Fault detection and prediction in Smart Grids

Christian Andre Andresen

Energy Systems Bendik Nybakk Torsæter Hallvar Haugdal
SINTEF Energy Research Energy Systems Department of Electric
Trondheim, Norway SINTEF Energy Research Power Engineering
[email protected] Trondheim, Norway NTNU
[email protected] Trondheim, Norway
Kjetil Uhlen [email protected]
Department of Electric Power
Engineering
NTNU
Trondheim, Norway
[email protected]

Abstract—Modern society is to a larger and larger extent overall electric system towards faults and disturbances. Without
dependant on electric energy, and hence the reliance on and the stabilizing nature of rotating masses cascading faults on a
utilization of the electric grid is increasing steadily. At the same system wide level may become more prevalent in the future. In
time the production and consumption patterns are changing from the transition towards a more flexible and adaptable Smart Grid,
large centralized generation of electric power and pure consumers it is essential that such challenges are handled. Enabling the grid
to distributed generation (DG) and more complex consumers. This to meet the resulting change in load profile and the increased
transition causes higher stress on an aging infrastructure and variability in demand is a massive challenge on a European level.
major investments are required over the coming years to maintain
a reliable supply of electric energy. Better monitoring solutions However, there is a parallel development that is helping to
and predictive methods can increase the possible utilization of the provide the means to meet this challenge. The development in
existing grid and reduce the fault frequency. This paper presents instrumentation, communication and data analysis has been
some current challenges in the grid and a possible monitoring significant, and novel solutions for monitoring and managing the
solution and fault prediction method. This is exemplified with grid is becoming reality [2, 3]. Of special interest to this paper is
statistics and field-measurements from the Norwegian power grid. the development in monitoring instrumentation, such as Power
Quality Analysers (PQA) and Phasor Measurement Units
Keywords—PMU, PQA, fault prediction, WAMS, statistical (PMU). In combination with the development of machine
learning
learning (ML) techniques, it will be possible to extract and
I. INTRODUCTION efficiently analyse the vast amounts of data generated from these
devices. This article argues that modern machine learning
The use of electric power has been increasing over the last algorithms may utilize the gathered data to predict or give early
decades and is expected to further increase globally in both the warnings about faults and instabilities in the grid both on a
short and long-term future [1]. At the same time the grid system-wide level, such as frequency oscillations between
infrastructure in Europe is aging and changing towards a Smart generators, and on the local level, such as impeding component
Grid architecture [2], where there is more flexibility in the grid and line failure. To do so, the faults and instabilities needs a
to meet varying demand. On the production side a substantial preceding signature to manifest itself in the measurements
amount of intermittent energy sources such as PV and wind has before the operational ability of the grid is affected. This paper
been introduced into the energy mix, to some degree replacing seeks to demonstrate how a sufficiently large dataset may be
controllable generation such as coal and nuclear-powered plants. collected for such signatures to be detected and used to give
On the consumption side the consumers have become more effective decision support for the grid operators.
complex in their behaviour. Many modern electric appliances,
such as induction heaters and chargers for electric vehicles (EV), It is beyond the scope of this paper to give a comprehensive
have a more challenging consumption pattern than traditional overview of challenges in the smart grid field. This is an active
appliances, which causes higher stress on the distribution grid. area of research and this paper seeks to give some insight into a
In addition, some end-users, previously considered pure possible route with special focus on the Norwegian grid as an
consumers, have installed local generation capability and are at example case.
times net producers of electric energy termed prosumers.
II. FAULT DETECTION AND MITIGATION
As a consequence of these trends traditional generation
The last few years the amount of measuring instruments and
methods that included the use of large rotating mass, such as
sensors in the power grid has increased significantly, through
large turbine generators, are increasingly being replaced by
PQAs, PMUs, smart meters (AMI – Advanced Metering
power electronics that digitally control the frequency and
Infrastructure) and other types of measurements (temperature,
voltage levels. This development gives less robustness in the
wind, humidity, etc.) and sensors (line angle, vibration, etc.). The

