounded and isolated neutral, different fault types (symmetric and asymmetric),
low- and high-impedance faults occurring at different locations, passive and active
networks. The accuracy and time-latency performances of the proposed method are
evaluated.

5.1 Introduction
The content of this chapter is based on the reference [113] and is a shared contribution
with the PhD dissertation of Marco Pignati that is the other co-main author of the
work. In the introduction of this dissertation, we illustrated that the SE outcome is
used by several control functions in existing EMSs. The high-refresh rate and low-
latency of PMU-based RTSE can also enable the development of new protection and
fault-location schemes exploiting the state estimates. In common practice, protective
relaying and fault location are separated functions that have different time require-
ments: the former is a real-time process, whereas the latter is executed off-line after
the fault clearing.

Nowadays, PMUs play an important role in several aspects of the operation of trans-
mission networks [19]. Protection schemes have not undergone major changes for
many years, but recent studies have shown that PMUs can be employed for protective

115
Chapter 5. A new method based on real-time state estimation for fault location
and protective relaying

relaying [114, 115]. PMU data can also be used for adaptive setting of distance relays,
as discussed in [116]. Fault-location methods are historically divided in two main
categories: impedance-based techniques (e.g., [117]) and algorithms based on the
traveling waves originated by the fault transient (e.g., [118]). These approaches can
be further classified in one-terminal, two-terminals, multi-terminals and wide-area
methods depending on the number of measurement points; additionally, synchronous
or asynchronous measurements can be considered. In the last decade, an increasing
number of PMU-based fault-location algorithms have been proposed (e.g., [119, 120]).
In [121], the authors propose a wide-area fault locator using wavelet where bad-data in
the measurement set are eliminated by means of LAV-SE. To the best of our knowledge,
there are only two works that use PMU-based SE to directly infer the fault location, i.e.,
[122, 123]. In [122], phasor measurements provided by PMUs at the two terminals of
the faulted line are used for SE. The state vector is augmented by the voltage at the
faulted point and by the fault distance. The approach proposed in [123] is similar to
that in [122], the difference lying in the fact that the state vector is composed only of
the voltage at the faulted point and by the fault distance.

Distribution networks have usually a radial structure and are equipped with a limited
number of breakers and measurement devices. For instance, a typical configuration
of medium-voltage distribution networks consists in having a breaker at the root of
each feeder, a PT in the primary substation and CTs on the secondary windings of
the high-to-medium voltage transformers and, sometimes, also at the root of each
feeder. Therefore, current practices related to protective relaying typically consist
of basic schemes that change from country to country and are different depending
upon the type of fault and network, as illustrated in [124]. Fault location usually
consist in on-site investigation by repair crew. The increasing integration of distributed
generation (DG) in distribution network is leading to substantial changes in the grid
operation. In particular, fault management might experience major transformations
as a gradual deployment of monitoring infrastructures is expected in the coming years
[124]. Accurate fault location in distribution networks can expedite service restoration,
thus reducing the duration of power outages. In the literature, a number of methods
for fault location in distribution networks have been proposed, which are based on
impedance measurements [125], traveling waves [126] or phasor measurements [127].
[124]

In this chapter, we present a new method for fault detection and faulted-line identifica-
tion that exploits the results of LWLS-based RTSE using synchrophasor measurements.
The formulation of the proposed method does not change regardless of the type of
network (radial or meshed as well as with grounded, compensated or isolated neutral),
the type of fault (symmetric or asymmetric), the fault impedance and the presence of
DG. This method is able to locate the faulted component in real-time; therefore, it may
theoretically combine the execution of protection and fault-location functions that are
usually separated processes.

116
 5.2. The proposed method for fault detection and faulted-line identification

In the following sections, we describe the new above-mentioned methodology and we

present a performance assessment in terms of fault-location accuracy and time-latency.
We will show that this method is suitable for the time requirements of protective
relaying.

5.2 The proposed method for fault detection and faulted-line

identification
Hereafter in this chapter, the term fault location is used with the meaning of faulted-
line identification. The proposed fault detection and location method relies on the
following assumptions:

1. Exact knowledge of the network model (topology and line parameters), i.e., H is
known with no errors;

2. A PMU is installed at every bus and measures voltage and current-injection

synchrophasors;

3. The presence of bad data in the measurement set is not considered as the proba-
bility of having a fault and bad data simultaneously is assumed to be negligible.

Observations. Later in this section, we will see that Assumption 2 is crucial to ensure
the network observability. Moreover, we assume to measure the current injections and
not the current flows in order to limit the number of measurements that is already
quite large for distribution networks.
We are aware of the fact that Assumption 1 is not perfectly valid in real systems and
large network-model errors can degrade the fault location capabilities. However, it
is reasonable to assume that the network topology and parameters are known with
good accuracy. The topology can be constructed at the data concentration point by
collecting the breaker/switch statuses that are streamed by PMUs as Boolean variables.
The parameters of the network components usually have a standard configuration
with known electrical parameters.
An extensive analysis with respect to network-model errors and a performance assess-
ment including also current flows will be the subjects of future research.

The concept behind the proposed fault-location method is that a fault on a line modi-
fies the network topology and can be modeled as an additional bus that absorbs the
fault current. We use the LWLS for this application as low computational time is an
important requirement; the DKF needs to employ the computational expensive PECE
method presented in Section 2.5 in order to rapidly track step-varying events like faults.
The method is explained in detail in the following.

117
Chapter 5. A new method based on real-time state estimation for fault location
and protective relaying

Let us consider a power system with n buses and m lines. We can define m LWLS
estimators that utilize the same measurement vector z, but each one of them uses
a different topology. The difference in the topology consists in the position of an
additional bus that we call virtual bus: the state estimator #u (u = 1, ..., m) has the
virtual bus in the middle of line #u. The state vector of each estimator, previously
defined in (1.1), is augmented by the state variables related to the virtual bus as follows:

xaug = [V1a,b,c
re
, ..., Vna,b,c
re
a,b,c
, Vn+1 re
, V1a,b,c
im
, ..., Vna,b,c
im
a,b,c T
, Vn+1 im
] (5.1)

a,b,c a,b,c
where Vn+1 re
and Vn+1 im
are respectively the real and imaginary parts of the voltage at
the virtual bus that is placed in a different line for each state estimator. The measure-
ment matrix H of each state estimator is modified accordingly.

Now it is clear why the network is observable for every state estimator only if we have a
PMU at every bus (i.e., if Assumption 2 holds). Indeed, let us consider a generic line #u
between buses #i and #h. If bus #i is not equipped with PMUs, the state estimator with
the virtual bus on line #u has two consecutive buses (bus #i and the virtual bus) with
no measurements; the observability criterion is not fulfilled for this state estimator.

During normal operating conditions, all the virtual buses absorb no current and the
different topologies do not lead to significant differences in the outputs of the m
state estimators: ∀u, xuaug  xtrue . Therefore, also the objective functions J of all the
estimators are similar (the LWLS objective function is defined in (1.10)).

When a fault occurs in line #f between buses #i and #h, a fault current is drawn from
an unknown position between these two buses. The topology of the state estimator #f
is close to the real topology even if the fault is not located exactly in the middle of the
line. Therefore, it should provide an estimated state close to the true one (xfaug  xtrue ),
whereas the estimated of the other estimators is more affected by the wrong topology
(∀u = f, xuaug = xtrue ). As a consequence, we assume that the faulted-line index
corresponds to the index of the state estimator characterized by the lowest objective
function. The phases of the virtual bus in which the estimated current differs from
zero are the ones affected by the fault; hence, also the fault type is identified (i.e., 3-, 2-
or 1-phase).

The detection of the fault is performed by comparing the mean of the objective func-
tions of the m estimators, called Jmean , calculated at two consecutive time-steps k and
k − 1. When there is no fault, Jmean has a value that is quite stable in time. We observed
that when a fault occurs, Jmean exhibits a sudden increase due to the increase of the
objective functions. A fault is detected when the difference between the values of Jmean
at two consecutive time-steps exceeds a certain threshold ε, i.e.,Jmean,k − Jmean,k−1 > ε
(see Section 5.4.1). A tuning stage is required to estimate the value of this threshold.

118
 5.2. The proposed method for fault detection and faulted-line identification

Start

Run m estimators and compute J1,..., Jm

no
mean(J1,...,Jm)|k
mean(J1,...,Jm)|k
1 >  ?
(k is the time-step index)

yes

Fault detected No fault

Jf = min(J1 ,...,Jm)  f = faulted line index

Estimated currents at virtual bus  fault type

End End

Figure 5.1 – Flowchart of the proposed fault detection and location method.

Further research is needed in order to design an algorithm for the estimation of this
threshold.

A flowchart of the procedure described above is given in Fig. 5.1 and is explained
in the following. For every new PMU-data set at a generic time-step k, we run m
state estimators and we compute their objective functions J1 , ..., Jm . The mean value
of these objective functions is Jmean,k . Then, we calculate the difference between
the values of Jmean at two consecutive time-steps k and k − 1; if it exceeds a certain
threshold ε, a fault is detected. If a fault is detected, the index of the state estimator
with the minimum J corresponds to the faulted-line index. Let us suppose that line #f
is identified as faulted, we infer the fault type from the magnitude of the current
estimates in the virtual bus provided by the state estimator #f .

In summary, the proposed method allows to:

119
Chapter 5. A new method based on real-time state estimation for fault location
and protective relaying

• detect a fault;

• identify the faulted line;

• identify the fault type;

Observations. As long as a PMU is installed at every bus, the proposed method has no
theoretical restrictions related to the type of network (meshed or radial with grounded,
compensated or isolated neutral) and the type of fault (symmetric or asymmetric).
Additionally, this method can be applied in the same way to any scenario, because
it does not need any model of loads/generators and it has no prior knowledge of the
fault type and of the neutral connection. The inputs consist only of the incoming set of
measurements and the network model (topology and line parameters).
Once the faulted line is identified, classic one- or two-terminals fault-location methods
can be employed in order to infer the fault location in the line. Indeed, the estimated
values of current flows at the line terminals are made available by SE.

5.3 Simulation set-up

As stated in the introduction of this chapter, an interesting application of the proposed
fault detection and location method is in distribution networks.Therefore, we evaluate
its performance by building a real-time model of a distribution feeder as well as of the
PMUs.

The considered network is a three-phase medium-voltage distribution feeder located

in the Netherlands and operated by Alliander. It has a rated voltage equal to 10-kV and
it is composed of 18 buses and 17 lines. The network schematic is depicted in Fig. 5.2.
The lines are underground cables with cross sections between 95 and 240 mm2 . The
lengths as well as the zero- and positive-sequence parameters of the cables are given in
Tab. 5.1. The rated power of the medium-to-low voltage transformers at the secondary
substations are given in Tab. 5.2.

A three-phase model of the network is built in SimPowerSystemTM and the simulations

are run by using an Opal-RT RTS. The lines are modeled as PI circuits. The equivalent
circuit of the upstream grid is modeled with a fixed voltage source in series with a
short-circuit impedance Zsc . The value of Zsc is computed by assuming a short-circuit
power of 1 GVA and a resistance-to-reactance ratio Rsc /Xsc = 1/10. Loads are present
at every bus except for bus #1 and are modeled as a star connection of impedances. In
normal operating conditions, they absorb approximately 1/4 of the rated power of the
secondary-substation transformers. In general, the power absorption is not balanced
in the three phases. The real feeder is operated with isolated neutral, but we performed
simulations with both grounded and isolated neutral.

120
 5.3. Simulation set-up

1 km

Zsc 12

11

10
1
18 9

2 17
16 8

15 14 13 7

3 4 5 6
Figure 5.2 – Schematic of the 18-bus distribution feeder.

