See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/330246445

Optimal Dynamic Line Rating Forecasts Selection Based on Ampacity

Probabilistic Forecasting and Network Operators' Risk Aversion

Article in IEEE Transactions on Power Systems · January 2019

DOI: 10.1109/TPWRS.2018.2889973

CITATIONS READS

39 143

3 authors, including:

Georges N. Kariniotakis
MINES ParisTech
278 PUBLICATIONS 8,528 CITATIONS

SEE PROFILE

All content following this page was uploaded by Georges N. Kariniotakis on 20 October 2022.

The user has requested enhancement of the downloaded file.

 Optimal Dynamic Line Rating Forecasts Selection Based
on Ampacity Probabilistic Forecasting and Network
Operators’ Risk Aversion
Romain Dupin, Andrea Michiorri, Georges Kariniotakis

To cite this version:

Romain Dupin, Andrea Michiorri, Georges Kariniotakis. Optimal Dynamic Line Rating Forecasts
Selection Based on Ampacity Probabilistic Forecasting and Network Operators’ Risk Aversion. IEEE
Transactions on Power Systems, Institute of Electrical and Electronics Engineers, In press, pp.1-1.
�10.1109/TPWRS.2018.2889973�. �hal-01998856�

HAL Id: hal-01998856

https://hal.archives-ouvertes.fr/hal-01998856
Submitted on 12 Feb 2019

HAL is a multi-disciplinary open access L’archive ouverte pluridisciplinaire HAL, est

archive for the deposit and dissemination of sci- destinée au dépôt et à la diffusion de documents
entific research documents, whether they are pub- scientifiques de niveau recherche, publiés ou non,
lished or not. The documents may come from émanant des établissements d’enseignement et de
teaching and research institutions in France or recherche français ou étrangers, des laboratoires
abroad, or from public or private research centers. publics ou privés.
 1

Optimal Dynamic Line Rating Forecasts Selection

Based on Ampacity Probabilistic Forecasting and
Network Operators’ Risk Aversion
Romain Dupin, Andrea Michiorri, Member, IEEE, and George Kariniotakis, Senior Member, IEEE

by national or grid regulations, including the definition of a

Abstract-- Real-time current-carrying capacity of overhead maximum temperature for the core of the line. Conductor
conductors is extremely variable due to its dependence on weather temperature is in turn dependent on the heat balance established
conditions, resulting in the use of traditionally conservative static between the heat dissipated by the Joule effect within the
ratings. This paper proposes a methodology for exploiting the conductor, and the heat exchange with the environment on its
latent current-carrying capacity of overhead transmission lines external surface. This heat exchange is strongly influenced by
taking into account line ampacity forecasts, power flow
wind speed, air temperature and solar radiation, parameters that
simulations and the network operator’s risk aversion.
The procedure can be described as follows: Firstly, probabilistic can vary considerably in the space of one day or even one hour.
forecasts for the current rating of transmission lines are generated, The link between the line temperature and weather parameters,
paying particular attention to the reliability of the lower part of which is described in detail in [1], [2], is used to set a current
the distribution. Secondly, a cost benefit analysis is carried out by rating in line with regulations .
solving a bilevel stochastic problem that takes into account the Traditionally, the parameters used to define the line rating
reduction in generation costs resulting from a higher power are set as static. The current rating obtained is called Static Line
transfer capacity and the increased use of reserves caused by Rating (SLR), which is usually defined so that it remains below
forecast errors. The risk appetite of the network operator is the real-line rating up to 99% of the time [3] for the duration of
considered in order to accept or penalize high-risk situations,
a season. Based on this definition, a significant margin of
depending on whether the network operator can be described as
risk neutral or risk averse. improvement exists between SLR and the real value of the line
rating, depending on weather characteristics. In this context, it
Index Terms-- Dynamic Line Rating, Numerical Weather was initially proposed in [4] to set the line rating according to
Predictions, Probabilistic Forecasts, Risk Management dynamic measures of weather parameters, the new rating being
qualified as the DLR.
I. INTRODUCTION The use of DLR instead of SLR for network operations
presents multiple benefits and has already been studied in the
C URRENTLY, power system operators face several
challenges, due to increased peak demand, electricity
market deregulation and increasing penetration of
literature. For example, [5] and [6] show the potential impact of
DLR on the economic dispatch, [7] shows how DLR can
contribute to increasing renewable energy penetration without
renewable energy sources. Due to these latter factors, grid
grid reinforcement, and [8] demonstrates how grid reliability
sections are often operated closer to their voltage or current
can be improved with DLR. These aspects could be taken into
limits. From a network operator’s point of view, network
account during planning operations, in which case DLR
reinforcement is often the default option for alleviating these
forecasts would be required.
problems. However, this is a capital-intensive option and, in
The majority of applications proposed in the literature
this context, the alternative solution of the Dynamic Line
feature the generation of probabilistic DLR forecasts. Most of
Rating (DLR) is gaining popularity.
the time, based on these forecasts, a conservative choice of an
DLR technology centres on the fact that the current-carrying
ampacity value is made for use in planning operations, and
capacity of overhead conductors is limited by thermal
quantile forecasts with a fixed nominal proportion τ are
constraints, in addition to voltage drops and stability limits. ̂τt+h|t made at an instant t
regularly employed [9]. Forecasts DLR
When the operating temperature rises, the conductors elongate,
approaching the ground or other objects. When the clearance for an instant t+h are then set to ensure the probability of being
distance between the line and the objects below is too short this higher than the future observation DLR t+h equal to τ:
can lead to potentially dangerous situations. In order to limit ̂τt+h|t > DLR t+h ) = τ
P(DLR (1)
disastrous consequences, a minimum clearance is determined