XXX-X-XXXX-XXXX-X/XX/$XX.00 ©20XX IEEE

increased amount of measurement data from the power grid, in related to extreme weather conditions are common. Such faults
combination with new methods within machine learning, gives can be earth faults and short-circuits caused by wind (vegetation)
new possibilities in terms of fault detection and mitigation. By a and icing. Even though such faults normally don't have a
system-wide deployment of PQA and PMU devices the grid may signature in advance of the decisive fault event, it could be
be monitored in real-time via an efficient communication possible to anticipate such faults and quantify the probability of
network. This enables the continuous evaluation of the current their occurrence by analysing power quality measurements in
state of the grid, indicating congestions, frequency oscillations combination with weather data.
and overall load distribution. It also enables the real-time
application of predictive algorithms for fault probability While it is assumed that some faults and disturbances can be
assessments that have the potential of reducing the fault prewarned, other fault events are unfeasible or prohibitively hard
frequency if appropriate mitigating actions can be found. to include in a system for fault prediction and mitigation. Such
faults are related to lightning strikes, human errors, animals etc.
A. Fault prediction and system requirements
Transformer faults have been pointed out by operators as a
The power grid is exposed to many possible types of motivation for deploying an early warning system. Severe
disturbances and faults, and all of them have different signatures transformer faults could cause long-lasting outages due to
and propagations. In order for a system to be able to detect and lengthy and complicated repairs, delivery and transport, and
predict a future fault, the fault needs to have a signature in occurs frequently enough to represent a significant cost [5].
advance of the decisive fault event. It is assumed that many of Reference [6] considers a transformer subject to mechanical
the most frequent fault events in the power grid develop over degradation and quantifies the resulting introduction of
time and have a fault development signature ahead of the actual harmonics in the current and voltage waveforms. This is a
malfunction such as investigated in [4]. As an example, an developing signature that could potentially be detected by a ML-
insulator failure leading to an earth fault or short circuit might based early warning system.
have discharges in advance that could be detected using
measuring instruments. Earth faults in the medium-voltage grid can cause dangerous
situations and disturbances if they are not handled efficiently. As
There is some uncertainty related to whether a prewarning an example, intermittent earth faults caused by e.g. vegetation
ahead of a fault makes it possible to mitigate and avoid a crucial and weather (snow, ice, wind etc.) can be difficult to detect and
fault event. The uncertainty is related to how much time that is avoid. In Figure 1, an intermittent earth fault in the medium-
needed to initiate preventive measures, and what kind of action voltage grid is shown. The figure shows that the earth fault is
that is needed in order to avoid the potential fault event and unstable for a period of more than 12 hours before it is cleared.
customer outage. According to Norwegian grid companies, a Such an intermittent earth fault could be detected and handled at
prewarning of at least a few minutes is needed for a grid operator an earlier stage using measurement data, which in turn could
to manually do a preventive network configuration, e.g. make fault detection and location more reliable and reduce the
disconnecting a faulty line. If the fault can't be mitigated from danger for grid customers.
the control room, a prewarning of minimum a few hours is
needed for a technician to solve the problem. In the above
paradigm it is thought that a predictive algorithm would give
decision support to a human operator, and that the operator
would have to make a decision and initiate an action manually.
The application of automatic corrective actions initiated by the
system itself is emerging. In such a paradigm, where a machine
makes alterations to the grid to mitigate faults and customers
outages, a warning horizon of seconds or milliseconds may
suffice.
The fault prewarning that is received in the operation centre
must be reliable in order for the warning system to be trusted by
Figure 1: RMS voltage in grid with intermittent earth fault
the grid operator. In addition, the system should contain detailed
information on the type of fault, whether the fault is imminent III. MEASUREMENT INFRASTRUCTURE
and how critical the fault potentially could be for grid customers.
Real-time monitoring schemes requires high-resolution
B. Predictable faults and disturbances measurements that are reported with a low time delay (latency)
Traditionally, monitoring based on measuring instruments in to a centralized computing unit. Two relatively extensively
the grid has been sparse. Due to this, existing research on the deployed measurement infrastructures in the Norwegian grid are
power quality signature of fault events and disturbances is based on PQAs and PMUs, which both answer relatively well to
limited. However, some types of faults and disturbances are the demands of real-time monitoring. The characteristics of
more likely to have a signature in advance of the decisive fault these are outlined in the following, along with examples on what
event than others. Problems related to component failure, caused information can be retrieved.
by humidity, salt, ageing etc., generally have a development over A. Power Quality Analysers
time. It is therefore assumed that it should be possible to detect
such faults prior to their occurrence using power quality Power Quality Analysers, also known as PQAs, are
measurements. In the Norwegian grid, faults and disturbances commonly used to analyse the power quality in the grid.
Traditionally, PQAs have been installed in important junctions