Table 5.1 – Line parameters: length L in km, resistance R in Ω/km, reactance X in

Ω/km, and susceptance B in μS/km. The subscripts 0 and 1 stand for zero and positive
sequence, respectively.

L R0 X0 B0 R1 X1 B1
L1,2 0.745 1.057 0.910 73.4 0.139 0.075 147.9
L2,3 0.929 1.057 0.910 76.0 0.139 0.075 147.2
L3,4 1.438 0.844 0.197 12.0 0.159 0.087 145.0
L4,5 1.813 1.041 0.855 71.2 0.141 0.076 149.4
L5,6 0.706 0.815 0.100 14.1 0.162 0.089 141.4
L6,7 0.320 0.815 0.100 14.1 0.162 0.089 141.4
L7,8 0.431 0.815 0.100 14.1 0.162 0.089 141.4
L8,9 0.592 0.815 0.100 14.1 0.162 0.089 141.4
L9,10 0.564 0.815 0.100 14.1 0.162 0.089 141.4
L10,11 0.454 0.815 0.100 14.1 0.162 0.089 141.4
L11,12 0.422 1.237 1.354 56.0 0.357 0.082 106.2
L5,13 0.511 1.060 0.920 73.8 0.139 0.075 147.7
L13,14 0.457 1.060 0.920 73.8 0.139 0.075 147.7
L14,15 0.472 1.060 0.920 73.8 0.139 0.075 147.7
L13,16 0.227 1.240 1.380 53.4 0.356 0.082 106.8
L16,17 0.218 1.240 1.380 53.4 0.356 0.082 106.8
L17,18 0.417 1.192 0.975 87.4 0.373 0.088 100.0

121
Chapter 5. A new method based on real-time state estimation for fault location
and protective relaying
Table 5.2 – Rated power of the transformers at the secondary substations.

Rated Power [kVA]

Bus 2 160
Bus 3 250
Bus 4 400
Bus 5 200
Bus 6 250
Bus 7 1000
Bus 8 400
Bus 9 400
Bus 10 400
Bus 11 400
Bus 12 400
Bus 13 400
Bus 14 250
Bus 15 250
Bus 16 250
Bus 17 250
Bus 18 250

We place a PMU at every bus, which measure nodal voltage and current-injection
phasors. The PMU is modeled in the SimPowerSystemTM environment as illustrated
in [128]. The PMU synchrophasor-estimation algorithm is an enhanced interpolated
discrete-Fourier-transform that is described in [97]. The use of simulated PMUs makes
the validation of the proposed method more realistic compared to the common prac-
tice of using the true phasors, especially concerning the time-latency assessment.
Indeed, the PMU response-time during transients is included in the analysis, which
is mainly affected by the acquisition-window length and by the position of the time-
stamp within the window. A comparison of the time evolution of the fault-current
magnitude estimated by the modeled PMU versus the idealized fault-current magni-
tude is given in Fig. 5.3. This PMU is characterized by an acquisition-window length of
3 periods and the time-stamp is centered in this window; consequently, the estimated
magnitude takes 4 time-steps to reach the pre-fault accuracy level. This aspect is taken
into account in the assessment of the total time-latency given in Section 5.4.5.

The simulated PMU introduces only the noise due to the synchrophasor-estimation
algorithm. It is known that the robustness of fault detection and location algorithms
against measurement uncertainties is a crucial aspect. Therefore, we added the errors
of the sensors to the phasor magnitude and phase-angle estimated by the simulated
PMU. We disregard the random noise, because we observed that the PMU measure-
ments taken in the real network are characterized by magnitude and phase-angle

122
 5.3. Simulation set-up

Fault: 1-phase, earthed neutral, 100 Ω

140
Idealized
Estimated by PMU
135

130

125
Imag 1a [A]
120

115

110

105

100
0.46 0.48 0.5 0.52 0.54 0.56 0.58 0.6 0.62
time [s]

Figure 5.3 – Comparison between the fault-current magnitude estimated by the simu-
lated PMU versus the idealized fault-current magnitude.

noises that are about a order of magnitude smaller than the systematic errors of volt-
age and current sensors.

We assume to use measurement PTs at every bus for the voltage measurements. Con-
cerning the current measurements, we assume to use measurement CTs at the sec-
ondary substations (buses #2–#18) and 5P-class protective CTs at the primary sub-
station (bus #1), because at bus #1 we need to measure the fault current. To assess
the fault location performance for different error levels, we simulated the errors of
0.1-class and 0.5-class measurement transformers.

During a single-phase low impedance fault in the network with isolated neutral, we
observed that the voltage in the faulted phase drops to 0.01 pu. In this specific case,
we assume that the error limits of the PTs in the faulted phase are 10 times larger
then those given above. For the 5P-class protective CTs, the standard [45] specifies a
5% limit for the composite error at the maximum current; we assume that this error
corresponds to ±3 % ratio error and ±40 mrad phase-displacement.

The limits of ratio error and phase-displacement for PTs and CTs are given in the
Standards IEC-61869-3 [46] and IEC-61869-2 [45], respectively; in Tab. 5.3, we report
the limits used for the simulations in Section 5.4.

The procedure used to assess the accuracy of the proposed fault-location method is
described in the following:

1. A model corresponding to a certain fault scenario is implemented and run in

SimPowerSystemTM by using an Opal-RT RTS. During a fault, the simulated PMUs

123
Chapter 5. A new method based on real-time state estimation for fault location
and protective relaying
Table 5.3 – Limits of ratio error and phase displacement for PTs and CTs.

Sensor class ratio error phase-displacement

protective CT 5P ±3% ±40 mrad
measurement PT 0.1 ±0.1% ±1.5 mrad
measurement CT 0.1 ±0.1% ±1.5 mrad
measurement PT 0.5 ±0.5% ±6 mrad
measurement CT 0.5 ±0.5% ±9 mrad

estimate the synchrophasors at 50 measurements-per-second;

2. a set of measurements is obtained by adding the errors of the sensors to the

phasor magnitude and phase-angle inferred in step 1. These errors are randomly
generated between the limits specified in Tab. 5.3.

3. The set of measurements computed in step 2 is given to the m state estimators

described in Section 5.2, which return m estimated states. Then, m values of J are
computed and the index of the estimator exhibiting the lowest J coincides with
the index of the inferred faulted-line. The fault-location method is successful if
the inferred faulted line coincides with the real faulted line;

4. Steps 2 and 3 are repeated M = 1000 times in order to simulate many different
combinations of sensor errors. The number of times the fault-location method is
successful is equal to Ms ;

5. The accuracy of the fault location method is represented by the percentage of

successes, which is computed as

Ms
Accuracy = · 100 . (5.2)
M

5.4 Results
The accuracy of the proposed fault location method has been extensively tested and
in this section we present the simulation results. The test scenarios refer to different
combinations of the following factors:

• Network with grounded or isolated neutral;

• Low- or high-impedance faults (1 Ω or 100 Ω);

• Symmetric faults (3-phase) or asymmetric faults (2-phase and 1-phase);

124
 5.4. Results

• Fault location: the fault is simulated in three lines (L4,5 , L9,10 , L13,16 ) at 1/4 or 1/2
of the line length;

• Different class of the measurement transformers (0.1 or 0.5);

• Other factors: different operating conditions including the presence of DG and

fault in a bus.

5.4.1 Fault detection

The fault-detection principle explained in Section 5.2 is proven here below. Fig. 5.4
shows the time evolution of the objective functions of the m = 17 state estimators for
the specific case of a 3-phase 100Ω fault at 1/4 of line L13,16 . The fault occurs between
0.5 and 0.52 seconds. Before the fault, the values of all the objective-functions are
always close to each other; in the time-steps after the fault, a quick separation of the
objective functions provokes a sudden change of the mean of the m objective-functions
computed at two consecutive time-steps. Thus, the difference between the values of
Jmean at two consecutive time-steps exceeds a pre-set threshold and a fault is detected
(see Fig. 5.1). A complete separation of the objective functions occurs three time-steps
after the fault, according to the PMU response-time (see Fig. 5.3).

During the post-fault condition, it is evident that the LWLS with the virtual bus in line
L13,16 maintains the lowest value of J; therefore, the fault location algorithm correctly
identifies the fault in line L13,16 . Note that J of the LWLS with the virtual bus in line
L13,16 has a small increase after the fault as well, because the position of the virtual bus
(1/2 of the line) does not coincide with the fault point (1/4 of the line).

It is worth observing that the identification of a fault in line L13,16 is challenging,

because line L13,16 and its neighbor lines (L5,13 , L13,14 and L16,17 ) are short (218 to 510
meters), and the virtual buses in these lines are all quite close to the fault. Indeed, in
Fig. 5.4 we can see that the objective functions of the respective state estimators are
lower than the others. As a consequence, when large measurement errors are present,
a mis-estimation of the faulted line is more likely to occur. However, this consideration
implies that when the algorithm fails to identify the faulted line, it locates the fault in
one of the adjacent lines, not at the opposite side of the network.

Fig. 5.5 is equivalent to Fig. 5.4, but for a fault at 1/2 of line L13,16 . We observe that J
of the LWLS with the virtual bus in line L13,16 maintains a similar value even after the
fault, because the position of the virtual bus and the fault point coincide; thus, this
state estimator is using the correct topology even during the post-fault condition.

125
Chapter 5. A new method based on real-time state estimation for fault location
and protective relaying
1-phase 100 Ω fault at 1/4 of line L13,16 (grounded neutral)
10 6
LWLS L1,2
LWLS L
2,3
LWLS L3,4
5
10 LWLS L4,5
LWLS L5,6
Objective function J

LWLS L6,7
LWLS L7,8
10 4
LWLS L8,9
LWLS L9,10
LWLS L10,11
10 3 LWLS L11,12
LWLS L5,13
LWLS L13,14
LWLS L14,15
10 2 LWLS L13,16
LWLS L16,17
LWLS L
17,18

10 1
0.44 0.46 0.48 0.5 0.52 0.54 0.56 0.58 0.6 0.62 0.64 0.66 0.68 0.7 0.72 0.74
time [s]

Figure 5.4 – Time evolution of the objective functions of the m = 17 state estimators
when a 3-phase 100Ω fault occurs at 1/4 of line L13,16 . The fault inception is between
0.5 and 0.52 seconds.

1-phase 100 Ω fault at 1/2 of line L13,16 (grounded neutral)

10 6
LWLS L1,2
LWLS L2,3
LWLS L3,4
5
10 LWLS L4,5
LWLS L5,6
Objective function J

LWLS L6,7
4 LWLS L
10 7,8
LWLS L8,9
LWLS L9,10
LWLS L10,11
3 LWLS L
10 11,12
LWLS L5,13
LWLS L13,14
LWLS L14,15
10 2 LWLS L13,16
LWLS L16,17
LWLS L17,18

10 1
0.44 0.46 0.48 0.5 0.52 0.54 0.56 0.58 0.6 0.62 0.64 0.66 0.68 0.7 0.72 0.74
time [s]

Figure 5.5 – Time evolution of the objective functions of the m = 17 state estimators
when a 3-phase 100Ω fault occurs at 1/2 of line L13,16 . The fault inception is between
0.5 and 0.52 seconds.