Part of this work was carried out thanks to an exchange mobility grant energy systems, CS 10207, 1 rue Claude Daunesse 06904 SophiaAntipolis
awarded to the first author from the project ELECTRA IRP (Grant No 609687), Cedex, France.
funded in part by the European Commission under the 7th Framework (e-mails: romain.dupin, andrea.michiorri, georges.kariniotakis each with
Programme. @mines-paristech.fr).
R. Dupin, A. Michiorri and G. Kariniotakis are with MINES ParisTech, PSL
Research University, PERSEE - Centre for processes, renewable energies and
 2

A state of the art of real applications of this kind in Europe case study used for the model evaluation. In section IV the main
is described in [10]. In [11] and [12], where the experiences of results are presented, and finally section V presents the
the NETFLEX project are described, probability τ is set as conclusions and perspective of this work.
equal to 2%. In [13], probability τ fixed levels of 1%, 5%, 10%
and 20% are tested. In [14], a probability τ of 1% is considered. II. MODELLING APPROACH
Other studies could also be quoted, such as [15] and [16], where The problem to be solved is the optimal operation of a grid on
τ probabilities of 10% and 2.5% are respectively considered. which some lines are equipped with DLR. It is similar to that
Recently, it has been proposed in the literature to no longer formulated in [18] and [22], which assess the DLR value in grid
set the value of DLR forecasts with a fixed level of frequency, management. Since several uncertainties are involved, a
but to use stochastic optimization instead. In [17] and [18], this stochastic optimisation approach is applied. Below, several
is done with two-stage stochastic optimization, which aims at problem formulations are proposed for setting the DLR
minimizing the sum of the generation costs, reserve allocation forecasts, according to different objectives set by the TSO.
and potential reserve activation due to DLR forecast errors. A A. Risk Neutral Strategy – Vertically Integrated Monopoly
similar approach is also proposed in [19], where instead of
We first consider that the electricity producers and electric
reserve activation, it is considered that the retroactions are system operators are the same actor. The problem addressed
achieved through voluntary and involuntary load reduction. here is described in [18] and can be modelled as a two-stage
References [20] and [21] both propose considering that the stochastic optimization problem.
constraint associated with the line ampacity could be replaced For a stochastic optimization problem taking into account
with a cost in the objective function in case of non-respect of DLR forecasts, different terms could be considered in the
the constraint. objective function. One important question concerns the
The above two approaches are radically different. The one introduction and use of penalties regarding situations in which
with the fixed quantiles favours a conservative approach the current is slightly higher than the thermal limits of the lines,
independent from opportunities to improve social welfare, as in [20], [21], [23]. Such cost definitions require knowledge
while ensuring a low risk level. On the contrary, with the second of potential incidents on the grid, their associated probability,
approach based on stochastic optimization, the authors consider and recourse actions when these costs are too high. In this
a risk-neutral approach and aim at maximizing the social paper, we consider a preventive approach regarding N-1
welfare. In this context, the term “social welfare” refers to the incidents, and penalties requiring a corrective approach are not
sum of the consumer and producer surpluses. If the load is introduced. It should be stressed that the simplifications
considered as inelastic towards prices, maximizing the social allowed with this preventive approach permits the use of lookup
welfare is equivalent to minimizing the operation costs. tables which are necessary to provide computations for a high
In this paper, we propose to evaluate the benefits of DLR number of observations in a reasonable time.
based on an intermediate strategy, which aims at improving The objective function is risk neutral, the objective being to
social welfare while ensuring low levels of risk, required for minimize production costs of generators in the system plus the
application on the field. The key contributions can be costs of reserve allocation and potential activation (2):
summarized as follows:
∑ (πfuelg . Pg + Ig . πfixg ) +
1) Several alternative strategies for integrating DLR into g=1..Ng
planning operations are proposed. They aim at improving the up
∑ (πhupg . Hg + πhdog . Hgdo ) +
economic benefits, (referred to here also as the “social min (2)
g=1..Ng
welfare”) for a case where electricity consumption is inelastic up
towards price, while maintaining low values for several ∑ ∑ ρs ∙ (πrupg . R g,s + πrdog . Rdo
g,s )
{s∈Ns g=1..Ng }
parameters to which the Transmission System Operator (TSO)
is averse. The evaluation of these strategies allows us to where Ng is a set of conventional generators; Ig is a binary
understand the significant benefits of using risk-averse variable with value 1 describing a committed generator, and 0
strategies for DLR probabilistic forecasts, compared to if not; πfuelg is the fuel cost for generator g (€/MWh); πfixg is
traditional approaches that use fixed quantiles or risk-neutral
the commitment price for a conventional generator g (€/h); Pg
strategies.
2) We propose a test case and a methodology to evaluate the is the scheduled output of generator g (MW); πhupg and πhdog
economic value of both DLR forecast models and proposed are the costs for maintaining up and down reserve for a
up
strategies for using DLR forecasts. This is based on the use of generator g (€/MWh); Hg and Hgdo are the up and down reserve
lookup table and allow considering grid characteristics. service holding amounts for generator g (MW); Ns is the set of
3) The proposed study takes into account a high number of potential future realizations of DLR and other stochastic
realistic DLR probabilistic forecasts, generated with weather variable observations, with each scenario having a probability
station observations and machine learning methods. This is a ρs of occurrence set with probabilistic forecasts; πrupg and
more complete approach than similar ones proposed in the up
literature, which generally consider only a single virtual DLR πrdog are the reserve activation costs (€/MWh); and R g,s and
probabilistic forecast. Rdo
g,s are the activated reserves from a generator g at scenario s
(MW).
The paper is structured as follows: first, in section II, the The constraints are those of a DC power flow, and are
proposed methodology is presented. Section III presents the described in [18] and [22]. Here, N-1 constraints are added to
 3