in the grid, such as the bus bar on the secondary side of a HV/MV
transformer. PQAs can give valuable information on all voltage
quality parameters, e.g. voltage variation, transients and
harmonic distortion. Compared to PMUs, PQAs are superior
when it comes to sampling rate. Some PQA devices have a
bandwidth of up to 25 kHz and higher and are able to efficiently
reproduce and store continuous waveforms for large periods of
time (multiple years without generating extreme data volumes).
This is achieved by compression/decompression of signals when
storing/reading. The extent to which this has a negative impact
on time delay must be kept in mind when designing a real-time
application, since fast-acting systems might be crucial to achieve
rapid alert and remedial action for certain types of faults. Figure 3: WAMS using PQAs in three different locations in the same
geographic region
In Figure 2, the RMS and waveform of a line voltage with
high harmonic distortion is depicted in the same pane. When During the measurement period displayed in the figure, a
analysing the RMS voltage, it appears that the voltage is stable voltage dip occurs in the grid. As can be seen from the figure,
and without significant noise. However, when the waveform of the voltage dip has different amplitude in the three locations. It
the voltage is analysed in detail, it is apparent that the voltage is apparent that the voltage dip influences the voltage level in the
waveform is highly distorted. Thus, if analysing only the RMS middle pane more than in the two others. Using this information,
voltage in this case, valuable information would be lost. This it is reasonable to assume that the PQA in the middle pane has a
clearly shows the advantages of a PQA compared to a PMU, as higher electrical proximity to the root cause of the voltage dip
the higher bandwidth makes it possible to analyse voltage quality than the two other PQAs.
on a much higher resolution than 50 Hz RMS.
Considering the degradation of a transformer mentioned in B. Phasor Measurement Units
Section II.B, monitoring the RMS would not yield sufficient
information to be able to detect a change in harmonic distortion. PMUs (as standardized in [8] and [9]) provide time-
For this type of signature, a higher resolution on the level of that synchronized voltage and current phasors at a reporting rate in
provided by a PQA would be required. the range of 10-60 Hz, usually 50 Hz in the Norwegian grid.
Each reported frame contains phasors constituted by an angle
RMS
and a magnitude, which corresponds to the phase and amplitude
of the measured quantity, usually three phase voltages and
currents. Due to the highly accurate time-synchronization, which
is often achieved by using GPS-signals, providing an accuracy
in the range of milliseconds, the angles of phasors situated at
widely separated geographical locations can be compared with
high accuracy. PMU measurements therefore gives an advantage
in the wide area perspective, making this infrastructure suitable
for monitoring the operational state and stability of the power
system at the transmission level.
PMUs report measurements to Phasor Data Concentrators,
which transmit the data further to other units, for instance a
computer in a control centre. The transfer protocol is very
efficient, resulting in a low latency in the range of milliseconds
Figure 2: RMS and waveform of line voltage with high harmonic distortion
[9]. PMUs are therefore suitable for real-time applications, and
monitoring applications based on PMU measurements are
already used extensively in power systems around the world
An increasing instrumentation in the power grid also [10].
provides a better fundament for wide area monitoring systems
(WAMS) [7]. Both DSOs and TSOs invest in a higher number Figure 4 and Figure 5 indicate how information and
of high-quality measuring instruments, which can be used to knowledge of a disturbance can be retrieved from PMU
anticipate and detect faults and disturbances, and to find their measurements recorded at three widely separated locations. The
location in the grid. In Figure 3, an example of wide area figures show frequency and power flow, respectively, in the
monitoring using PQAs is presented. The figure displays the different locations during a loss of load.
measured line voltage in three different locations, using three The increasing frequency in all locations indicates surplus
different PQAs. The PQAs are placed in three different MV production of power, i.e. increased generation or loss of load.
grids, but they are all in the same geographic region. Looking at the power flows in Figure 5, which flow from
Location 2 to Location 1, it can be deduced that a loss of load
probably occurred near Location 1. Further, looking closer at the
frequencies in the three locations, it is observed that the IV. MEASUREMENT RESOLUTION ASPECTS
frequency in Location 1 rises quickly before the other The requirements for fault logging with respect to PQA and
frequencies. This also indicates that the loss of load occurs close PMU data varies dramatically. It is not always immediately clear
to Location 1. what is the required data quality needed for event identification.
The trained expert would easily be able to perform the above However, modern power quality analysers enable logging at
reasoning by studying measurements retrospectively. However, least as rapidly as 1024 samples per cycle using standard
if such knowledge could instead be retrieved by an online ML- commercially available products. Logging strategies also differ
based algorithm, the operators could potentially be notified between event-based logging and continuous logging. Event-
almost immediately about the root cause of the recorded based logging enables high-resolution sampling of parameters
deviations. such as voltage and current in the immediate time interval
surrounding some event, e.g. a voltage transient. Such logging
50.25 often does not allow the determination of the development of the
50.2
fault, and root causes can be harder to identify lacking such
information. In addition, continuous time series allows for a
Frequency [Hz]

