Professional Documents
Culture Documents
Renewable Energy
journal homepage: www.elsevier.com/locate/renene
a r t i c l e i n f o a b s t r a c t
Article history: Hydropower plants are one of the most convenient option for power generation, as they generate energy
Received 2 October 2019 exploiting a renewable source, they have relatively low operating and maintenance costs, and they may
Received in revised form be used to provide ancillary services, exploiting the large reservoirs of available water. The recent ad-
19 December 2020
vances in Information and Communication Technologies (ICT) and in machine learning methodologies
Accepted 16 February 2021
Available online 22 February 2021
are seen as fundamental enablers to upgrade and modernize the current operation of most hydropower
plants, in terms of condition monitoring, early diagnostics and eventually predictive maintenance. While
very few works, or running technologies, have been documented so far for the hydro case, in this paper
Keywords:
Hydropower plants
we propose a novel Key Performance Indicator (KPI) that we have recently developed and tested on
Self-organizing maps operating hydropower plants. In particular, we show that after more than one year of operation it has
Control charts been able to identify several faults, and to support the operation and maintenance tasks of plant oper-
Machine learning ators. Also, we show that the proposed KPI outperforms conventional multivariable process control
Neural networks charts, like the Hotelling t 2 index.
Condition monitoring © 2021 Elsevier Ltd. All rights reserved.
https://doi.org/10.1016/j.renene.2021.02.102
0960-1481/© 2021 Elsevier Ltd. All rights reserved.
A. Betti, E. Crisostomi, G. Paolinelli et al. Renewable Energy 171 (2021) 246e253
by 2030 over half of the world’s hydropower plants will be due for general consensus that this could soon become a very active area of
upgrade and modernization, or will have already been renovated, research [5,16]. In this paper we show that the proposed KPI per-
also driven by investments for new concession renewals. Still ac- forms better than a standard multivariate process control tool like
cording to Ref. [5], the main reason why major works seem around the Hotelling control chart, over a test period of more than one year
the corner is that most industries in this field wish to adopt best (from April 2018 to July 2019).
practices in operations and asset management plans, or in other This paper is organized as follows: Section 2 describes more in
words, share a desire for optimized performance and increased detail the case study of interest, and the data used to tune the
efficiency. In combination with the quick pace of technological proposed indicator. Section 3 illustrates the proposed indicator.
innovation in hydropower operations and maintenance, together Also, the basic theory of the Hotelling multivariate control chart is
with the increased ability to handle and manage big amounts of recalled, as it is used for comparison purposes. The obtained results
data, a technological revolution of most hydropower plants is ex- are provided and discussed in Section 4. Finally, in Section 5 we
pected to take place soon. conclude our paper and outline our current lines of research in this
topic.
1.2. State of the art
2. Case study
In hydropower plants, planned periodic maintenance has been
for a long time the main, if not the only one, adopted maintenance Throughout the paper, we will refer to two hydroelectric power
method. Condition monitoring procedures have been often plants, called plant A and plant B, which have an installed power of
reserved for protection systems, leading to shutting down the 215 MW and 1000 MW respectively. The plants are located in Italy
plants when single monitored signals exceeded pre-defined as shown in Fig. 1, and both use Francis turbines. Plant A is a storage
thresholds (e.g., bearings with temperature and vibration protec- power plant, while plant B is of type pumped-storage. More details
tion). In this context, one of the earliest works towards predictive are provided in the following subsections.
maintenance has been [6], where Artificial Neural Networks (ANNs)
were used to monitor, identify and diagnose the dynamic perfor- 2.1. Hydropower plants details
mance of a prototype of system. Predictive maintenance methods
obviously require the measurement and storage of all the relevant Plant A is located in Northern Italy and consists of four gener-
data regarding the power plant. An example of early digitalization ation units moved by vertical axes Francis turbines, with a power of
is provided in Ref. [7] where a Wide Area Network for condition 60 MVA for each unit. The machinery room is located 500 m inside
monitoring and diagnosis of hydro power stations and substations the mountain. The plant is powered by two connected basins: the
of the Gezhouba Hydro Power Plant (in China) was established. main basin, with a daily regulation purpose with a capacity of 5.9
Thanks to measured data, available in real-time, more advanced 106 m3, and a second basin, limited by a dam, with a seasonal
methods that combine past history and domain knowledge can
provide more efficient monitoring services, advanced fault prog-
nosis, short- and long-term prediction of incipient faults, pre-
scriptive maintenance tools, and residual lifetime and future risk
assessment. Benefits of this include, among other things, prevent-
ing (possibly severe) faults from occurring, avoiding unnecessary
replacements of components, more efficient criteria for scheduled
maintenance.