126
 5.4. Results

Table 5.4 – Accuracy for 3-phase Table 5.5 – Accuracy for 3-phase
1 Ω fault (neutral grounded) 100 Ω fault (neutral grounded)

Sensor’s class Sensor’s class

Fault position Fault position
0.1 0.5 0.1 0.5
1/4 100% 100% 1/4 100% 82%
L4,5 L4,5
1/2 100% 100% 1/2 100% 88%
1/4 100% 100% 1/4 98% 51%
L9,10 L9,10
1/2 100% 100% 1/2 99% 52%
1/4 100% 100% 1/4 71% 24%
L13,16 L13,16
1/2 100% 100% 1/2 97% 27%

Table 5.6 – Accuracy for 2-phase Table 5.7 – Accuracy for 2-phase
1 Ω fault (neutral grounded) 100 Ω fault (neutral grounded)

Sensor’s class Sensor’s class

Fault position Fault position
0.1 0.5 0.1 0.5
1/4 100% 100% 1/4 100% 69%
L4,5 L4,5
1/2 100% 100% 1/2 100% 74%
1/4 100% 100% 1/4 90% 36%
L9,10 L9,10
1/2 100% 100% 1/2 95% 40%
1/4 100% 100% 1/4 63% 13%
L13,16 L13,16
1/2 100% 100% 1/2 84% 15%

Table 5.8 – Accuracy for 2-phase Table 5.9 – Accuracy for 2-phase
1 Ω fault (neutral isolated) 100 Ω fault (neutral isolated)

Sensor’s class Sensor’s class

Fault position Fault position
0.1 0.5 0.1 0.5
1/4 100% 100% 1/4 99% 69%
L4,5 L4,5
1/2 100% 100% 1/2 100% 69%
1/4 100% 100% 1/4 91% 34%
L9,10 L9,10
1/2 100% 100% 1/2 96% 35%
1/4 100% 100% 1/4 64% 12%
L13,16 L13,16
1/2 100% 100% 1/2 85% 13%

127
Chapter 5. A new method based on real-time state estimation for fault location
and protective relaying
Table 5.10 – Accuracy for 1-phase Table 5.11 – Accuracy for 1-phase
1 Ω fault (neutral grounded) 100 Ω fault (neutral grounded)

Sensor’s class Sensor’s class

Fault position Fault position
0.1 0.5 0.1 0.5
1/4 100% 100% 1/4 100% 78%
L4,5 L4,5
1/2 100% 100% 1/2 100% 96%
1/4 100% 100% 1/4 94% 41%
L9,10 L9,10
1/2 100% 100% 1/2 98% 42%
1/4 100% 100% 1/4 89% 23%
L13,16 L13,16
1/2 100% 100% 1/2 96% 26%

Table 5.12 – Accuracy for 1-phase Table 5.13 – Accuracy for 1-phase
1 Ω fault (neutral isolated) 100 Ω fault (neutral isolated)

Sensor’s class Sensor’s class

Fault position Fault position
0.1 0.5 0.1 0.5
1/4 100% 99% 1/4 99% 72%
L4,5 L4,5
1/2 100% 100% 1/2 100% 84%
1/4 100% 97% 1/4 91% 34%
L9,10 L9,10
1/2 100% 99% 1/2 96% 34%
1/4 100% 97% 1/4 79% 14%
L13,16 L13,16
1/2 100% 99% 1/2 91% 18%

128
 5.4. Results

5.4.2 Faulted line identification

The accuracies of the proposed fault-location method for every test scenario are given
in Tabs. 5.4–5.13. The accuracy values were approximated to the lowest integer. The
tables include the results for 0.1-class and 0.5-class measurement transformers (see
Tab. 5.3).

Using 0.1-class sensors, the proposed method guarantees an excellent accuracy for
low-impedance faults and a satisfactory accuracy for most of the high-impedance
faults. Note that high-impedance faults are typically very difficult to locate and require
dedicated techniques. If we increase the errors of the sensors by using 0.5-class sen-
sors, the accuracy remains almost unchanged for low-impedance faults, whereas it is
considerably degraded for high-impedance faults.

The above-mentioned results apply to faults at 1/4 and 1/2 of the line, proving the
fact that the method is still effective even if the virtual bus does not exactly coincides
with the fault position. A fault in proximity of a bus is a critical case, because little
measurement errors can cause the fault locator to fail in identifying the faulted line;
however, we recall that even when the fault locator fails, the fault is very likely to be
identified in one of the lines adjacent to the faulted line. Therefore, we are still able to
limit the estimated fault location to an area that is close to the fault point.

As a conclusion, we can state that the proposed algorithm is able to locate the faulted
line irrespectively of the neutral connection, fault type, fault impedance and fault
position. Moreover, the use of precise sensors enhances considerably the fault location
accuracy.

5.4.3 Faulted bus identification

A fault in a bus can be easily detected and identified by directly using the measurements
of the PMU installed at the faulted substation. For example, let us consider a network
with isolated neutral and let us assume that a 1-phase fault occurs in a bus. The
faulted bus is identified by means of an algorithm that analyze the voltage and current
measurements of the PMU at the faulted bus: the voltage zero-sequence component
becomes suddenly different from zero and the current magnitude in the faulted phase
has a sudden jump of tens of Amperes.

5.4.4 Influence of operating conditions and distributed generation

The SE results, and consequently the proposed fault locator, should be marginally
affected by the network operating conditions. Therefore, we performed numerical
simulations with different operating conditions that include the presence of DG. We

129
Chapter 5. A new method based on real-time state estimation for fault location
and protective relaying

created two scenarios, which are listed here below:

1. The powers absorbed by the loads are tripled;

2. The powers absorbed by the loads are reduced to one tenth and three gener-
ators are placed at buses #4, #10, #17 so that the feeder becomes active. The
generators consist in variable-pitch wind turbines driving 160 kW squirrel-cage
asynchronous generators running at nominal speed. The generator production
is abundantly larger than the load demand and the feeder exports power towards
the upstream grid.

The fault-location accuracies found for these scenarios are very similar to the values in
Tabs. 5.4–5.13.

5.4.5 Computational time and latency

The assessment of the speed of the algorithm in identifying the faulted line is a metric
of interest when comparing fault-location algorithms. In this section, we focus on two
time latencies:

1. the computational time of the proposed method, i.e., the time needed to perform
the procedure depicted in the flowchart of Fig. 5.1;

2. the total time-latency of a real system implementing the proposed method.

First, we just consider the computational time of the procedure described in Section
5.2, which is implemented in MATLAB R
2014b running on an Apple MacBook Pro with
a 2.5 GHz CPU, 16 GB RAM. The average computational time for the specific case of
Section 5.3 is 11.0 ms. It is worth mentioning that the 17 state estimators are computed
in series. Note also that if we consider other case studies, this time is affected by the
network size and the number of measurements.

Second, the total time-latency of a real fault-location system represents the time
between the occurrence of a fault and the identification of its location. In order to
obtain reliable and correct estimates of the post-fault synchrophasors, the PMUs have
to process a dataset of waveforms that does not contain the step that occurs at the
fault inception. To clarify this aspect, Fig. 5.6 shows that when a fault occurs (e.g., in
the gray area), three acquisition windows (W1 , W2 , W3 ) contain a step in the waveform;
therefore, the phasor estimation resulting from these windows is corrupted and not
reliable. We remind that the adopted synchrophasor-estimation algorithm [97] uses a
window containing three periods of the fundamental frequency. W4 is the first post-
fault acquisition window, which contains a step-free waveform so that the respective

130
 5.4. Results

W1
W2
W3
W4

T1 T 2 T 3 T 4

T1 = 0 ÷ 20 ms T2 = 30 ms T3 = 61 ms T4 = 11 ms
Figure 5.6 – Overall time-latency of the proposed method in identifying faults.

synchrophasor estimate is correct. After these considerations, we list the four main
contributions to the total latency in the following:

1. T1 : the time between the fault event and the first sample of window W4 . Depend-
ing on when the fault occurs in the gray area of Fig. 5.6, T1 can vary between 0
and 20 ms;

2. T2 : the time corresponding to half of the acquisition window length used by

the synchrophasor-estimation algorithm. With reference to the synchrophasor-
estimation algorithm described in [97], T2 is equal to 30 ms at 50 Hz;

3. T3 : the time between the center of the acquisition window and the moment the
set of measurements is made available to the state estimators. This time includes
many stages, such as the encapsulation of the synchrophasor packets done in the
PMUs, the latency of the telecommunication network, the decapsulation of the
packets at the data collection point, and the time-alignment of the synchropha-
sors. The latency of the telecommunication network is the key aspect of T3 , which
can be considerably different if we use a wired or a wireless network and if the
network is fully dedicated to the PMU data streaming or it is shared with other
functions. In [28], a detailed analysis on a real feeder equipped with PMUs shows
that T3 is equal to 61 ms for that specific set-up. The communication layer is a
dedicated network consisting of a mix of telephone cables and fiber;

4. T4 : the time needed to execute the procedure depicted in Fig. 5.1. As mentioned
above, the mean value of T4 for the specific case of Section 5.3 is 11 ms.

Therefore, the total latency can vary between 102 and 122 ms, depending on the
instant of the fault inception. This latency is compatible with the requirements of

131
Chapter 5. A new method based on real-time state estimation for fault location
and protective relaying

protective relaying so that we can place the proposed methodology also in the context
of protections. In order to maintain the same time performance also for networks
of much larger size, the network should be split into smaller sub-networks and the
proposed method can be applied separately to each sub-network.

5.5 Conclusions
The new fault detection and faulted-line identification method proposed in this chap-
ter exploits the knowledge of the network state made available by low latency and
high refresh-rate RTSE based on PMUs. This method does not need any model of
loads/generators and it has no prior knowledge of the fault type and of the neutral
connection. We assumed to have a PMU at every bus and we compared the objective
functions of multiple LWLS estimators. The performance evaluation was carried out
for a real distribution feeder modeled in a RTS. The simulated scenarios included
feeders with grounded and isolated neutral, symmetric and asymmetric faults with
low- and high-impedance at different locations. We showed that the proposed method
is able to locate the faulted line with very good accuracy for low- and high-impedance
faults when 0.1-class sensors are used. Whereas, if 0.5-class sensors are employed,
the accuracy is significantly decreased for high-impedance faults. Additionally, this
method is also resilient to the presence of high amount of DG.

Finally, we showed that the total latency of a real system implementing the proposed
method ranges from 102 to 122 ms for the considered network. It is well known that
protections are a crucial power-system component and much more detailed tests have
to be performed in order to claim that this method can be used for protective relaying.
However, we can state that both accuracy and latency performance do match the
requirements of protection schemes and are likely to be further improved in the coming
years thanks to the advancement of PMU and telecommunication technologies.

132
Conclusions
An increasing number of PMUs are being installed in power systems in order to en-
hance the situational awareness and to develop new categories of applications based
on synchrophasor measurements. In the near future, we expect to have more networks
becoming observable just with synchrophasor measurements. PMUs are commonly
associated with transmission systems, but are gaining consideration also in the context
of distribution networks that need fast control schemes due to the presence of highly
volatile DG or for fault location purposes.

In this dissertation, we present and analyze the advantages of using a measurement

infrastructure exclusively composed of PMUs for power-system SE. The latter becomes
a linear and not iterative process that uses a set of phase-aligned synchrophasor mea-
surements and is characterized by high refresh-rate (tens of estimates-per-second)
and sub-second time-latency. We studied and compared different linear SE algorithms
(LWLS, LAV and DKF) with particular focus on the DKF. Indeed, the literature lacks
of a thorough study on the DKF in this specific context. It is important to point out
that we did not only performed tests in a simulation environment, but we also imple-
mented RTSE in real networks that are pervasively equipped with PMUs. Therefore,
we discussed the results of the above-mentioned linear state estimators using real
synchrophasor measurements.