50.15 much richer analysis of the power system enabling a broader

50.1
range of model-building and statistical learning.

50.05
In the following, an earth fault is identified in a 22 kV line in
the Norwegian power grid. Figure 6 illustrates what information
50 would be available for analysis of such a fault with different
0 5 10 15 20 25
logging strategies. In the top pane cycle-by-cycle RMS values of
Time [s]
the phase voltages are shown. The pane shows that the
Location 1 Location 2 Location 3
magnitude of the first event (at about 4800 seconds) is taken to
be of approximately the same magnitude as the main earth fault
Figure 4: Frequency excursion measured by PMUs at three locations (at about 7000 seconds). In the middle pane, 1-second averages
approximately 500 km apart. A loss of load occurs at Location 1, causing the are shown. In this case, the preceding event is displayed to be of
frequency to rise rapidly in this area, before Locations 2 and 3 starts significant lower magnitude than the later earth fault. In the
following. bottom pane, 1-minute averaged RMS values are shown. In this
case, it is impossible to see that there was an earth fault in the
700 grid at all. It should be noted that if max/min values for each
600 sample period for each resolution were plotted, the information
would be more complete even for the larger time-period
Power Flow [MW]

500

400
averages. It is assumed that higher time resolution is more
300
beneficial for the application of algorithms for fault prediction.
200

100

0
0 5 10 15 20 25
Time [s]

Power Flow 1 Power Flow 2

Figure 5: Power flows (derived from PMU measurements) from Location 2 in

the direction of Location 1, where a loss of load occurred.

A benefit of wide area information is that it provides for

improved situational awareness, helping operators to quickly
understand the criticality of disturbances and take correct
actions. Looking further, it is envisioned that WAMS will be
used in automatic control systems to perform very fast and
targeted remedial actions. Actions including fast ramping, or
shedding of generation or loads, will then be performed without
the operators' intervention, thus greatly reducing the risk of
blackouts. Solutions for system protection or control schemes
that utilize synchronized phasor measurements are commonly
referred to as wide area protection (WAPS) or wide area control
systems (WACS) or combined WAMPAC [11].
Figure 6: An earth fault in the Norwegian 22 kV grid displayed with three
levels of resolution. In the top pane cycle-by-cycle RMS values are shown, in
the middle pane 1-second RMS averages are shown and in the bottom pane 1-
minute RMS values are shown.