account for the risks of line failure, as proposed in [24], which It should be stressed that the risk aversion towards risk
investigates DLR use for transfer capacity setting. activation costs defined here is a trade-off solution: risk
aversion towards other parameters, such as load shedding,
B. Risk-Neutral Strategy – Separate Grid Operators and would be more representative of TSOs’ behaviour. However, it
Electricity Producers would make the problem more complex, and impossible to
In the previous strategy, the generation planning is affected compute for a very high number of scenarios and forecasts, as
by the potential recourse actions set by the TSO. It is however it is the case here. By considering these reserve costs as similar
unlikely that electricity producers would modify their day- to risk indices, it is thus possible to evaluate first different risk
ahead decisions according to the risks and potential costs of aversion functions, and also determine the economic value of
DLR forecast errors, which are assumed by the TSO. DLR forecast models in a latter approach.
In order to decouple the decision of the generator-planned The shape of the function vβ is arbitrarily set by the TSO,
production levels from the decision of the DLR forecast value, depending on its objectives. Here, three functions are proposed
a bilevel stochastic optimization problem is used here, with a inspired from [20], [21], where virtual costs associated with line
leader problem (3)-(4) and a follower problem (5): ampacity violations by currents are defined:
minx∈X,y∈Y F(x, y) • A linear function:
(3)
vβ (x) = (1 + β) ∙ x β≥0 (8)
s. t.
Gi (x, y) ≤ 0 for i ∈ {1,2, … , I} (4)
• A quadratic function:

y ∈ arg minz∈Y {f(x, z): g j (x, z) ≤ 0, j ∈ {1,2, … , J}} vβ (x) = x + β ∙ x² β≥0 (9)
(5)

For a risk-neutral approach, the upper-level objective • An exponential function:

x if β = 0
function F is the same as that described in (2), and the
constraints Gi are those described in the previous section. vβ (x) = { eβ.x − 1 (10)
if β > 0
The lower-level objective function f is the objective eβ − 1
function defined according to the goals of the electricity
producers. Considering no uncertainties except those for which The parameter β aims at describing the TSO’s aversion to the
the TSO is responsible and which in this case are only the DLR risk of observing a DLR forecast error with a high cost. When
forecast errors, this can be written as (6): β is set equal to 0, the TSO exhibits risk-neutral behaviour.
D. Simulation Approach
f(x, z) = { ∑ (πfuelg . Pg + Ig . πfixg )} (6) The solution of the problem described above represents a
g=1..Ng
significant computational challenge due to the high number of
In a future work, reserve activation and allocation terms constraints generated by the stochastic nature of the problem,
could be added in (6), provided that other uncertain variables the respect of the N-1 criterion, and the presence of discrete
such as renewable energy levels of production are added, for values related to the activation of generators. This paper aims
which the producers assume the forecast error risks. The at carrying out a prospective analysis of the benefits of different
constraints g j are the same as those of the leader problem, minus DLR forecast strategies over a long period of time and
the constraints associated with the values of reserve allocations considering a high number of different parameters, here the
and activations. The upper-level decision vector includes the values of β and the nature of the penalty functions.
values of allocated reserves, activated reserves and DLR This requires solving a high number of stochastic problems.
forecasts. The lower-level decision vector includes the planned As an example, considering 100 values of β, 45000 forecasts
production levels and the list of activated generators. Only the and three different penalty functions, 13.5 million simulations
DLR forecast value, set with the leader problem, has an impact would be required. It is therefore necessary to reduce the
on the constraints of the follower problem. computational time taken to resolve the optimisation problems.
To do so, the problem is simplified considering the following
C. Risk Averse Strategy hypotheses and properties:
In the previous section, the objective function (2) was • Since the DLR forecast may remain relatively close to the
defined such that the TSO was considered risk neutral in terms SLR, and that up and down reserves must be allocated in
of reserve activation risks. Here, we propose to modify this case of any forecast error scenarios, it is unlikely that DLR
problem by introducing a risk appetite function vβ (x), to take would have a significant impact on the list of activated
into account TSOs’ risk aversion: generators. The list of activated generators Ig is thus defined
∑ (πfuelg . Pg + Ig . πfixg ) +
considering that all DLR values are equal to the SLR.
g=1..Ng • For the case presented in this paper, only line failures and
up
∑ (πhupg . Hg + πhdog . Hgdo ) +
DLR forecast uncertainties are considered. Uncertainties
min g=1..Ng (7) related to renewable energy sources and electric demand are
not considered. Recourse actions involving activating
up
vβ ( ∑ ∑ ρs ∙ (πrupg . R g,s + πrdog . Rdo
g,s )) reserves can only be carried out for inaccurate DLR
{ s∈Ns g=1..Ng } forecasts, while other potential incidents (which may face a
similar preventive approach) are outside the scope of this
work.
 4

For a test case such as the one described in [18], the property 4) For a scenario s associated with a DLR realization Δl,s , if
of reversibility could be considered regarding the states of Δl,s is lower than inferior to the DLR forecasted value Δl
generator production towards DLR values, where all selected by the TSO, the activated reserves are found by
conventional generators participate in recourse actions, and the computing the differences in the outputs of the different
reserve activation costs are proportional to the day-ahead generators associated with the solutions of the SCOPFs
prices. With the properties described above, the problem can be made with the line ratings equal to Δl,s and Δl .
resolved by solving different Security Constrained Optimal 5) Then, considering the levels of probability ρs of a DLR
Power Flows (SCOPF) according to the following steps: observation Δl,s , it is possible to select an optimal value for
1) A first SCOPF is resolved to initialize the problem, where the DLR forecast Δl .
the line ratings are set at the value that corresponds to the By considering a limited set of possible states of load L1 and
SLR, to define the list of the activated generators Ig . L2 (in this work, 20 for each load area, going from 0% of the
2) After setting Ig , a SCOPF is computed for each possible nominal load to 100%), DLR planned forecasts (100, going
value of line rating Δl . from 100% of the SLR to 200%) and DLR potential
3) For each forecast Δl , the values of allocated reserves observations (100, going from 100% of the SLR to 200%), it is
up
Hg /Hgdo are set by: a) running a SCOPF with a line rating possible to compute the different SCOPFs for each state in
set as Δl , b) calculating the difference between generator which DLR improvements bring benefits. Thus, lookup tables
outputs in this situation and generator outputs after the can be built (Fig. 1), each value of the table being associated to
SCOPF in step 1. This method implies that the problem a tuple of scenario, and the use of this tool allows reducing
associated with the SCOPF is strictly convex, which may considerably the computation time for a period associated to a
not be the case with a linear approach. In the study proposed high number of observations (Fig. 2).
here, a negligible quadratic component is added to the
generator cost function, in order to ensure strict convexity.