In Chapter 2, we extensively studied the DKF algorithm with the goal of improving
the estimation accuracy by filtering the measurement noise. We assumed that the
measurement model is known and the measurements are affected only by random
noise while no systematic errors are present. We adopted a persistent process-model
that is characterized by unknown and time-varying uncertainties, because the power-
system operating conditions are continuously varying. The power-system state is
usually characterized by smooth variations (quasi-static conditions), but exhibits also
step changes associated, for instance, with the connection/disconnection of loads,
generators and transmission lines. We designed a heuristic method that continuously
updates the process-noise variances used by the DKF; it is effective in quasi-static
conditions and is prone to real-time implementation. Then, we proposed a method,
called prediction-error covariance estimation (PECE), that, in addition to effectively

133
Conclusions

filter the measurement noise in quasi-static conditions, is able to follow the state during
step changes. The PECE method estimates the prediction-error covariances by means
of a constrained convex optimization-problem that has a unique solution. It ensures
the positive semi-definiteness of the estimated prediction-error covariance matrix,
which is a necessary requirement for the DKF numerical stability. Both the heuristic
and PECE methods use a set of previous state estimates and innovations, respectively,
contained in a moving time-window of N time-steps. It is worth noting that N is the
only parameter to be set and can be inferred from off-line simulations. With respect
to the heuristic method, the PECE method is based on a rigorous theory and is able
to cope with step-varying processes, but it has a significantly higher computational
burden. Moreover, we provided a theorem that theoretically proves the better accuracy
of the DKF with respect to the LWLS as long as the process model is correct.

The accuracy and computational performances of the two above-presented methods

were extensively tested via numerical simulations. We considered the New England
39-bus system and the 123-bus test feeder and we reproduced quasi-static conditions
as well as step changes in the system state. First, we make reference to networks
in quasi-static conditions and to the DKF employing the heuristic method: (1) we
statistically validated the adequacy of the persistent process-model; (2) we numerically
validated the theoretical finding about the better accuracy of the DKF with respect to
the LWLS. Second, we simulated a step change in the state variables and we showed
that the DKF employing the PECE method accurately tracks the state, whereas the DKF
employing the heuristic method exhibits a significant delay. The computational time
of the DKF employing the heuristic method is comparable with the LWLS one for the
adopted case studies, whereas the PECE method is computationally more expensive.

The information associated with zero-injection buses (that are very common in power
systems) can be treated in different ways for SE. In the WLS algorithm, zero injections
are usually represented as virtual measurements (i.e., measurements with zero value
and low variance) or are embedded as equality constraints by means of the method of
Lagrange multipliers. However, both methods can be improved for the case of LWLS
with only linear equality constraints. In Chapter 3, we proposed a method based on
LQD that strictly satisfies the equality constraints and reduces the dimension of the
state vector by the number of constraints. The structure of the NE is preserved and the
coefficient matrix is positive definite. Therefore, the problem becomes more computa-
tionally efficient as there are less state variables. Standard techniques exploiting the
characteristics of the NE, such as QRD or the LNR test, can be applied in the same way.
Additionally, we demonstrated that the problem becomes less ill-conditioned. Tests
were conducted on the New England 39-bus system that has 12 zero-injection buses.
The LQD-based method outperforms other commonly-employed methods (a LWLS
including virtual measurements and a LWLS using the method of Lagrange multipliers)
in terms of numerical stability and computational time.

134
 Conclusions

In Chapter 4, we illustrated the field implementation of RTSE in two real power-systems:

the EPFL-campus distribution network and the sub-transmission network of the city
of Lausanne. We deployed PMUs at every bus in order to achieve the network observ-
ability and a high redundancy level by using exclusively synchrophasor measurements.
We described the components of the measurement infrastructure comprising sensors,
PMUs, telecommunication network and PDC. We showed that RTSE can run at 50
estimates-per-second with a total time-latency below 70 ms. Note that this is valid for
networks of similar size and for measurement infrastructures with the same latency
performance.

Then, we presented and extensively discussed the results of different linear SE algo-
rithms (LWLS, LAV and DKF) for the two case studies. Therefore, we had the possibility
to verify the theoretical findings of Chapters 1, 2, 3 with real-scale experiments. We no-
ticed that the measurement noise contained in synchrophasor measurements is much
lower than the systematic errors of the sensors. As the calibration sheets of the sensors
were not available, we could not compensate these systematic errors; consequently,
we are violating the assumption made in Chapter 2 according to which systematic
errors in the measurements are negligible. Indeed, we observed that the DKF and
LWLS estimates are similar. Although the DKF filtering action is visible, the accuracy
improvement is marginal. The numerical results of Chapter 2 are confirmed for the
case of step changes in the state variables: the DKF employing the PECE method reacts
immediately, whereas the DKF employing the proposed heuristic method has a delay
of a few time-steps. For the distribution-feeder case, SE enabled the identification of
an incorrect installations of the sensors. For the sub-transmission network case, many
LWLS normalized residuals systematically exceed the selected threshold for bad-data
identification. Most probably, this is due to line-parameter errors that highly affect the
residuals related to current-flow measurements.

The accuracy level, refresh rate and time-determinism of the estimated state provided
by PMU-based RTSE enable to improve or even completely rethink existing power-
system applications. Fault management may belong to the latter category, as we
demonstrated in Chapter 5. We proposed a fault detection and faulted-line identifica-
tion method based on LWLS-RTSE. It does not need any model of loads/generators and
has no prior knowledge of the fault type and of the neutral connection, even though
it requires that a PMU is installed at every bus. We assume that every PMU measure
the voltage and current-injection phasors at the respective bus. The performance
evaluation was carried out for a real distribution feeder modeled in a RTS. The simu-
lated scenarios included feeders with grounded and isolated neutral, symmetric and
asymmetric faults with low- and high-impedance at different locations. We showed
that the proposed method is able to locate the faulted line with very good accuracy for
low- and high-impedance faults when 0.1-class sensors are used. Whereas, if 0.5-class
sensors are employed, the accuracy is significantly decreased for high-impedance
faults. Additionally, the proposed method maintain the same accuracy irrespectively

135
Conclusions

from the network operating conditions, even in presence of large amount of DG. Finally,
we showed that the total latency of a real system implementing the proposed method
ranges from 102 to 122 ms for this case study. It is well known that protections are a
crucial power-system component and much more detailed tests have to be performed
in order to claim that this method can be used for protective relaying. However, we
can state that both accuracy and latency performance do match the requirements of
protection schemes and are likely to be further improved in the coming years thanks
to the advancement of PMU and telecommunication technologies.

Possible directions for future research:

• A missing point of the study on the DKF is the treatment of bad data, which is
also marginally addressed in the literature. Future research may focus on the
development of bad-data detection and identification algorithms specifically
designed for DKF applied to power-system SE, which can exploit both residuals
and innovations.

• Concerning the analysis of the SE results for the two real networks, we are cur-
rently investigating the problem linked to the high normalized residuals related
to current-flow measurements in the sub-transmission network. Moreover the
availability of such PMU-based measurement infrastructures enable many differ-
ent practical and theoretical future studies, such as optimization of the latency of
each component, PMU-data conditioning, line-parameter estimation, analysis
of other recorded events, etc.

• Regarding fault location using PMU-based RTSE, future work may focus on the
validation of the proposed method for meshed networks, the inclusion of current-
flow measurements and the evaluation of the impact of line-parameter errors on
the fault-location accuracy. Another important aspect may be the development
of a method that relaxes the assumption of having a PMU at every bus; then, it
automatically follows the study on the optimal placement of PMUs.

136
A Appendix

A.1 Derivation of the measurement matrix H

In this appendix we derive the elements of the measurement matrix H defined in
(1.5) for three types of measurements: voltage phasors, current-injection phasors and
current-flow phasors. The three blocks of H related to these measurements are called
HV , HIinj and HIflow , respectively.

The matrix block related to the voltage-phasor measurements HV is simply composed

of zeros and ones. The elements of HV are defined in what follows:

H1V H2V
HV = (A.1)
H3V H4V

where


ip,re 1, if i = h and p = l
HV1 hl,re = (A.2)
0, if i = h or p = l
ip,re
HV2 hl,im = 0 (A.3)
ip,im
HV3 hl,re = 0 (A.4)

ip,im 1, if i = h and p = l
HV4 hl,im = (A.5)
0, if i = h or p = l

In (A.2)–(A.5), the superscripts refer to the bus i, the phase p and the real re or imaginary
im part of the measurements; whereas, the subscripts refer to the bus h, the phase l
ip,re
and the real re or imaginary im part of the state variables. For instance, HV2 hl,im is the
p l
scalar that links the measurement Vi,re with the state variable Vh,im .

137
Appendix A. Appendix

The current-injections are related to the voltages through the so-called network admit-
tance matrix denoted by Y. Therefore, we need to briefly recall the admittance matrix
structure (a detailed description is given in [1, Chapter 8] and [6, Chapter 2.2]). For a
three-phase network of s buses, Y is a 3s × 3s complex matrix:
⎡ ⎤
Y11 Y12 ··· Y1s
⎢ ⎥
⎢Y21 Y22 ··· Y2s ⎥
⎢ ⎥
Y=⎢ . .. .. .. ⎥ (A.6)
⎢ .. . . . ⎥
⎣ ⎦
Ys1 Ys2 ··· Yss

By considering the three phases (a, b, c), the generic element Yih can be written as the
following 3 × 3 complex matrix:
⎡ ⎤
Yihaa Yihab Yihac
⎢ ⎥
Yih = ⎢ ba
⎣Yih Yihbb Yihbc ⎥
⎦ (A.7)
Yihca Yihcb Yihcc

The admittance matrix can be also written in terms of the real and imaginary parts:

Y = G + jB (A.8)

where j denotes the imaginary unit of complex numbers.

The current-injection phasor Iip at bus i ∈ S in phase p ∈ P can be expressed in terms

of the state variables (i.e., the voltage phasors) as


s 
Iip = Yihpl Vhl . (A.9)
h=1 l∈P

The real and imaginary parts of Iip are

s 
 
p
Ii,re = Gpl V l
ih h,re − B pl l
ih h,im ,
V (A.10)
h=1 l∈P

s 
 
p
Ii,im = Gpl V l
ih h,im + B pl l
ih h,re .
V (A.11)
h=1 l∈P

The block of H related to the current-injection phasor measurements can be derived

138
 A.1. Derivation of the measurement matrix H

in a straightforward way from (A.10) and (A.11):

H1Iinj H2Iinj
HIinj = (A.12)
H3Iinj H4Iinj

where

ip,re
HI1inj = Gpl
ih (A.13)
hl,re
ip,re pl
HI2inj = −Bih (A.14)
hl,im
ip,im pl
HI3inj = Bih (A.15)
hl,re
ip,im
HI4inj = Gpl
ih (A.16)
hl,im

In (A.13)–(A.16), the superscripts and subscripts have the same meaning as in (A.2)–
(A.5).

The expression of the current-flow phasor measurements as a function of the state

variables can be derived by considering the two-port π-model of a generic three-phase
network branch u ∈ D between buses i and h, as shown in Fig. A.1. The model
parameters are 3 × 3 complex matrices; therefore, the π-longitudinal impedance zih,L
and the two π-transverse admittances yih,T and yhi,T can be expressed in rectangular
coordinates as

zih,L = rih,L + jxih,L (A.17)

yih,T = gih,T + jbih,T (A.18)
yhi,T = ghi,T + jbhi,T (A.19)

where

• rih,L and xih,L are, respectively, the π-longitudinal resistance and reactance;

• gih,T and bih,T are, respectively, the π-transverse conductance and susceptance
from the side of bus i.

• ghi,T and bhi,T are, respectively, the π-transverse conductance and susceptance
from the side of bus h;

139
Appendix A. Appendix

i
zih,L h

yih,T yhi,T

Figure A.1 – Two-port π-model of a generic three-phase network branch. Note that the
model parameters are 3 × 3 complex matrices.