V. GRID FAULT STATISTICS

Faults and disturbances in the power grid are recorded by
several actors, such as end-users, DSOs and TSOs. In the
Norwegian power grid there is a mandatory responsibility given
to the DSOs and the TSO to report faults and disturbances via a
national reporting structure called FASIT [12]. The report must
specify the type of incident, the time and duration, the number Figure 8: Number of events reported in the Norwegian 33-420 kV grid in 2016
of affected phases, the voltage level and plausible root cause and and the resulting energy not delivered (ILE) as function of root cause.
contributing causes. This reporting gives a statistical basis for
national reports on the frequency of faults and disturbances in
VI. DISCUSSION
the Norwegian grid both regarding voltage levels and causes.
Every year the Norwegian TSO Statnett publishes statistics As has been demonstrated in the sections above, there is a
regarding faults in the 1-22 kV grid and in the 33-420 kV grid significant development in the available grid monitoring data
[5], and the Norwegian governmental directorate NVE publishes alongside an expectation that the loading profile of the grid will
statistics of end-user power quality and disruptions [13]. In become more challenging for the operators in the future. This
Figure 7 and Figure 8 respectively, the number of events development goes alongside an unprecedented evolution in the
reported and the resulting energy not delivered is shown for the fields of big-data analysis and machine learning, such as deep
1-22 kV grid and for the 33-420 kV grid. Data is publicly neural networks [14] [15]. These developments are made
available and collected from Statnett's web pages [9]. As can be possible by better computational resources (such as HPC and
seen, the number of events and consequences is far larger in the GPU utilization), better machine learning algorithms and more
1-22 kV grid. Vegetation and wind is two of the major root accessible libraries making the application of the methods
causes [5]. Both these categories lead to faults and disturbances considerably easier and more efficient for researchers in fields
where a significant portion is expected to be of a developing outside computer science and mathematics.
nature where there are indications of such events before the There are several initiatives seeking to exploit modern
critical fault occurs. By analysing historical PQA and PMU data machine learning methods to both detect and predict faults in the
in combination with the recorded faults that have both time, grid, such as GRIP [16], EarlyWarn [17] and others. One of the
location and cause information, it may be possible to apply central challenges in the machine learning domain is the large
modern machine learning algorithms to enable prediction of the amount of training data that is required to train the algorithms.
occurrence of such faults as discussed below. In traditional grid fault analysis, experts study each fault
occurrence separately and seek the root cause and/or mitigating
actions to be implemented. There is a limit to the number of
cases that can be studied by a person or team in this manner. In
addition, manual analysis is not suitable for real-time
applications. The training of machine learning algorithms
typically applies thousands or millions of examples before
robust functionality is achieved. Each example needs to be
labelled (annotated) with the desired outcome of the algorithm
for that particular example. As an example, algorithms designed
to do picture recognition have utilized millions of pictures that
have been manually labelled with their class ("cat", "dog",
"horse" etc.) for the algorithms to be trained to the level where
Figure 7: Number of events reported in the Norwegian 1-22 kV grid in 2016 they can automatically distinguish between different classes of
and the resulting energy not delivered (ILE) as function of root cause. images.
For the utilization of machine learning towards fault
classification and prediction, a similar dataset of labelled input
will be needed. The above fault reporting structure combined
with the continuous logging of the PQA and PMU units
described enables the automatic generation of such training
material. Since high-resolution time series are becoming
available for periods over years for a large number of locations,
training datasets can be generated with the required time
resolution as needed by the algorithm. Manually extracting fault
statistics for thousands or millions of fault occurrences would
require enormous resources and raise data-right issues. There are
however a number of methods to automatically detect and
classify faults in PQA and PMU data series [3]. By running such VIII. REFERENCES
classification algorithms over continuously collected time-series
data for an extensive grid area (such as the Norwegian grid) over [1] IEA, “World Energy Outlook 2017,” IEA, 2017.
a prolonged period of time, a sufficient number of events can be [2] N. IqtiyaniIlham, M. Hasanuzzaman and M. Hosenuzzamana,
collected to facilitate the application of machine learning “European smart grid prospects, policies, and challenges,” Renewable
methods such as deep neural networks. It is important to note and Sustainable Energy Reviews, vol. 67, pp. 776-790, 2016.
that time series of "non-events" (time series where there are no [3] S. M. Muyeen and S. Rahman, Communication, Control and Security