Fig. 1. For a load scenario (L1, L2), a DLR forecast Δl and a DLR observation Δl,s , computation process for setting the associated value in the lookup table.

Fig. 2. Flowchart associated to the computations made with the lookup tables.
 5

QRF tends to generate bias errors, the model being less

E. DLR Forecast Generation
effective at explaining complex relationships, and even if such
Here we use the same weather station data as in [15]. For a methods have not been used yet to provide DLR forecasts, a
line located in the UK, several weather stations have been XGBoost model or a Light Gradient Boosting Machine model
installed all along the line for DLR use. In this study we [28] could be preferred. However, for this case study, the
consider 6 of them. For each station, weather characteristics differences between the two models would be little, due to the
were measured every 5 minutes for the period 2009-2010. little amount of features being used. Moreover, a QRF is easy
Using Numerical Weather Predictions (NWP) provided by the to configure, the model outputs converging when the number of
European Centre for Medium-Range Weather decisional trees becomes high, and a single QRF directly
Forecasts (ECMWF), day-ahead probabilistic forecasts are provides all the quantile forecasts. On the contrary, with
generated using a machine learning method. gradient boosting methods, a different model should be trained
The available data set is split into training and testing sets. For for each quantile, and methods as Bayesian optimization [29]
this study, these sets are generated with sliding windows, the and cross validation should be used for each model in order to
models are updated every month, and the training sets used avoid errors due to a non-optimal configuration, thus making
correspond to observations associated with the whole year the training of the model much longer. To give an order of
preceding the moment of the model update. magnitude, instead of a single QRF with the bagging method,
The models are trained to provide forecasts every day at 12:00 999 models would be required considering boosting, in addition
for horizons going from 24 hours to 47 hours. The following to a high number of models being trained for each one to
inputs are used to generate the forecasts: provide optimal parameters. Due to these reasons, the QRF is
• The DLR observation at the instant t the forecast is made. here preferred to gradient boosting methods.
• The mean value of the DLR observations made at the
instants t+h-48 hours, t+h-72 hours and t+h-96 hours.
• NWPs for the zonal and meridional wind speeds at 10m (U
and V), the surface solar radiation downwards (S) and the
ambient temperature at 2m (T) for the four positions that
frame the location of the weather station. The spatial
resolution of the considered NWPs is 0.125°. Since the
ECMWF forecasts are only made for 3-hour time steps and
DLR forecasts are requested with hourly resolution, linear
interpolation is applied to obtain hourly values.
• Four DLR forecasts associated with the ECMWF forecasts
are provided at the instant where the forecasts are made for
the considered horizon. For a given position, the DLR
forecast is provided with the CIGRE physical model [1]
using the forecast values of U, V, S and T.
No selection of the transmission lines’ critical span is made Fig. 3. Illustration of the process of a quantile forecast generation with a QRF
here. Thus, the data from each weather station are studied in the method.
same way as if each station were set at the position of the critical
span as proposed in [15] in order to increase the amount of
observations used.
Regarding the used machine learning method, a Quantile
Regression Forest model (QRF) is selected [25]. This model is
the same as the one used in [15] for DLR forecasting. This
machine learning method is an ensemble method, based on the
generation of k decisional trees, each one trained with a
randomly selected subset of features and of data. For the
forecast generation, the outputs of all the trees are concatenated
and sorted, and the quantile forecasts are drawn from the sorted
list (Fig. 3).
The choice of such a machine learning method could be
discussed. Regarding the latest competition associated to the Fig. 4. Example of day-ahead DLR probabilistic forecasts for station 1, the
use of machine learning in energy fields, the Global Energy different probability intervals (int.) being shown. The 0.1% and 1% quantile
forecasts (Pred.) are the lower limits of the 99.8% and 98% probability
Forecasting Competition [26], the methods using bootstrap intervals.
aggregating processes (bagging) as the QRF appear as being
less effective than the ensemble methods associated to boosting, This model is set up with 2000 trees, each of which is trained
as the XGBoost model [27]. Compared to such methods, the with 5 randomly selected features and with a maximal leaf size
 6