The matrix of the π-longitudinal impedance zih,L has the following structure:
⎡ aa + jxaa ab + jxab ac + jxac ⎤
rih,L ih,L rih,L ih,L rih,L ih,L
⎢ ba ⎥
zih,L = ⎢ ba
⎣rih,L + jxih,L
bb + jxbb
rih,L ih,L
bc + jxbc ⎥
rih,L ih,L ⎦ (A.20)
ca + jxca
rih,L cb + jxcb
rih,L cc + jxcc
rih,L
ih,L ih,L ih,L

where a, b and c are the three phases. The diagonal entries are the self-impedances
of each phase and the off-diagonal entries are the mutual impedances between the
phases. The matrix of the π-transverse admittance yih,T is written as follows (note that
yhi,T has the same structure):
⎡ aa + jbaa ab + jbab ac + jbac ⎤
gih,T ih,T gih,T ih,T gih,T ih,T
⎢ ba ⎥
yih,T = ⎢ ba
⎣gih,T + jbih,T
bb + jbbb
gih,T ih,T
bc + jbbc ⎥
gih,T ih,T ⎦ (A.21)
ca + jbca
gih,T cb + jbcb
gih,T cc + jbcc
gih,T
ih,T ih,T ih,T

We can also define the π-longitudinal admittance yih,L as the inverse of zih,L :

yih,L = z−1
ih,L . (A.22)

In general, the current-flow is different at the two ends of the branch. Therefore, we
p
indicate as Iih the current-flow phasor at branch u in phase p ∈ P from the side of bus
p p
i, and as Ihi the current-flow phasor from the side of bus h. The expression of Iih as a
function of the state variables (i.e., the voltage phasors) is

p
 
Iih = l
yih,L (Vil − Vhl ) + yih,T
l
Vil . (A.23)
l∈P

p
The real and imaginary parts of Iih are:

p
 pl
Iih,re = [gih,L l
(Vi,re − Vh,re
l
) − bpl l l pl l pl
ih,L (Vi,im − Vh,im ) + gih,T Vi,re − bih,T Vi,im ] (A.24)
l

l∈P

140
 A.2. Proof of Theorem 2

p
 pl
Iih,im = [gih,L l
(Vi,im − Vh,im
l
) + bpl l l pl l pl
ih,L (Vi,re − Vh,re ) + gih,T Vi,im + bih,T Vi,re ] (A.25)
l

l∈P

The block of H related to the current-flow phasor measurements can be derived in a

straightforward way from (A.24) and (A.25):

H1Iflow H2Iflow
HIflow = (A.26)
H3Iflow H4Iflow

where

ihp,re pl pl
HI1flow il,re = gih,L + gih,T (A.27)
ihp,re pl
HI1flow hl,re = −gih,L (A.28)
ihp,re
HI2flow il,im = −(bpl pl
ih,L + bih,T ) (A.29)
ihp,re
HI2flow hl,im = bpl
ih,L (A.30)
ihp,im
HI3flow il,re = bpl pl
ih,L + bih,T (A.31)
ihp,im
HI3flow hl,re = −bpl
ih,L (A.32)
ihp,im pl pl
HI4flow il,im = gih,L + gih,T (A.33)
ihp,im pl
HI4flow hl,im = −gih,L (A.34)

In (A.27)–(A.34), the superscripts refer to the two terminal buses i and h of the branch,
the phase p and the real re or imaginary im part of the measurements; whereas, the
subscripts refer to the bus i or h, the phase l and the real re or imaginary im part of the
ihp,re p
state variables. For instance, HI2flow is the scalar that links the measurement Iih,re
il,im
l .
with the state variable Vi,im

A.2 Proof of Theorem 2

In this Appendix, we give the proofs of Theorem 2 in Section 2.5.1.

The objective function is convex and the set of feasible Σ is convex [129, Chapter
7.1.1, p. 355–357]. Furthermore, the infimum of the objective function cannot occur
when λmin (Σ) → 0 (where λmin denotes the smallest eigenvalue) because the objective
function becomes infinite when λmin (Σ) → 0. As the feasible set is bounded (because

141
Appendix A. Appendix

of the condition In − Σ ≥ 0), it follows that the optimization problem has a finite
minimum, which is attained for one or several values of Σ. Furthermore, log[det(Σ)] is
strictly concave [130, Lemma 6.2.2 p.101] and trace(ΣE) is linear in Σ, therefore the
objective function is strictly convex. It follows that the minimum is reached at one
unique value of Σ.

(QED)

A.3 Proof of Theorem 3

In this Appendix, we give the proofs of Theorem 3 in Section 2.5.1.

First, note that the innovations form a Gaussian random vector with zero mean and
covariance matrix given by (2.22). By [129, Chapter 7.1.1, p. 355–357], it follows that the
ML estimation of the covariance matrix of the innovation is obtained as the optimal
value of T in the following optimization problem (recall that H and R are fixed and
known):
 
min log[det(T)] + trace(T−1 Ck ) (A.35)
T,P

subject to: T = HPHT + R

P real symmetric and P  0.

Second, we show that the feasible sets of the optimization problems (2.23) and (A.35)
are equivalent. More precisely, there is a one-to-one mapping between any feasible Σ
of Problem (2.23) and a feasible (T, P) pair of Problem (A.35). In one direction, given
Σ, T and P are obtained by

 
P = U−1 Σ−1 − In U−T (A.36)
T = HPHT + R (A.37)

We need to prove that if Σ satisfies the conditions of Problem (2.23), then P  0. To

show this, observe that Σ  0 and In − Σ  0 therefore Σ−1 − In  0 [131, Appendix
 
C]. This in turn implies that P = U−1 Σ−1 − In U−T  0.

In the reverse direction, given T and P, Σ is obtained by

 −1
Σ = UPUT + In (A.38)

Similarly, we need to show that if P satisfies the condition of Problem (A.35) then Σ
is well defined, Σ  0 and In − Σ  0. First observe that P  0 therefore UPUT  0

142
 A.3. Proof of Theorem 3

and UPUT + In  In . Thus UPUT + In  0 and is invertible. It follows that Σ is well

defined and Σ  0. Furthermore, Σ−1 = UPUT + In  In and thus Σ−1  In [131,
Appendix C].

Third, we show that the values of the objective functions differ by a constant when Σ
and (T, P) are mapped by the above correspondence. Indeed, by (A.37)
1 1 1 1
R− 2 TR− 2 = R− 2 HPHT R− 2 + Im
 
U  
=V P U 0m−n,n VT + Im
T
0m−n,n
 
UPUT 0n,m−n
=V V T + Im
0m−n,n 0m−n,m−n
 
UPUT + In 0n,m−n
=V VT (A.39)
0m−n,n Im−n

Therefore, noticing that det(V) det(VT ) = 1 and using rules for block-diagonal matri-
ces:
 
det(R)−1 det(T) = det UPUT + In = det(Σ)−1 (A.40)

Taking the inverse of (A.39) gives

1 1
R 2 T−1 R 2 =
 −1
UPUT + In 0n,m−n
=V VT
0m−n,n Im−n
 
Σ 0n,m−n
=V VT (A.41)
0m−n,n Im−n

and thus
 
− 12 Σ 0n,m−n 1
T−1 Ck = R V V T R− 2 Ck (A.42)
0m−n,n Im−n

Using the property that trace(AB) = trace(BA) it comes

trace(T−1 Ck ) =
  
Σ 0n,m−n T − 12 − 12
= trace V R Ck R V
0m−n,n Im−n
  
Σ 0n,m−n Ek E k
= trace
0m−n,n Im−n E k E k
= trace(ΣEk ) + trace(E k ) (A.43)

143
Appendix A. Appendix

where E k , E k , E k are sub-blocks of appropriate sizes and Ek is defined in (2.25).

Putting together (A.40) and (A.43) gives the following relation between the objective
functions:

log det(T) + trace(T−1 Ck ) = − log det(Σ) + trace(ΣEk ) + α (A.44)

where α = log det(R) + trace(E k ) is a constant.

It follows from all the above that Σ is optimal for Problem (2.23) if and only if (T, P) is
optimal for Problem (A.35), when Σ and (T, P) are mapped by (A.36) to (A.38). (QED)

A.4 Covariance estimation method of Myers and Tapley

In [89], Myers and Tapley propose a simple covariance-matching method to assess
both Q and R. We assume that R is known, so that only Q has to be estimated. At
time-step k, the value of Qk is inferred by using the following procedure:

• Compute N vectors (i = 1, ..., N ):

k−i|k−i − Ak−i−1 x
gi = x k−i−1|k−i−1 ; (A.45)

• Then, compute the sample covariance matrix:

 k = cov(g1 , ..., gi , ..., gN ) ;

M (A.46)

• Compute the estimated process-noise covariance matrix:

N +1  
  1 
Qk = M k − Ak−i Pk−i|k−i Ak−i − Pk−i+1|k−i+1 .
T
(A.47)
N
i=2

In stationary conditions, Q approaches the true value as the parameter N increases. In

case of variations of the process stochastic parameters, the smaller N , the faster the KF
reaction.

144
Bibliography
[1] J. Grainger and W. Stevenson, Power system analysis. McGraw-Hill, 1994.

[2] F. Schweppe and J. Wildes, “Power system static-state estimation, part I: Exact
model,” Power Apparatus and Systems, IEEE Transactions on, vol. PAS-89, pp. 120–
125, Jan 1970.

[3] F. Schweppe and D. Rom, “Power system static-state estimation, part II: Approxi-
mate model,” Power Apparatus and Systems, IEEE Transactions on, vol. PAS-89,
pp. 125–130, Jan 1970.

[4] F. Schweppe, “Power system static-state estimation, part III: Implementation,”

Power Apparatus and Systems, IEEE Transactions on, vol. PAS-89, pp. 130–135,
Jan 1970.

[5] A. Monticelli, State estimation in electric power systems: a generalized approach,

vol. 507. Springer Science & Business Media, 1999.

[6] A. Abur and A. Exposito, Power system state estimation: theory and implementa-
tion, vol. 24. CRC, 2004.

[7] A. Monticelli and F. F. Wu, “Network observability: Identification of observable

islands and measurement placement,” IEEE Transactions on Power Apparatus
and Systems, vol. PAS-104, pp. 1035–1041, May 1985.

[8] A. Monticelli and A. Garcia, “Reliable bad data processing for real-time state
estimation,” Power Apparatus and Systems, IEEE Transactions on, vol. PAS-102,
pp. 1126–1139, May 1983.

[9] T. Van Cutsem, M. Ribbens-Pavella, and L. Mili, “Hypothesis testing identifica-

tion: A new method for bad data analysis in power system state estimation,”
Power Apparatus and Systems, IEEE Transactions on, vol. PAS-103, pp. 3239–3252,
Nov 1984.

[10] K. A. Clements and P. W. Davis, “Multiple bad data detectability and identifi-
ability: A geometric approach,” IEEE Transactions on Power Delivery, vol. 1,
pp. 355–360, July 1986.

145
Bibliography

[11] F. Wu and W.-H. Liu, “Detection of topology errors by state estimation (power
systems),” Power Systems, IEEE Transactions on, vol. 4, pp. 176–183, Feb 1989.

[12] W. H. E. Liu, F. F. Wu, and S. M. Lun, “Estimation of parameter errors from

measurement residuals in state estimation (power systems),” IEEE Transactions
on Power Systems, vol. 7, pp. 81–89, Feb 1992.

[13] J. Zhu and A. Abur, “Identification of network parameter errors,” IEEE Transac-
tions on Power Systems, vol. 21, pp. 586–592, May 2006.