faults) also needs to be collected for the algorithms to learn to Challenges for the Smart Grid, Institution of Engineering and
accept normal variations in the field data. The above-reported Technology, 2017.
statistics indicate the number of events that can be expected to [4] B. D. Russell, R. M. Cheney, T. L. Anthony, C. L. Brenner, C. F.
be found in time series from the Norwegian grid. The data Wallis and W. E. Muston, “Reliability improvement of distribution
suggests that there are a few thousand events per year, and feeders through real-time, intelligent monitoring,” in Power & Energy
Society General Meeting, 2009. PES '09. IEEE, Calgary, 2009.
continuous monitoring has been conducted for more than a
decade at some locations. Keeping in mind that a balanced [5] Statnett, “Statnett feilstatistikk,” Statnett, [Online]. Available:
http://www.statnett.no/Kraftsystemet/Nedlastingssenter/Feilstatistikk/.
dataset of time series with and without faults should be collected, [Accessed 14 May 2018].
a training set of tens of thousands of samples should be feasible.
[6] A. S. Masoum, N. Hashemnia, A. Abu-Siada, M. A. S. Masoum and S.
It remains to be established what time resolution and duration M. Islam, “Online transformer internal fault detection based on
that is needed for robust fault detection and prediction. instantaneous voltage and current meassurements considering impact
of harmonics,” IEEE Trans. Power Deliv., vol. 32, no. 2, pp. 587-598,
A distinction should be drawn between the task of fault 2017.
identification and fault prediction. In the former case of [7] A. Chakrabortty and P. P. Khargonekar, “Introduction to Wide-Area
identification, the task is to quickly determine that a fault has Control of Power Systems,” in 2013 AMERICAN CONTROL
occurred, where it has occurred and what type of fault that has CONFERENCE (ACC), Washington, DC, 2013.
occurred. This information then needs to be communicated to an [8] IEEE, IEEE standard for synchrophasor meassurements for power
operator or system that takes appropriate action. Time is of systems, IEEE Standard C37.118.1-2011, 2011.
significant importance in the execution of this task. Traditionally [9] IEEE, IEEE standard for synchrophasor data transfer for power
some instrumentation has had batching of data before systems, IEEE Standard C37.118.2-2011, 2011.
communicating it to a central where any fault detecting activities [10] U. ANNAKAGE, A. RAJAPAKSE, B. BHARGAVA, N. R.
would be carried out. This batching could give a latency of 5 CHAUDHURI, S. RIBEIRO CAMPOS ANDRADE, V.
minutes or more in commonly found industrial applications. PATHIRANA, K. V. KATSAROS, A. CHAKRABORTTY and V.
Recent improvements in communication could potentially bring TERZIJA, “Application of Phasor Measurement Units for Monitoring
Power System Dynamic Performance,” in Cigré, Paris, 2017.
this latency down to milliseconds. Employing rapid algorithms
for fault identification could enable the location and [11] V. Terzija, G. Valverde, D. Y. Cai, P. Regulski, V. Madani, J. Fitch, S.
Skok, M. M. Begovic and A. Phadke, “Wide-Area Monitoring,
classification in near real-time. In the case of prediction, the time Protection, and Control of Future Electric Power Networks,”
scale of importance varies according to the fault to be predicted, PROCEEDINGS OF THE IEEE, vol. 99, no. 1, pp. 80-93, 2011.
as discussed above. However, the same argument regarding low [12] FASIT, “FASIT - Reporting structure for faults and disturbances in the
latency data communication and rapid algorithms holds for this Norwegian power grid,” [Online]. Available: www.fasit.no. [Accessed
case. Machine learning algorithms have demonstrated very rapid 2018 2018 2018].
processing times once they have been trained properly, and it is [13] NVE, “NVE - Avbruddsstatistikk,” NVE, [Online]. Available:
likely that they will contribute to bringing rapid identification https://www.nve.no/reguleringsmyndigheten-for-energi-rme-marked-
and prediction to the power grid as well. og-
monopol/nettjenester/leveringskvalitet/leveringspalitelighet/avbruddsst
VII. CONCLUSION atistikk/. [Accessed 14 May 2018].
[14] Y. LeCunn, Y. Bengio and G. Hinton, “Deep Learning,” Nature, vol.
This paper presents the foreseen challenges arising the in the 521, pp. 436-44, 2015.
operation of the grid, and some instrumentation and monitoring [15] F. Chollet, Deep learning with Python, New York: Manning
solutions that may help in the operation of the grid. The Publications Co., 2018.
increasing availability of long time series of high quality [16] SLAC National Accelerator Laborotory, [Online]. Available:
monitoring data with low latency combined with advances in https://www6.slac.stanford.edu/news/2017-09-14-slac-led-project-
big-data analysis and machine learning gives rise to efficient will-use-artificial-intelligence-prevent-or-minimize-electric-grid.
fault identification and prediction. Developments in this research [Accessed 14 May 2018].
area is expected to enable a more robust and efficient utilization [17] S. E. AS, “EarlyWarn,” [Online]. Available:
of the grid in the transition towards a smarter grid. https://www.sintef.no/en/projects/earlywarn/. [Accessed 14 May
2018].
ACKNOWLEDGMENT
The authors would like to thank the Research Council of
Norway and industry partners for the support in writing this
paper under project 268193/E20 EarlyWarn.