set as equal to 5. In most of the studies associated with DLR for the up-reserve activation.
forecasts, the forecasts are used with fixed quantiles varying The computations of DC optimal power flows are carried out
from 1% to 10%, which thus requires using the forecast model using PyPower [32], and the security constraints are added by
to provide percentile forecasts. However, here, these forecasts generating constraints obtained with Benders cuts. It is
are used as inputs of risk-averse strategies, which aim at considered that no line failures could occur on line 8-7, this part
providing results that are similar to 1%-quantile DLR forecasts of the grid being operated in a radial configuration.
in terms of risks. For this reason, the model provides a set of Lines 8-9 and 8-10 are considered to be equipped with DLR
999 quantile forecasts, for quantiles ranging from 0.1% to functionality, and to have exactly the same DLR value at each
99.9% in 0.1% steps. An example of such forecasts is shown time. These two lines are indeed connected from bus 8 to buses
in Fig. 4. 9 and 10, the latter two being at the same position. The reason
The Mean Absolute Percentage Error (MAPE) and reliability for choosing these two lines is the fact that they appear to be
criteria are considered here to evaluate the performance of the congested when computing the Total Transfer Capacity from
forecast models with regard to state-of-the-art models. First, area 1 to area 2.
regarding the MAPE, a comparison can be made with the
persistence model, i.e. a model where a forecast is set as equal
to the observations made at the moment of the forecast.
For this case, the MAPE for the QRF and the persistence
model are equal to 11.1% and 16.1% respectively. The relative
difference of 31% is in line with the results found in [15].
In terms of reliability, the frequency of overestimation for
quantiles 0.1%, 1% and 2% is equal to 0.8%, 1.8% and 2.7%
respectively. With binomial laws, considering perfectly reliable
forecasts and the number of considered forecasts, it is easy to
find that such frequencies should respectively be inferior to
0.14%, 1.11% and 2.16% with a level of probability of 99%.
The proposed forecast model is then not reliable, even if the last
two frequencies are in line with the results found in [14] and
[30].
Despite this weakness, this model is considered as
acceptable. The first reason is based on recent research made on
DLR forecasts, which indicate that the performances of the
model are close to the models currently used by the TSOs.
Secondly, the reliability criterion is here used to evaluate the
model, but there is no evidence in the literature that this Fig. 5. IEEE 24-bus reliability test system. The lines equipped with DLR are
parameter is the most important one regarding DLR forecasts, marked in red.
and other criteria as the sharpness might be more important. As
an example, with the traditional use of a fixed quantile, it might Lines 8-9 and 8-10 are considered to be equipped with DLR
be considered that a TSO might prefer to use 1% quantile functionality, and to have exactly the same DLR value at each
forecasts from a non-reliable but sharp forecast model than 2% time. These two lines are indeed connected from bus 8 to buses
quantile forecasts from another model, more reliable but less 9 and 10, the latter two being at the same position. The reason
sharp. Such questions are still not answered, and they are to be for choosing these two lines is the fact that they appear to be
addressed by using simultaneously DLR forecasts with congested when computing the Total Transfer Capacity from
different statistical properties and economic models providing area 1 to area 2.
an evaluation of the value of the forecasts. In future work, the The line dynamic ampacity is set as equal to its initial value
development of the tool presented in this paper based on the use provided in the grid data, multiplied by the DLR divided by the
of lookup tables aims at realizing such a research of the required SLR value. For this study, it cannot be below its nominal value.
properties of DLR forecast models. Three SLR values are used, one for each season, and are set
such as to below the observed DLR for 99% of the time for the
III. CASE STUDY year 2009. The seasons are defined as winter (December to
February), spring/autumn (March to May and September to
A. Network Description November) and summer (June to August). Different values of
The IEEE 24-bus grid is considered here (Fig. 5), with the SLR are defined for each weather station.
same characteristics as the ones in [31]. It is split into two areas,
B. Load Series
each of which is considered with different load values.
The reserve allocation and activation costs are set similarly In most studies involving grid simulations using DLR
to [18]: the costs for up and down reserve allocation, πhupg and forecasts, only a given configuration of the grid is used.
πhdog , are set as equal to 6 €/MWh for every generator g; the However, the grid configuration has a significant impact on the
DLR impact and the selection of the optimal forecast.
reserve activation costs are set as equal to the fuel price for the
We therefore propose generating load time series with a
down-reserve activation, and as equal to 150% of the fuel price
bootstrap process, considering only congestion situations, and
 7

taking the frequencies of the different situations from a study of other congestions have appeared. These other congestions
historical time series. ENTSO-E data for the year 2010 are could be new congestions on other lines or changes in
exploited for this goal. Historical data from France (area 1) and congestion natures from thermal limitations to stability or
Germany (area 2) are used, since these data feature a high voltage drop limitations [34].
variety of congestion issues for the bootstrap process.

IV. RESULTS
In order to evaluate the value of the different strategies, the
cost reductions from the initial total system costs, LSLR (€), to
the final total system costs obtained when DLR forecasts are
used, LDLR (€), are considered as benefits. They are normalized
by dividing them by the total LSLR costs associated with the
situations where SLR is used:

LSLR− LDLR
Benefits = 100% ∙ (11)
LSLR
Fig. 6. Evolution of the benefits as a function of the values of the selected fixed
In addition to the benefits, financial losses and risks are also quantiles. The two yellow circles are associated to the use of fixed quantiles
evaluated, using three indices: equal to 1% and 20%.
1) The frequency of events requiring reserve activations due
The observed benefits are computed for the studied period
to overestimations of DLR forecasts (%).
and shown on Table I, with the associated risk and loss indices
2) The total cost of reserve activation Cres act (€) for the
associated being also shown.
studied period. As with the benefits, this cost is normalized
The expected benefits, which are equal to the average of the
using the total costs of the system when the SLR is used:
Cres act expected benefits for the whole set of resolved optimization
Cres act,n = 100% ∙ (12) problems, were also computed, and for every situation
LSLR considered in this paper, we found that the relative difference
between the expected and observed benefits was less than 2%.
3) The frequency of errors with a cost above a defined
This is not surprising, considering the high amount of
threshold f_reserve(threshold) (%). The idea behind this is
observations used and the fact that the stochastic part of the
to address the fact that TSOs may aim at maintaining the
benefits, which is the total costs of the activated reserves, is
DLR error magnitude below a certain threshold most of the
lower than 10% and 3% of the total benefit, respectively with
time. This is especially the case when the reserve may be
the use of risk-neutral optimization and a 1% quantile DLR
hard to call, which could be considered as when it is costly.
forecast. We can thus consider that the expected and observed
The threshold is arbitrarily defined equal as to €1500.
benefits are equal. This allows us to compare forecast models
Situations involving such events are qualified as incidents.
and DLR forecast setting strategies by setting the same levels
A. Traditional Strategies of observed benefits, which are considered as known because
A first study is made by considering traditional methods, i.e. equal to the expected benefits.
the use of fixed quantiles and risk-neutral strategies. Three
TABLE I
strategies are tested: STRATEGY COMPARISON: OBSERVED BENEFITS, FREQUENCY OF ERRORS AND
1) Low arbitrary quantile τ of 1% is selected. FREQUENCY OF INCIDENTS.
2) High arbitrary quantile τ of 20% is selected. Strategy Benefit Overestimation Reserve f_reserve(1500€)
3) The value of the DLR forecast is dynamically set to used (%) frequency (%) costs (%) (%)
minimize the expected operational costs, with a risk- SLR 0.00 0.00 0.000 0.00
neutral strategy.
τ = 1% 0.61 0.72 0.016 0.17
Regarding the choice of the quantile of 20%, it is to stress
that when the selected quantile becomes superior to 20%, the τ = 20% 0.97 4.17 0.105 1.30
benefits converge toward a value of 0.97%. This is shown on
Figure 3, where the evolution of the observed benefits with the Risk-neutral
1.01 3.62 0.085 1.10
strategy
value of a fixed selected quantile is represented.
The results shown on Figure 3 are to be compared to [33], Perfect
1.10 0.00 0.000 0.00
where a similar figure is provided. Contrary to the Figure 3, [33] forecast
shows that benefits do not converge but drop sharply when
quantiles reach values superior to 20%. This difference in Several results are noteworthy. We start by comparing the
behaviour between the results presented here and [33] is due to results with existing literature. Regarding levels of benefits, we
the fact that the grid characteristics are considered in our study, consider study [5], where DLR is considered for dispatching on
and that when DLR forecast values increase above a given an IEEE 30-bus grid and benefits of around 1% of the
threshold no further improvements are observed on the grid, operational costs are found. Here, as shown in Table I, we find
either because the congestions have been erased or because
 8

benefit values ranging from 0.2% to 1.1%. Regarding total

operational costs, we find a mean value of 40k€ per hour.
Results comparable to the literature are also found if we
consider the ratio between the activated reserve costs and the
benefits. For a risk-neutral approach, the ratio between reserve
costs (0.085%) and benefits (1.01%), i.e. a ratio of 1 to 12, is
close to that found in [22].
Compared to the use of 1% quantile forecasts, the use of a
high-quantile (i.e. =20%) or risk-neutral strategy increases the
benefits by 60%, going from 0.61% to 0.97% and 1.01%
respectively. Unfortunately, this is linked with a seven-fold
increase of the charges paid by the TSO and of the number of
large accidents (expressed by the last column in Table I).
In terms of benefits, a dynamic risk-neutral strategy provides
slightly better benefits compared to using high quantiles but a
quite significant reduction (by ~15%) of the number of
incidents (from 1.3% down to 1.1%).
Thus, the dynamic selection of a level of probability for the
forecast brings advantages compared to a fixed quantile.
However, although the required activated amount of reserve is
reduced, it is still far higher than that resulting from the use of
a conservative DLR forecast (i.e. =1%), with the TSO’s
charges and the number of incidents respectively multiplied by
5 (for the risk-neutral strategy) and 7 (for =20%).
B. Evaluation with Risk-Averse Strategies
In this second part, DLR forecast are calculated as before
(optimised fixed quantile τ and variable quantile) with the Fig. 7. Total error DLR forecast costs in function of economic benefits,
depending on the use of a fixed quantile (red) or a dynamic selection, set with
difference that the expected revenue function is filtered with the linear penalties (green), quadratic penalties (orange) and exponential penalties
vβ function described in equation (8-10). (blue). The yellow circle is associated with the use of 1% DLR quantile
The evolution of operational cost linked to the DLR is forecasts. The lower figure is a zoom of the upper one for the benefits in the
area between 0.3% and 0.8%.
followed by modifying the values of the probability τ and the
parameter β. The values of β are selected with the following Fig. 8 shows the evolution of the frequency of incidents with
numerical sequence: the observed benefits.
𝑘−𝛼
βk = 10 20 In contrast to what we observed with the evolution of total
(13)
costs, ranking the different risk-averse strategies is complex
here. Although better results are observed with the use of
For each function vβ , the parameters α and the amount of β
quadratic and exponential penalties compared to linear
values are set such that for the first 5 and last 5 problems solved, penalties, this ranking is not as clear as that observed in Fig. 7,
the benefits observed do not vary. especially for benefit levels close to those associated with the
In Fig. 7, for each level of benefits observed with the use of 1% quantile forecasts. This is mainly due to the quality
variations of β, the value of the total DLR forecast error costs
of DLR forecasts for very low quantiles (≤ 1%). Whereas
is represented.
linear penalties penalize the different potential overestimates
As we can see, for each level of benefits, linear penalties
with linear weights, and thus are only slightly sensitive to the
generate significant improvements in terms of reducing the total
quality of extreme quantile forecasts, this is not the case with
error costs. On the other hand, no improvement is observed with
quadratic and exponential penalties. The forecasts used in this
the use of quadratic penalties whilst the application of
paper have been designed to provide results similar to those
exponential penalties results in a worsening of the performance.
presented in the literature, with no focus on the quality of the
forecasts for probability levels lower than 1%. It is thus not
surprising that the quadratic and exponential strategies perform
poorly for benefit levels associated with the use of low-quantile
DLR forecasts (≈ 0.61%).
 9

penalty functions have been arbitrary defined. Further work

should be carried out in the future to define them so as to reflect
in a more realistic way the the objectives of a specific TSO.

ACKNOWLEDGEMENT
The European Centre for Medium-Range Weather Forecasts
(ECMWF) is acknowledged for the provision of the Numerical
Weather Predictions. The authors would like to thank Prof. M.
Matos of INESC TEC for our useful discussions.