[14] L. L. Grigsby et al., The electric power engineering handbook. CRC, 2012.

[15] F. F. Wu, “Power system state estimation: a survey,” International Journal of

Electrical Power & Energy Systems, vol. 12, no. 2, pp. 80–87, 1990.

[16] F. F. Wu, K. Moslehi, and A. Bose, “Power system control centers: Past, present,
and future,” Proceedings of the IEEE, vol. 93, no. 11, pp. 1890–1908, 2005.

[17] A. Garcia, A. Monticelli, and P. Abreu, “Fast decoupled state estimation and bad
data processing,” IEEE Transactions on Power Apparatus and Systems, vol. PAS-
98, pp. 1645–1652, Sept 1979.

[18] A. Phadke and J. Thorp, Synchronized Phasor Measurements and Their Applica-
tions. Power Electronics and Power Systems, Springer US, 2008.

[19] N. Data and N. S. G. Network Management Task Team, “Naspi 2014 survey of
synchrophasor system networks - results and findings,” Tech. Rep. TR 95-041,
NASPI Technical Report, July 2015.

[20] M. Zhou, V. A. Centeno, J. S. Thorp, and A. G. Phadke, “An alternative for includ-
ing phasor measurements in state estimators,” IEEE Trans. Power Syst., vol. 21,
pp. 1930–1937, November 2006.

[21] T. Yang, H. Sun, and A. Bose, “Transition to a two-level linear state estima-
tor—part I: Architecture,” IEEE Transactions on Power Systems, vol. 26, no. 1,
pp. 46–53, 2011.

[22] T. Yang, H. Sun, and A. Bose, “Transition to a two-level linear state estima-
tor—part II: Algorithm,” IEEE Transactions on Power Systems, vol. 26, no. 1,
pp. 54–62, 2011.

[23] N. M. Manousakis, G. N. Korres, J. N. Aliprantis, G. P. Vavourakis, and G.-C. J.

Makrinas, “A two-stage state estimator for power systems with PMU and SCADA
measurements,” in PowerTech, 2013 IEEE Grenoble, pp. 1–6, IEEE, 2013.

[24] M. Glavic and T. Van Cutsem, “Tracking network state from combined SCADA
and synchronized phasor measurements,” in Bulk Power System Dynamics and
Control-IX Optimization, Security and Control of the Emerging Power Grid (IREP),
2013 IREP Symposium, pp. 1–10, IEEE, 2013.

146
 Bibliography

[25] M. Göl and A. Abur, “A hybrid state estimator for systems with limited number of
PMUs,” IEEE Transactions on Power Systems, vol. 30, no. 3, pp. 1511–1517, 2015.

[26] K. D. Jones, J. S. Thorp, and R. M. Gardner, “Three-phase linear state estimation

using phasor measurements,” in IEEE PES Gen. Meet., Vancouver Canada, July
21-25 2013.

[27] L. Zhang, A. Bose, A. Jampala, V. Madani, and J. Giri, “Design, testing, and imple-
mentation of a linear state estimator in a real power system,” IEEE Transactions
on Smart Grid, vol. PP, no. 99, pp. 1–8, 2016.

[28] M. Pignati, M. Popovic, S. Barreto, R. Cherkaoui, G. Dario Flores, J.-Y. Le Boudec,

M. Mohiuddin, M. Paolone, P. Romano, S. Sarri, T. Tesfay, D.-C. Tomozei, and
L. Zanni, “Real-time state estimation of the EPFL-campus medium-voltage grid
by using PMUs,” in Sixth Conf. on Innov. Smart Grid Techn. (ISGT 2015), pp. 1–5,
Washington, DC, USA, Feb 17-20 2015.

[29] A. Phadke, J. Thorp, and K. Karimi, “State estimlatjon with phasor measure-
ments,” IEEE Transactions on Power Systems, vol. 1, no. 1, pp. 233–238, 1986.

[30] A. Abur, “Impact of phasor measurements on state estimation,” in Electrical

and Electronics Engineering, 2009. ELECO 2009. International Conference on,
pp. I–3–I–7, Nov 2009.

[31] J. Zhu and A. Abur, “Effect of phasor measurements on the choice of reference
bus for state estimation,” in IEEE PES Gen. Meet., Tampa, Florida, June 24-28
2007.

[32] T. E. Dielman, “Least absolute value regression: recent contributions,” Journal of

Statistical Computation and Simulation, vol. 75, no. 4, pp. 263–286, 2005.

[33] R. E. Kalman, “A new approach to linear filtering and prediction problems,”

Trans. of the ASME-Journal of Basic Eng., pp. 33–45, 1960.

[34] M. Göl and A. Abur, “LAV based robust state estimation for systems measured by
PMUs,” IEEE Transactions on Smart Grid, vol. 5, pp. 1808–1814, July 2014.

[35] M. Göl and A. Abur, “A robust PMU based three-phase state estimator using
modal decoupling,” IEEE Transactions on Power Systems, vol. 29, pp. 2292–2299,
Sept 2014.

[36] M. Gol, Incorporation of PMUs in power system state estimation. PhD thesis,
Northeastern University, 2014.

[37] J. Zhang, G. Welch, G. Bishop, and Z. Huang, “A two-stage Kalman filter approach
for robust and real-time power system state estimation,” IEEE Transactions on
Sustainable Energy, vol. 5, pp. 629–636, April 2014.

147
Bibliography

[38] M. E. Baran and A. W. Kelley, “A branch-current-based state estimation method

for distribution systems,” IEEE Trans. Power Syst., vol. 10, pp. 483–491, February
1995.

[39] C. A. Fantin, M. R. C. Castillo, B. E. B. de Carvalho, and J. B. A. London, “Using

pseudo and virtual measurements in distribution system state estimation,” in
Transmission Distribution Conference and Exposition - Latin America (PES T
D-LA), 2014 IEEE PES, pp. 1–6, Sept 2014.

[40] J. Liu, J. Tang, F. Ponci, A. Monti, C. Muscas, and P. Pegoraro, “Trade-offs in PMU
deployment for state estimation in active distribution grids,” Smart Grid, IEEE
Transactions on, vol. 3, no. 2, pp. 915–924, 2012.

[41] E. Caro, A. J. Conejo, and R. Mínguez, “Power system state estimation considering
measurement dependencies,” IEEE Trans. Power Syst., vol. 24, no. 4, pp. 1875–
1885, 2009.

[42] C. Muscas, M. Pau, P. A. Pegoraro, and S. Sulis, “Effects of measurements and

pseudomeasurements correlation in distribution system state estimation,” IEEE
Trans. on Instr. and Meas., vol. 63, no. 12, pp. 2813–2823, 2014.

[43] B. I. des Poids et Mesures, C. électrotechnique internationale, and O. interna-

tionale de normalisation, Guide to the Expression of Uncertainty in Measurement.
International Organization for Standardization, 1995.

[44] M. Paolone, A. Borghetti, and C. A. Nucci, “A synchrophasor estimation algorithm

for the monitoring of active distribution networks in steady state and transient
conditions,” in Power Syst. Comp. Conf. (PSCC), Wroclaw, Poland, August 18-22
2014.

[45] Instrument transformers Part 2: Additional Requirements for current transformers.

IEC Std 61869-2, 2012.

[46] Instrument transformers Part 3: Additional Requirements for inductive voltage

transformers. IEC Std 61869-3, 2011.

[47] “IEEE Standard for Synchrophasor Measurements for Power Systems,” IEEE Std
C37.118.1-2011 (Revision of IEEE Std C37.118-2005), pp. 1–61, 2011.

[48] F. Milano, Advanced Techniques for Power System Modelling, Control and Stability
Analysis. London, U.K.: The Institution of Engineering and Technology - IET,
2015.

[49] K. Nishiya, J. Hasegawa, and T. Koike, “Dynamic state estimation including

anomaly detection and identification for power systems,” in IEE Proceedings
C-Generation, Transmission and Distribution, vol. 129, pp. 192–198, IET, 1982.

148
 Bibliography

[50] G. Valverde and V. Terzija, “Unscented Kalman filter for power system dynamic
state estimation,” IET Gener., Transm. and Distr., vol. 5, pp. 29–37, January 2011.

[51] A. Leite da Silva, M. C. Filho, and J. Cantera, “An efficient dynamic state estima-
tion algorithm including bad data processing,” Power Engineering Review, IEEE,
vol. PER-7, pp. 49–49, Nov 1987.

[52] K. A. Clements, G. R. Krumpholz, and P. W. Davis, “Power system state estimation

residual analysis: An algorithm using network topology,” IEEE Transactions on
Power Apparatus and Systems, vol. PAS-100, pp. 1779–1787, April 1981.

[53] M. Göl and A. Abur, “Observability and criticality analyses for power systems
measured by phasor measurements,” IEEE Transactions on Power Systems, vol. 28,
pp. 3319–3326, Aug 2013.

[54] A. Monticelli, “Electric power system state estimation,” Proceedings of the IEEE,
vol. 88, no. 2, pp. 262–282, 2000.

[55] L. Mili, M. G. Cheniae, N. S. Vichare, and P. J. Rousseeuw, “Robust state estimation

based on projection statistics (of power systems),” IEEE Transactions on Power
Systems, vol. 11, pp. 1118–1127, May 1996.

[56] G. Welch and G. Bishop, “An introduction to the Kalman filter,” Tech. Rep. TR
95-041, Dep. of Computer Science, University of North Carolina, July 2006.

[57] A. S. Debs and R. E. Larson, “A dynamic estimator for tracking the state of a
power system,” IEEE Trans. Power Appar. and Syst., vol. PAS-89, pp. 1670–1678,
September 1970.

[58] A. L. Da Silva, M. Do Coutto Filho, and J. De Queiroz, “State forecasting in electric

power systems,” in IEE Proceedings C-Generation, Transmission and Distribution,
vol. 130, pp. 237–244, IET, 1983.

[59] A. K. Sinha and J. K. Mandal, “Dynamic state estimator using ANN based bus
load prediction,” IEEE Transactions on Power Systems, vol. 14, pp. 1219–1225,
Nov 1999.

[60] K.-R. Shih and S.-J. Huang, “Application of a robust algorithm for dynamic state
estimation of a power system,” IEEE Transactions on Power Systems, vol. 17,
pp. 141–147, Feb 2002.

[61] E. A. Blood, M. D. Ilic, J. Ilic, and B. H. Krogh, “A Kalman filter approach to quasi-
static state estimation in electric power systems,” in 2006 38th North American
Power Symposium, pp. 417–422, Sept 2006.

[62] E. Blood, B. Krogh, and M. Ilic, “Electric power system static state estimation
through Kalman filtering and load forecasting,” in Power and Energy Society

149
Bibliography

General Meeting-Conversion and Delivery of Electrical Energy in the 21st Century,

2008 IEEE, pp. 1–6, IEEE, 2008.

[63] A. Jain and N. R. Shivakumar, “Impact of PMU in dynamic state estimation of

power systems,” in 2008 40th North American Power Symposium, pp. 1–8, Sept
2008.

[64] A. Jain and N. R. Shivakumar, “Phasor measurements in dynamic state estimation

of power systems,” in TENCON 2008 - 2008 IEEE Region 10 Conference, pp. 1–6,
Nov 2008.

[65] A. Jain and N. R. Shivakumar, “Power system tracking and dynamic state esti-
mation,” in Power Syst. Conf. and Expos. (PSCE), 2009, Seattke, WA, IEEE, March
15-18 2009.

[66] F. C. Schweppe and R. D. Masiello, “A tracking static state estimator,” IEEE Trans-
actions on Power Apparatus and Systems, vol. PAS-90, pp. 1025–1033, May 1971.