The equipment required for predictive maintenance in hydro
generators is also described in Ref. [8], where the focus was to gain
the ability to detect and classify faults regarding the electrical and
the mechanical defects of the generator-turbine set, through a
frequency spectrum analysis. More recent works (e.g. Ref. [9])
describe condition monitoring and fault-diagnostics (CMFD) soft-
ware systems that use recent Artificial Intelligence (AI) techniques
(in this case a Support Vector Machine (SVM) classifier) for fault
diagnostics. In Ref. [9] a CMFD has been implemented on a hy-
dropower plant with three Kaplan units. Another expert system has
been developed for an 8-MW bulb turbine downstream irrigation
hydropower plant in India, as described in Ref. [10]. An online
temperature monitoring system was developed in Ref. [11], and an
artificial neural network based predictive maintenance system was
proposed in Ref. [12]. The accuracy of early fault detection system is
an important feature for accurate reliability modeling [13e15].
247
A. Betti, E. Crisostomi, G. Paolinelli et al. Renewable Energy 171 (2021) 246e253
regulation capability. The plant is part of a large hydraulic system warning is triggered when the KPI drops below a threshold and an
and it has been operative since 1951. At full power, the plant em- automatic notification is sent to the operator. The operator, guided
ploys a flow of 88 m3/s, with a net head of 284 m; in nominal by the provided warning, checks and possibly confirms the nature
conditions (1 unit working 24/7, 3 units working 12/7) the main of the detected anomaly. Then, the sequence of data during the fault
basin can be emptied in about 24 h. The 2015 net production was of is eventually removed from the log, so that it will not be included in
594 GWh, serving both the energy and the service market thanks to any historical dataset and it will not be used in the future retraining
the storage capability. stages.
Plant B is representative of pumped-storage power plants; po-
wer is generated by releasing water from the upper reservoir down 3. Methodologies
to the power plant which contains four reversible 250 MW Francis
pump-turbine-generators. After power production, the water is The proposed approach consists in training a self-organizing
sent to the lower reservoir. The upper reservoir, formed by an map (SOM) neural network with the scope to build a model of
embankment dam, is located at an elevation of 643 m in the the nominal behaviour of the system, using an historical dataset of
Province of Isernia. Both the upper and lower reservoirs have an nominal observations, that we denote as training dataset. In fact,
active storage capacity of 60 106 m3. The difference in elevation this is a very peculiar problem where most of the measured data
leads to a hydraulic head of 495 m. The plant has been in operation correspond to nominal behaviours, and very few cases of anoma-
since 1994, and its net production in 2015 was 60 GWh. This plant is lous patterns are usually measured. Asymmetric distributions of
strategical for its pumping storage capability and sells mainly to the correct and faulty patterns advise against the usage of other classic
services market. supervised learning methodologies (e.g., such as Feed-Forward
Neural Networks or Support Vector Machine methods), while it is
2.2. Dataset usually more efficient to adopt unsupervised learning methodolo-
gies to represent the structure and the distribution of nominal data.