REFERENCES
[1] W. CIGRE, 12, Thermal Behaviour of Overhead Conductors, Tech.
Brochure, 2002.
[2] “IEEE Standard for Calculating the Current-Temperature Relationship
of Bare Overhead Conductors,” IEEE Std 738-2012 Revis. IEEE Std
738-2006 - Inc. IEEE Std 738-2012 Cor 1-2013, pp. 1–72, Dec. 2013.
[3] T. O. Seppa, “Guide for selection of weather parameters for bare
overhead conductor ratings,” CIGRE WG B, vol. 2, 2006.
[4] M. W. Davis, “A new thermal rating approach: The real time thermal
rating system for strategic overhead conductor transmission lines – Part
I: General description and justification of the real time thermal rating
system,” IEEE Trans. Power Appar. Syst., vol. 96, no. 3, pp. 803–809,
May 1977.
[5] M. A. Bucher, M. Vrakopoulou, and G. Andersson, “Probabilistic N-1
security assessment incorporating dynamic line ratings,” in IEEE
Power Energy Society General Meeting, Vancouver, Canada, 2013.
[6] M. Khaki, P. Musilek, J. Heckenbergerova, and D. Koval, “Electric
power system cost/loss optimization using dynamic thermal rating and
linear programming,” in Electric Power and Energy Conference
Fig. 8. Frequency of incidents in function of economic benefits, depending on (EPEC), 2010 IEEE, Halifax, Canada, 2010.
the use of a fixed quantile (red) or a dynamic selection, set with linear penalties [7] C. J. Wallnerström, Y. Huang, and L. Söder, “Impact From Dynamic
(green), quadratic penalties (orange) and exponential penalties (blue). The Line Rating on Wind Power Integration,” IEEE Trans. Smart Grid, vol.
yellow circle is associated with the use of 1% DLR quantile forecasts. The 6, no. 1, pp. 343–350, Jan. 2015.
lower figure is a zoom of the upper one for the benefits in the area between [8] D. M. Greenwood and P. C. Taylor, “Investigating the Impact of Real-
0.3% and 0.8%. Time Thermal Ratings on Power Network Reliability,” IEEE Trans.
Power Syst., vol. 29, no. 5, pp. 2460–2468, Sep. 2014.
[9] A. Michiorri et al., “Forecasting for dynamic line rating,” Renew.
V. CONCLUSIONS Sustain. Energy Rev., vol. 52, pp. 1713–1730, Dec. 2015.
[10] I. Albizu, E. Fernandez, A. J. Mazon, K. J. Sagastabeitia, M. T.
Recently, the use of two-stage stochastic optimization Bedialauneta, and J. G. Olazarri, “Overhead line rating forecasting for
problems for setting DLR forecasts has been investigated in the the integration of wind power in electricity markets,” in 2015
literature. However, most of these studies propose risk-neutral International Conference on Clean Electrical Power (ICCEP),
strategies, in which the DLR forecast setting only aims at Taormina, Italy, 2015, pp. 382–388.
[11] H.-M. Nguyen, J.-L. Lilien, and P. Schell, “Dynamic line rating and
optimizing social welfare, and the TSO’s risk aversion is not ampacity forecasting as the keys to optimise power line assets with the
considered. As a result of this last point, a TSO may still prefer integration of res. The European project Twenties Demonstration inside
to continue setting DLR forecasts using fixed quantiles, thus Central Western Europe,” presented at the CIRED, Stockholm,
disregarding the potential benefits of more developed Sweden, 2013, p. 0946.
[12] H.-M. Nguyen, J.-J. Lambin, F. Vassort, and J.-L. Lilien, “Operational
strategies. experience with Dynamic Line Rating forecast-based solutions to
In this paper, we have proposed new methods for setting increase usable network transfer capacity,” Proc. 45th Sess. Counc.
DLR using DLR probabilistic forecasts and the solving of a Large Electr. Syst. CIGRE, 2014.
bilevel stochastic optimization problem, taking into account a [13] H. E. Hoekstra, C. P. J. Jansen, J. Hagen, J. W. Van Schuylenburg, J. S.
P. Wisse, and W. J. Zitterrsteijn, “Weather forecasted thermal line
risk aversion function. The results are promising: while rating model for the Netherlands,” in CIGRE Session, Paris, France,
ensuring low costs for the TSO or a low frequency of incidents, 2012.
the benefits associated with DLR forecasts could be relatively [14] T. Ringelband, P. Schäfer, and A. Moser, “Probabilistic ampacity
improved to the order of 20%. forecasting for overhead lines using weather forecast ensembles.,”
Electr. Eng., vol. 95, no. 2, 2013.
This work opens up several research topics. First, the [15] J. L. Aznarte and N. Siebert, “Dynamic Line Rating Using Numerical
proposed methodology uses DLR quantile forecasts with Weather Predictions and Machine Learning: A Case Study,” IEEE
probability levels below 1%. Such DLR forecasts have not been Trans. Power Deliv., vol. 32, no. 1, pp. 335–343, Feb. 2017.
investigated in the literature to date, and the interest of such [16] “DLR Forecasting Presented at FERC - Nexans.” [Online]. Available:
http://www.nexans.us/eservice/US-en_US/navigatepub_0_-
forecasts is demonstrated here. Secondly, the methodology 32903/DLR_Forecasting_Presented_at_FERC.html. [Accessed: 08-
requires several forecasts of load and energy production, and a May-2018].
frame would need to be defined for such applications with other [17] M. A. Bucher and G. Andersson, “Robust Corrective Control Measures
stochastic features. Finally, we have shown that the choice of in Power Systems With Dynamic Line Rating,” IEEE Trans. Power
Syst., vol. 31, no. 3, pp. 2034–2043, May 2016.
penalty function has an important impact on the final result. For [18] Y. Chen, F. Teng, R. Moreno, and G. Strbac, “Impact of dynamic line
the illustration of the potential benefits, evaluation indices and rating with forecast error on the scheduling of reserve service,” in
 10