[67] S. J. Julier and J. K. Uhlmann, “New extension of the Kalman filter to nonlinear
systems,” in AeroSense’97, pp. 182–193, International Society for Optics and
Photonics, 1997.

[68] F. Aminifar, M. Shahidehpour, M. Fotuhi-Firuzabad, and S. Kamalinia, “Power

system dynamic state estimation with synchronized phasor measurements,”
IEEE Transactions on Instrumentation and Measurement, vol. 63, pp. 352–363,
Feb 2014.

[69] A. Sharma, S. C. Srivastava, and S. Chakrabarti, “Testing and validation of power

system dynamic state estimators using real time digital simulator (RTDS),” IEEE
Transactions on Power Systems, vol. 31, pp. 2338–2347, May 2016.

[70] I. Arasaratnam and S. Haykin, “Cubature Kalman filters,” IEEE Transactions on

Automatic Control, vol. 54, pp. 1254–1269, June 2009.

[71] L. Zanni, S. Sarri, M. Pignati, R. Cherkaoui, and M. Paolone, “Probabilistic as-

sessment of the process-noise covariance matrix of discrete Kalman filter state
estimation of active distribution networks,” in Intern. Conf. Probab. Meth. Appl.
to Power Syst. (PMAPS), Durham, UK, July 7-10 2014.

[72] S. Sarri, L. Zanni, M. Popovic, J.-Y. Le Boudec, and M. Paolone, “Performance

assessment of linear state estimators using synchrophasor measurements,” IEEE
Transactions on Instrumentation and Measurement, vol. 65, pp. 535–548, March
2016.

[73] R. K. Mehra, “Approaches to adaptive filtering,” IEEE Transactions on Automatic

Control, vol. 17, pp. 693–698, Oct. 1972.

150
 Bibliography

[74] D. Alspach, “A parallel filtering algorithm for linear systems with unknown time
varying noise statistics,” IEEE Transactions on Automatic Control, vol. 19, pp. 552–
556, Oct. 1974.

[75] C. Hilborn Jr and D. G. Lainiotis, “Optimal estimation in the presence of unknown

parameters,” IEEE Transactions on Systems Science and Cybernetics, vol. 5, pp. 38–
43, Jan. 1969.

[76] R. L. Kashyap, “Maximum likelihood identification of stochastic linear systems,”

IEEE Transactions on Automatic Control, vol. 15, pp. 25–34, Feb. 1970.

[77] T. Bohlin, “Four cases of identification of changing systems,” in: R. K. Mehra and
D. G. Lainiotis, eds., System identification advances and case studies, vol. 126,
Academic Press, New York, 1976.

[78] V. A. Bavdekar, A. P. Deshpande, and S. C. Patwardhan, “Identification of process

and measurement noise covariance for state and parameter estimation using
extended Kalman filter,” Journal of Process control, vol. 21, pp. 585–601, Apr.
2011.

[79] R. K. Mehra, “On the identification of variances and adaptive Kalman filtering,”
IEEE Transactions on Automatic Control, vol. 15, pp. 175–184, Apr. 1970.

[80] B. Carew and P. R. Belanger, “Identification of optimum filter steady-state gain

for systems with unknown noise covariances,” IEEE Transactions on Automatic
Control, vol. 18, pp. 582–587, Dec. 1973.

[81] P. R. Bélanger, “Estimation of noise covariance matrices for a linear time-varying

stochastic process,” Automatica, vol. 10, pp. 267–275, May 1974.

[82] R. G. Reynolds, “Robust estimation of covariance matrices,” IEEE Transactions

on Automatic Control, vol. 35, pp. 1047–1051, Sep. 1990.

[83] B. J. Odelson, M. R. Rajamani, and J. B. Rawlings, “A new autocovariance least-

squares method for estimating noise covariances,” Automatica, vol. 42, pp. 303–
308, Feb. 2006.

[84] B. J. Odelson, A. Lutz, and J. B. Rawlings, “The autocovariance least-squares

method for estimating covariances: application to model-based control of chem-
ical reactors,” IEEE Transactions on Control Systems Technology, vol. 14, pp. 532–
540, May 2006.

[85] B. M. Åkesson, J. B. Jørgensen, N. K. Poulsen, and S. B. Jørgensen, “A generalized

autocovariance least-squares method for Kalman filter tuning,” Journal of Process
control, vol. 18, pp. 769–779, Aug. 2008.

151
Bibliography

[86] M. R. Rajamani and J. B. Rawlings, “Estimation of the disturbance structure

from data using semidefinite programming and optimal weighting,” Automatica,
vol. 45, pp. 142–148, Jan. 2009.

[87] Y. Zhen and J. Harlim, “Adaptive error covariances estimation methods for en-
semble Kalman filters,” Journal of Computational Physics, vol. 294, pp. 619–638,
Aug. 2015.

[88] G. Noriega and S. Pasupathy, “Adaptive estimation of noise covariance matrices

in real-time preprocessing of geophysical data,” IEEE Transactions on Geoscience
and Remote Sensing, vol. 35, pp. 1146–1159, Sep. 1997.

[89] K. Myers and B. D. Tapley, “Adaptive sequential estimation with unknown noise
statistics,” IEEE Transactions on Automatic Control, vol. 21, pp. 520–523, Aug.
1976.

[90] V. Krishnan, Nonlinear filtering and smoothing: an introduction to martingales,

stochastic integrals, and estimation. Courier Dover Publications, 2005.

[91] A. Pai, Energy function analysis for power system stability. Norwell, MA: Kluwer,
1989.

[92] J.-Y. L. Boudec, Performance Evaluation of Computer and Communication Sys-

tems. EPFL Press, 2010.

[93] W. Kersting, “Radial distribution test feeders,” in Power Engineering Society

Winter Meeting, 2001. IEEE, vol. 2, pp. 908–912 vol.2, 2001.

[94] L. Zanni, J. Y. L. Boudec, R. Cherkaoui, and M. Paolone, “A prediction-error

covariance estimator for adaptive Kalman filtering in step-varying processes:
Application to power-system state estimation,” IEEE Transactions on Control
Systems Technology, vol. PP, no. 99, pp. 1–15, 2016.

[95] J. Lofberg, “YALMIP: A toolbox for modeling and optimization in MATLAB,” in

Computer Aided Control Systems Design, 2004 IEEE International Symposium on,
pp. 284–289, IEEE, 2004.

[96] K.-C. Toh, M. J. Todd, and R. H. Tütüncü, “SDPT3—a MATLAB software package
for semidefinite programming, version 1.3,” Optimization methods and software,
vol. 11, no. 1-4, pp. 545–581, 1999.

[97] P. Romano and M. Paolone, “Enhanced interpolated-DFT for synchrophasor esti-

mation in FPGAs: Theory, implementation, and validation of a PMU prototype,”
Instrumentation and Measurement, IEEE Transactions on, vol. 63, pp. 2824–2836,
Dec 2014.

152
 Bibliography

[98] L. Zanni, J. Y. L. Boudec, R. Cherkaoui, and M. Paolone, “Linear power-system

state estimation with zero-injection equality constraints via LQ-decomposition,”
Submitted to IEEE Transactions on Power Systems.

[99] C. L. Lawson and R. J. Hanson, Solving least squares problems, vol. 15. SIAM,
1995.

[100] F. Aschmoneit, N. Peterson, and E. Adrian, “State estimation with equality con-
straints,” in Proc. 10th PICA Conf, pp. 427–430, sn, 1977.

[101] F. F. Wu, W. H. E. Liu, and S. M. Lun, “Observability analysis and bad data process-
ing for state estimation with equality constraints,” IEEE Transactions on Power
Systems, vol. 3, pp. 541–548, May 1988.

[102] L. Holten, A. Gjelsvik, S. Aam, F. F. Wu, and W. H. E. Liu, “Comparison of differ-

ent methods for state estimation,” IEEE Transactions on Power Systems, vol. 3,
pp. 1798–1806, Nov 1988.

[103] A. Simoes-Costa and V. H. Quintana, “A robust numerical technique for power

system state estimation,” IEEE Transactions on Power Apparatus and Systems,
vol. PAS-100, pp. 691–698, Feb 1981.

[104] J. W. Gu, K. A. Clements, G. R. Krumpholz, and P. W. Davis, “The solution of ill-

conditioned power system state estimation problems via the method of Peters
and Wilkinson,” IEEE Transactions on Power Apparatus and Systems, vol. PAS-102,
pp. 3473–3480, Oct 1983.

[105] G. N. Korres, “A robust method for equality constrained state estimation,” IEEE
Transactions on Power Systems, vol. 17, pp. 305–314, May 2002.

[106] Y. Guo, W. Wu, B. Zhang, and H. Sun, “An efficient state estimation algorithm
considering zero injection constraints,” IEEE Transactions on Power Systems,
vol. 28, pp. 2651–2659, Aug 2013.

[107] “EPFL Smart Grid.” http://smartgrid.epfl.ch. Accessed: 2017-02-13.

[108] “SCCER - FURIES, Shaping the future swiss electrical infrastructure.” http://
sccer-furies.epfl.ch. Accessed: 2017-02-13.

[109] “IEEE Standard for Synchrophasor Data Transfer for Power Systems,” IEEE Std
C37.118.2-2011 (Revision of IEEE Std C37.118-2005), pp. 1–53, 2011.

[110] A. Derviskadic, P. Romano, M. Pignati, and M. Paolone, “Architecture and experi-

mental validation of a low-latency phasor data concentrator,” IEEE Transactions
on Smart Grid, vol. PP, no. 99, pp. 1–1, 2016.

153
Bibliography

[111] L. Zanni, D. Colangelo, R. Cherkaoui, and M. Paolone, “Impact of synchrophasor

measurement types and uncertainties on the accuracy of distribution system
linear state estimators,” in PowerTech, 2015 IEEE Eindhoven, pp. 1–6, IEEE, 2015.

[112] “National Instrument Grid Automation System.” http://www.ni.com/

white-paper/52556/en/. Accessed: 2016-09-08.

[113] M. Pignati, L. Zanni, P. Romano, R. Cherkaoui, and M. Paolone, “Fault de-

tection and faulted line identification in active distribution networks using
synchrophasors-based real-time state estimation,” IEEE Transactions on Power
Delivery, vol. 32, pp. 381–392, Feb 2017.

[114] P. Navalkar and S. Soman, “Secure remote backup protection of transmission

lines using synchrophasors,” Power Delivery, IEEE Transactions on, vol. 26, pp. 87–
96, Jan 2011.

[115] J. O’Brien, A. Deronja, A. Apostolov, A. Arana, M. Begovic, S. Brahma, G. Brunello,

F. Calero, H. Faulk, Y. Hu, G. Kobet, H. Kirkham, Y. Liao, C. W. Liu, Y. Lu, D. Lukach,
K. Martin, J. Mooney, J. Murphy, K. Narendra, D. Novosel, M. Patel, E. Price, S. Say-
gin, V. Skendzic, R. Taylor, D. Tziouvaras, and S. Ward, “Use of synchrophasor
measurements in protective relaying applications,” in 2014 67th Annual Confer-
ence for Protective Relay Engineers, pp. 23–29, March 2014.

[116] F. Chunju, L. Shengfang, Y. Weiyong, and K. K. Li, “Study on adaptive relay

protection scheme based on phase measurement unit (PMU),” in 2004 Eighth
IEE International Conference on Developments in Power System Protection, vol. 1,
pp. 36–39 Vol.1, April 2004.

[117] S. Das, S. Santoso, A. Gaikwad, and M. Patel, “Impedance-based fault location in

transmission networks: theory and application,” IEEE Access, vol. 2, pp. 537–557,
2014.