The dataset of plant A consists of 630 analog signals with a Hotelling’s control charts are an example of them. Other unsuper-
sampling time of 1 min. The dataset of plant B consists of 60 analog vised learning methods include clustering and vector quantization
signals, since a smaller number of sensors is installed in this system. algorithms, and among these SOMs are very convenient as they
The signals are collected from several plant-components, for operate a map from the original multi-dimensional space to a two-
instance, the water intake, penstocks, turbines, generators, and HV dimensional space that preserves the same topology of the original
transformers. The acquisition system has been in service since the data (i.e., points that were close to each other in the input space
1st of May 2017. In this work, we used data from the 1st of May 2017 correspond to cells that are still close to each other in the two-
to the 31st of March 2018, as the initial training set. Then, the model dimensional output space). Accordingly, SOMs are an excellent
was tested online from the 1st of April 2018 until the end of July candidate when it is necessary to provide an accurate model of a
2019. During the online testing phase, we retrained the model multivariate distribution of data, and the nonlinear map towards
every two months to include the most recent data. the output space allows us to introduce a number of very useful
As usual in this kind of applications, before starting the training tools for data analysis, such as the measurement of distortion that
phase, an accurate pre-processing was required to improve the has been proposed in this work. In fact, SOMs have been widely
quality of measured data. In particular: used for condition monitoring applications in other contexts. In this
manuscript, an original KPI based on a distortion measure has been
1. several measured signals had a large number of not regular data, introduced on purpose for our specific application of interest.
such as missing, or “frozen” samples (i.e., where the signal In particular, the trained SOM is used to calculate a parameter,
measured by the sensor does not change in time), values out of denoted as the average distortion measure of the training dataset,
physical or operative limits, spikes, and statistical outliers in which represents the average weighted distance between each
general; training point and each neuron of the SOM. During the monitoring
2. The training dataset did not contain information about historical stage, new state observations are presented to the SOM and are
anomalies occurred in the plants; similarly, Operation and classified as “in control” or “out-of-control”. For this purpose, we
Maintenance (O&M) logs were not available. calculate the distortion measure of the single new observation and
we compare it against the previously computed average distortion
Accordingly, we first implemented a classic procedure of data measure of the training dataset. The procedure is now illustrated in
cleaning (see for instance Ref. [17]). This procedure has two ad- more detail.
vantages itself: first of all, it allows any data-based condition
monitoring methodology to be tuned upon nominal data corre- 3.1. Self-organizing map neural network based Key Performance
sponding to a correct functioning of the plant; in addition, the Indicator
plants’ operators were informed of those signals whose percentage
of regular samples was below a given threshold, so that they could Self-organizing maps (SOMs) are popular artificial neural
evaluate whether it could be possible to mitigate the noise with network algorithms, belonging to the unsupervised learning cate-
which data were recorded, or whether the sensor was actually gory, which have been frequently used in a wide range of appli-
broken. cations [18,19]. Given a high-dimensional input dataset, the SOM
The second problem was mainly due to the fact that in many algorithm produces a topographic mapping of the data into a
operating plants, the O&M logs are often not integrated with the lower-dimensional output set.
historical databases. Thus, it may be subtle to distinguish nominal The SOM output space consists of a fixed and ordered bi-
and faulty behaviours occurred in the past. In this case, we found it dimensional grid of cells, identified by an index in the range 1; …;
useful to iteratively merge analytic results with feedbacks from on- D, where a distance metric dðc; iÞ between any two cells of index c
site operators to reconstruct correct sequences of nominal data. In and i is defined [18]. Each cell of index i is associated with a model
particular, as we shall see in greater detail in the next section, the vector mi 2R1n that lies in the same high-dimensional space of
output of our proposed procedure is a newly proposed KPI, which the input patterns r2D, where the matrix D2RNn represents the
monitors the functioning of the hydropower plant. In particular, a training dataset to be analyzed, containing N observations of row
248
A. Betti, E. Crisostomi, G. Paolinelli et al. Renewable Energy 171 (2021) 246e253
vectors r2R1n . After the training, the distribution of the model 3.2. The contribution of individual variables to the SOM-based KPI
vectors resembles the distribution of the input data, with the
additional feature of preserving the grid topology: model vectors When the SOM-based KPI deviates from its nominal pattern, it is
that correspond to neighbouring cells shall be neighbours in the desired to identify the individual variables that most contribute to
high-dimensional input space as well. When a new input sample r the KPI variation. This allows the operator not only to identify a
is presented to the network, the SOM finds the best matching unit possible malfunctioning in the hydropower plant, but also the
(BMU) c, whose model vector mc has the minimum Euclidean specific cause, or location, of such a malfunctioning. For this pur-
distance from r: pose, we first calculate an average contribution to DM of individual
variables using the data in the nominal dataset D. We then compare
c ¼ argmini fkr mi kg: the contribution of individual variables of newly acquired patterns
to the average contribution of nominal training patterns.