Power and Energy Society General Meeting (PESGM), Boston, USA, works at the MINES ParisTech PERSEE Centre as a senior scientist and head
2016. of the Renewable Energies and Smart Grids Group. He has authored more than
[19] K. Kopsidas, A. Kapetanaki, and V. Levi, “Optimal Demand Response 220 scientific publications in journals and conferences. He has been involved
Scheduling With Real-Time Thermal Ratings of Overhead Lines for as participant or coordinator in more than 40 R&D projects in the fields of
Improved Network Reliability,” IEEE Trans. Smart Grid, vol. 8, no. 6, renewable energies and distributed generation. Among them, he was the
pp. 2813–2825, 2017. coordinator of some major EU projects in the field of wind power forecasting
[20] B. Banerjee, D. Jayaweera, and S. M. Islam, “Alleviating post- such as Anemos, Anemos.plus and SafeWind projects. His scientific interests
contingency congestion risk of wind integrated systems with dynamic include among others timeseries forecasting, decision making under
line ratings,” in Power Engineering Conference (AUPEC), 2014 uncertainty, modeling, management and planning of power systems.
Australasian Universities, Perth, Australia, 2014.
[21] B. Banerjee, S. M. Islam, and D. Jayaweera, “Monte Carlo based
method for managing risk of scheduling decisions with dynamic line
ratings,” in Power & Energy Society General Meeting, Denver, USA,
2015.
[22] F. Teng, R. Dupin, A. Michiorri, G. Kariniotakis, Y. Chen, and G.
Strbac, “Understanding the Benefits of Dynamic Line Rating under
Multiple Sources of Uncertainty,” IEEE Trans. Power Syst., vol. 33, no.
3, pp. 3306–3314, 2018.
[23] B. Banerjee, D. Jayaweera, and S. M. Islam, “Optimal scheduling with
dynamic line ratings and intermittent wind power,” in PES General
Meeting| Conference & Exposition, 2014 IEEE, National Harbor, USA,
2014.
[24] M. Miura, T. Satoh, S. Iwamoto, and I. Kurihara, “Application of
dynamic rating to evaluation of ATC with thermal constraints
considering weather conditions,” in 2006 IEEE Power Engineering
Society General Meeting, Montreal, Canada, 2006.
[25] N. Meinshausen, “Quantile Regression Forests,” J. Mach. Learn. Res.,
vol. 7, no. Jun, pp. 983–999, 2006.
[26] Y. Zhang and J. Wang, “GEFCom2014 probabilistic solar power
forecasting based on k-nearest neighbor and kernel density estimator,”
in IEEE Power Energy Society General Meeting, Denver, USA, 2015.
[27] T. Chen and C. Guestrin, “Xgboost: A scalable tree boosting system,”
in Proceedings of the 22nd acm sigkdd international conference on
knowledge discovery and data mining, 2016, pp. 785–794.
[28] G. Ke et al., “Lightgbm: A highly efficient gradient boosting decision
tree,” in Advances in Neural Information Processing Systems, 2017, pp.
3146–3154.
[29] J. Snoek, H. Larochelle, and R. P. Adams, “Practical bayesian
optimization of machine learning algorithms,” in Advances in neural
information processing systems, 2012, pp. 2951–2959.
[30] X. Sun, P. B. Luh, K. W. Cheung, and W. Guan, “Probabilistic
forecasting of dynamic line rating for over-head transmission lines,” in
IEEE Power Energy Society General Meeting, Denver, USA, 2015.
[31] C. Grigg et al., “The IEEE Reliability Test System-1996. A report
prepared by the Reliability Test System Task Force of the Application
of Probability Methods Subcommittee,” IEEE Trans. Power Syst., vol.
14, no. 3, pp. 1010–1020, Aug. 1999.
[32] R. Lincoln, PyPower. 2017.
[33] R. Dupin, A. Michiorri, and G. Kariniotakis, “Dynamic line rating day-
ahead forecasts - cost benefit based selection of the optimal quantile,”
presented at the CIRED Workshop, Helsinki, Finland, 2016, pp. 122–
126.
[34] L. Dawson and A. Knight, “Applicability of Dynamic Thermal Line
Rating for Long Lines,” IEEE Trans. Power Deliv., vol. 33, no. 2, pp.
719–727, 2018.

Romain Dupin is a graduate of Ecole Centrale Lille, where he took a Chair of

Electrical Networks option, and from the University of Lille 1, where he gained
a Master’s degree in electrical engineering in 2014. Since September 2014, he
has been a PhD student at the MINES ParisTech Center for Processes,
Renewable Energies and Energy Systems (PERSEE) in Sophia Antipolis,
France. He works on developing methods for forecasting dynamic line rating
and studies the impacts of dynamic line rating on power system management.
Andrea Michiorri received an Eng. degree in Mechanical Engineering with a
specialization in Energy from the University of Rome ‘La Sapienza’ in 2005.
He then obtained a PhD from the University of Durham in 2010 with a
dissertation on the thermal state estimation of power system components. He is
currently an associate professor at the MINES-ParisTech Center for Processes,
Renewable Energies and Energy Systems (PERSEE) in Sophia Antipolis,
France, working on the integration of renewable resources and distributed
generators into the power system, with a particular focus on the aspects of
decision-making under uncertainty.
George Kariniotakis (S’95-M’02-SM’11) was born in Athens, Greece. He
received his Eng. and M.Sc. degrees from Greece in 1990 and 1992
respectively, and his PhD from Ecole des Mines de Paris in 1996. He currently

View publication stats