[118] F. H. Magnago and A. Abur, “Fault location using wavelets,” IEEE Transactions on
Power Delivery, vol. 13, pp. 1475–1480, Oct 1998.

[119] K.-P. Lien, C.-W. Liu, C.-S. Yu, and J. A. Jiang, “Transmission network fault loca-
tion observability with minimal PMU placement,” IEEE Transactions on Power
Delivery, vol. 21, pp. 1128–1136, July 2006.

[120] Q. Jiang, X. Li, B. Wang, and H. Wang, “PMU-based fault location using voltage
measurements in large transmission networks,” IEEE Transactions on Power
Delivery, vol. 27, pp. 1644–1652, July 2012.

[121] M. Korkalı and A. Abur, “Robust fault location using least-absolute-value estima-
tor,” IEEE Transactions on Power Systems, vol. 28, pp. 4384–4392, Nov 2013.

154
 Bibliography

[122] M. Shiroei, S. Daniar, and M. Akhbari, “A new algorithm for fault location on
transmission lines,” in Power Energy Society General Meeting, 2009. PES ’09. IEEE,
pp. 1–5, July 2009.

[123] A. Öner and M. Göl, “Fault location based on state estimation in PMU observable
systems,” in 2016 IEEE Power Energy Society Innovative Smart Grid Technologies
Conference (ISGT), pp. 1–5, Sept 2016.

[124] “Protection of distribution systems with distributed energy resources,” tech. rep.,
Joint Working Group B5/C6.26/CIRED, March 2015.

[125] J. Mora-Flòrez, J. Melèndez, and G. Carrillo-Caicedo, “Comparison of impedance

based fault location methods for power distribution systems,” Electric Power
Systems Research, vol. 78, no. 4, pp. 657 – 666, 2008.

[126] A. Borghetti, M. Bosetti, C. Nucci, M. Paolone, and A. Abur, “Integrated use

of time-frequency wavelet decompositions for fault location in distribution
networks: Theory and experimental validation,” Power Delivery, IEEE Trans. on,
vol. 25, pp. 3139–3146, Oct 2010.

[127] J. Ren, S. Venkata, and E. Sortomme, “An accurate synchrophasor based fault
location method for emerging distribution systems,” Power Delivery, IEEE Trans-
actions on, vol. 29, pp. 297–298, Feb 2014.

[128] P. Romano, M. Pignati, and M. Paolone, “Integration of an IEEE std. c37.118 com-
pliant PMU into a real-time simulator,” in the 2015 IEEE PowerTech Conference,
2015.

[129] S. Boyd and L. Vandenberghe, Convex Optimization. Cambridge University Press,

2004.

[130] B. Gärtner and J. Matousek, Approximation algorithms and semidefinite pro-

gramming. Springer Science & Business Media, 2012.

[131] A. Van den Bos, Parameter estimation for scientists and engineers. New York City,
NY: John Wiley & Sons, 2007.

155
 LORENZO ZANNI
Chemin du Bochet 33  1025 Saint Sulpice, Switzerland
+41 078 679 67 18  lorenzo.zanni@epfl.ch

RESEARCH INTERESTS

· Smart grids
· Real-time monitoring of power systems via phasor measurement units (PMUs)
· Power-system state estimation
· Fault location and protective relaying

EDUCATION

École Polytechnique Fédérale de Lausanne (EPFL) April 2013 - Present

PhD in Energy, SCI-STI-RC Group and DESL Lausanne, Switzerland
· Thesis title: “Power-System State Estimation based on PMUs: Static and Dynamic Approaches –
from Theory to Real Implementation” (supervisor: Dr. Rachid Cherkaoui, co-supervisor: Prof. Mario
Paolone).

Northeastern University (NEU) Oct. 2015 - Dec. 2015

Visiting researcher, Prof. Ali Abur’s Power Systems Group Boston, USA
· Research related to the PhD thesis.

University of Bologna Sep. 2010 - Dec. 2012

M.Sc. in Electrical Engineering Bologna, Italy
· Thesis title: “Advanced state estimation for unbalanced distribution networks” (supervisor: Prof. Carlo
Alberto Nucci).
· Final mark: 110/110 with honors.

University of Bologna Sep. 2007 - Dec. 2010

B.Sc. in Electrical Engineering Bologna, Italy
· Thesis title: “Calculation of induced over-voltages in overhead lines due to indirect lightnings via the
finite-element method” (supervisor: Prof. Mario Paolone).
· Final mark: 110/110 with honors.

PROFESSIONAL EXPERIENCE
ALSTOM TIS (Transport Information Solution) Jun. 2010 - Jul. 2010
Internship Bologna, Italy
· Design of the electrical substations for the supply of high-speed trains.

SCIENTIFIC PUBLICATIONS
Book Chapters
1. M. Paolone, J.-Y. Le Boudec, S. Sarri and L. Zanni, “Static and recursive PMU-based state
estimation processes for transmission and distribution grids”, Chapter 6 of the book: F. Milano,
“Advances in Power System Modelling, Control and Stability Analysis”, Institute of Engineering
and Technology (IET), 2016.
Journal papers
2. S. Sarri, L. Zanni, M. Popovic, J.-Y. Le Boudec, and M. Paolone, “Performance assessment of
linear state estimators using synchrophasor measurements,” IEEE Trans. on Instrumentation and
Measurement, vol. 65, no. 3, pp. 535–548, Mar. 2016.
3. M. Pignati, L. Zanni, P. Romano, R. Cherkaoui and M. Paolone, “Fault Detection and Faulted-
Line Identification in Active Distribution Networks using Synchrophasors-based Real-Time State
Estimation,” in IEEE Trans. on Power Delivery, vol. 32, no. 1, pp. 381–392, 2017.
4. L. Zanni, J.-Y. Le Boudec, R. Cherkaoui and M. Paolone, “A Prediction-Error Covariance Es-
timator for Adaptive Kalman Filtering in Step-Varying Processes: Application to Power-System
State Estimation,” in IEEE Trans. on Control Systems Technology, vol.-, no.-, pp.-, 2016.
5. L. Zanni, J.-Y. Le Boudec, R. Cherkaoui and M. Paolone, “Linear Power-System State Estimation
with Zero-Injection Equality Constraints via LQ-Decomposition,” Submitted to IEEE Trans. on
Power Systems.
Conference papers
6. M. Paolone, M. Pignati, P. Romano, S. Sarri, L. Zanni, and R. Cherkaoui, “A Hardware-in-the-
Loop Test Platform for the Real-Time State Estimation of Active Distribution Networks using
Phasor Measurement Units,” in Proc. Cigr SC6 Colloquium, Yokohama, Japan, 6–9 Oct. 2013.
7. L. Zanni, S. Sarri, M. Pignati, R. Cherkaoui, and M. Paolone, “Probabilistic assessment of the
process-noise covariance matrix of discrete Kalman filter state estimation of active distribution
networks,” in Proc. 13th International Conference on Probabilistic Methods Applied to Power
Systems (PMAPS), Durham, U.K., 7–10 Jul. 2014, pp. 1–6.
8. M. Pignati, L. Zanni, S. Sarri, R. Cherkaoui, J.-Y. Le Boudec and M. Paolone, “A pre-estimation
filtering process of bad data for linear power systems state estimators using PMUs,” in Proc. 18th
Power Systems Computation Conference (PSCC), Wroclaw, Poland, 18–22 Aug. 2014, pp. 1–8.
9. M. Pignati, M. Popovic, S. Barreto, R. Cherkaoui, G. D. Flores, J.-Y. Le Boudec, M. Mohiuddin,
M. Paolone, P. Romano, S. Sarri, T. Tesfay, D.-C. Tomozei and L. Zanni, “Real-time state esti-
mation of the EPFL-campus medium-voltage grid by using PMUs,” in Proc. 6th IEEE PES ISGT
USA, Washington, DC, USA, 18–20 Feb. 2015, pp. 1–5.
10. L. Zanni, D. Colangelo, R. Cherkaoui, and M. Paolone, “Impact of Synchrophasor Measurement
Types and Uncertainties on the Accuracy of Distribution System Linear State Estimators,” in
Proc. IEEE PowerTech, Eindhoven, Netherlands, 29 Jun.–02 Jul. 2015, pp. 1–6.
11. D. Colangelo, L. Zanni, M. Pignati, P. Romano, M. Paolone, J.-P. Braun, and L.-G. Bernier,
“Architecture and Characterization of a Calibrator for PMUs Operating in Power Distribution
Systems,” in Proc. IEEE PowerTech, Eindhoven, Netherlands, 29 Jun.–02 Jul. 2015, pp. 1–6.
12. S. Sarri, M. Pignati, P. Romano, L. Zanni and M. Paolone, “A hardware-in-the-loop test plat-
form for the performance assessment of a PMU-based real-time state estimator,” in Proc. IEEE
PowerTech, Eindhoven, Netherlands, 29 Jun.–02 Jul. 2015, pp. 1–6.
13. M. Hoefling, F. Heimgaertner, M. Menth, K. V. Katsaros, P. Romano, L. Zanni, and G. Kamel,
“Enabling resilient smart grid communication over the information-centric C-DAX middleware,”
in Proc. International Conference and Workshops on Networked Systems, Cottbus, Germany, 9–12
Mar. 2015, pp. 1–8.
AWARDS
• Best Paper Award at the 13th International Conference on Probabilistic Methods Applied to Power
Systems (PMAPS), 2014.
L. Zanni, S. Sarri, M. Pignati, R. Cherkaoui, and M. Paolone, “Probabilistic assessment of the
process-noise covariance matrix of discrete Kalman filter state estimation of active distribution
networks,” in Proc. 13th International Conference on Probabilistic Methods Applied to Power
Systems (PMAPS), Durham, U.K., 7–10 Jul. 2014, pp. 1–6.
• Best Poster Award at the 2nd Annual Conference SCCER-FURIES, 25 Nov. 2015.
P. Romano, A. Derviskadic, L. Zanni, M. Pignati, J.-Y. Le Boudec, and M. Paolone, “Real-time
state estimation of the EPFL-campus medium voltage grid by using PMUs”.

RESEARCH PROJECTS
NanoTera S3 Grid – SmartGrids April 2013 - April 2017
· Development of new technologies dedicated to the real-time monitoring and management of smart grids
with validation in the EPFL-campus distribution network (smartgrid.epfl.ch).

RT-PMU January 2014 - January 2017

· Real-time monitoring of the sub-transmission network of the city of Lausanne (operated by Services
industriels de Lausanne – SiL) via state estimation based on phasor measurement units.

SUPERVISED STUDENTS
• Derviskadic, A., “Development of a PMU-based RTSE of sub-transmission networks: theory and
experimental validation based on the Lausanne 125 kV grid”, M.Sc. Thesis project, Spring
Semester 2014-2015.

TEACHING EXPERIENCE
• Analyse and Numérique, Undergraduate class, EPFL, 2015.
• Smart-grid Technologies, Graduate class, EPFL, 2015-present.
• Distribution de l’énergie électrique, Undergraduate class, EPFL, 2016-present.
• Conversion d’énergie, Undergraduate class, EPFL, 2016-present.

PEER REVIEWS
• IEEE Trans. on Industrial Informatics, since 2015
• IEEE Trans. on Control Systems Technology, since 2016
• Elsevier Electric Power Systems Research Journal (EPSR), since 2015
• Elsevier Sustainable Energy, Grids and Networks Journal (SEGAN), since 2015
• 19th Power Systems Computation Conference (PSCC), 2016
• IEEE PES General Meeting, 2016
• Innovative Smart Grid Technologies Conference (ISGT), 2016

LANGUAGE SKILLS
Italian Native
English Fluent
French Fluent