It is known that the goal of the SOM training algorithm is to
For each pattern r2D, we calculate the following average dis-
minimize the following distortion measure:
tance vector dðrÞ2R1n :
1X X
D
DMD ¼ wci jjr mi jj; (1) 1 XD
N dðrÞ ¼ w ðr mi Þ; cr2D: (6)
r2D i¼1 D i¼1 ci
where the function Then we calculate the vector of squared components of dðrÞ
! normalized to have norm 1, named dn ðrÞ2R1n , as
dðc; iÞ2
wci ¼ exp ; (2)
2s2 dðrÞ+dðrÞ
dn ðrÞ ¼ ; cr2D; (7)
kdðrÞk2
is the neighbourhood function, c is the BMU corresponding to
input sample r, and s is the neighbourhood width. The distortion where the symbol + denotes the Hadamard (element-wise) prod-
measure indicates the capacity of the trained SOM to fit the data uct. Finally, we compute the average vector of the normalized
maintaining the bi-dimensional topology of the output grid. The squared distance components for all patterns in D:
distortion measure relative to a single input pattern r is computed
as: 1X
dnD ¼ dn ðrÞ: (8)
N r2D
X
D
DMðrÞ ¼ wci kr mi k; (3) When a new pattern r is acquired during the monitoring phase,
i¼1
we calculate the following Hadamard ratio:
from which it follows that DMD , as defined in (1), is the average
of the distortion measures of all the patterns in the training data r2 dn ðrÞ
¼ ½cr1 cr2 …crn (9)
D. dnD
In order to assess the condition of newly observed state patterns
r to be monitored, we introduce the following KPI: where the contribution ratios cri , i ¼ 1…n represent how individual
variables of the new pattern influence the DM compared to their
1 influence in non-faulty conditions. If the new pattern actually
KPIðrÞ ¼ : (4)
corresponds to a non-faulty condition, cri takes values close to one.
1 þ 1 DMðrÞ
DMD If the new pattern deviates from the nominal behaviour, some of
the cri exceed the unitary value. An empirical threshold 1.3 was
Roughly speaking, the rationale behind the previous KPI defi- selected, as a trade-off between false positives and true positives.
nition is as follows: if the acquired state r corresponds to a normal Regarding the reliability and the accuracy of the proposed approach
behaviour, its distortion measure DMðrÞ should be similar to the to evaluate the contribution of individual variable, we can
average distortion measure of the nominal training set DMD (which comment as follows:
consists of non-faulty states), and the ratio DMðrÞ
DMD should be close to
one, which in turn gives a KPIðrÞ value to be close to one as well. On The proposed method has demonstrated a good behavior in the
the other hand, if the acquired state r actually corresponds to an real world applications where it has been used.
anomalous behaviour, DMðrÞ should substantially differ from DMD , the definition of the contribution ratios involves the calculation
leading to values of KPIðrÞ considerably smaller than one. In this of the average vector in equation (8), which depends on all the
way, values of the KPI near to one indicate a nominal functioning, observations appearing in the training dataset. Accordingly, it
while smaller values indicate that the plant is going out of control. can be expected that the reliability and accuracy of the proposed
A critical aspect is the choice of the threshold to discriminate a approach can only increase in time, as more data are gathered,
correct and a faulty functioning: for this purpose, we compute the and a richer variety of the possible operating conditions of the
system are observed.
average value mKPI and the variance s2KPI of the filtered KPI values of
all the points in the training set D, and we define the threshold as a
lower control limit (LCL) as follows:
3.3. Thresholds and warning levels
LCLkpi ¼ mkpi 3skpi : (5)
Depending on the expected severity of an anomalous pattern,
If the measured data are noisy, the proposed KPI may present a we have identified two different levels of alarm as follows:
noisy nature as well. For this purpose, in our work we filtered the
KPI using an exponentially weighted average filter over the last 12 h 1. First level warning: the global KPI of the monitored system goes
of consecutive KPI values. below the LCLkpi defined in equation (5);
249
A. Betti, E. Crisostomi, G. Paolinelli et al. Renewable Energy 171 (2021) 246e253
Table 2
List of anomalous behaviours that have been noticed during 16 months of test on the two hydropower plants. Lines 14 and 20 correspond to the two faults that have been
illustrated in this paper.
4.1. Case study 1: plant B - generator temperature signals shown in Fig. 4. After inspection of the operators, they informed us
that similar anomalous situations had occurred in the past as well,
In October 2018, an anomalous behavior was reported by our but had not been tagged as faulty behaviours. As soon as the time
system, which indicated an anomaly regarding a sensor measuring occurrences of the similar past anomalies had been notified by a
the iron temperature of the alternator. It was then observed that plant operator, we proceeded to remove the corresponding signals
the temperature values of such a sensor were higher than usual, as from the training set. Then we recomputed our KPI based on the
shown in Fig. 2, and also higher than the measurements taken by revised corrected historical dataset, and the KPI retrospectively
other similar sensors. However the temperature values did not yet found out that the ongoing faulty pattern had actually started one
exceed the warning threshold of the condition-monitoring systems month earlier. This updated information was then validated by
already operative in the plant. analyzing the output of an already installed gas monitoring system,
From an inspection of the sensor measurements it was possible that continuously monitors a composite value of various fault gases
to establish that the exact day when this anomaly started occurring, in ppM (Parts per Million) and tracks the oil-moisture. The gas
the proposed SOM based KPI sharply notified a warning, as shown monitoring system had been measuring increasing values, with
in Fig. 3; for this case, the time-extension of the training dataset respect to the historical ones, since the 22nd of April 2018, but no
was nine months, from 1 January 2018 to 1 September 2018. After warning had been generated by that system. In this case, this was
receiving the warning alert, the plant operators checked the sensor not however a critical fault, as the level of gasses in the transformer
and confirmed the event as a relevant anomaly. In particular, they oil was not exceeding the maximum feasible limit. However, the
acknowledged that this was a serious problem, since a further warning triggered by our system was used to schedule mainte-
degradation of the measurement could have eventually led to the nance actions that restored the nominal operating conditions. In
stop of the generation unit. For this reason, timely actions were addition, the feedback from the plant operators was very useful in
taken: operators restored the nominal and correct behavior of the order to tune the SOM-based monitoring system, and to increase its
sensor starting on the 12th of October 2018. The saved costs related early detection capabilities.
to the prevented stopping of the generation unit were estimated in In this case, the Hotelling control chart realized of the anoma-
the range between 25 kV and 100 kV. lous behaviour only 20 days after our SOM-based KPI.
While also the Hotelling control chart noticed an anomalous The aforementioned case study allows us to demonstrate the
behaviour of the sensor in the same time frame, still it would have greater predictive power of our proposed method compared to
given rise to several false positives in the past, as shown in Fig. 3. other classic methods. Hotelling’s control charts are considered as
one of the most powerful, and well-established, tools for process
control. Accordingly, the shown superiority of the SOM-based
4.2. Case study 2: plant A - HV transformer anomaly
systems can be considered as a non-obvious and significant
result. In practical terms, detecting a symptom of a possible
The SOM-based KPI detected an anomaly on the HV transformer
anomalous behaviour of the system 20 days in advance (with
of one of the Generation Units at the beginning of June 2018 as
respect than using another method) allows for anticipated inter-
vention times, and in some circumstances may give rise to signif-
Table 3 icant savings if eventual failures are prevented from occurring.
Two failures reported by the predictive system and discussed in detail. Indeed, an essential part of predictive maintenance is the ability to
Case Study Plant Warning name Warning date detect (e.g., through multivariate analysis) anomalous behaviors
that differ from nominal patterns, even if these behaviors do not
1 B Generator Temperature 10/01/2018
2 A HV transformer gasses 04/22/2018 correspond yet to faulty conditions.
251
A. Betti, E. Crisostomi, G. Paolinelli et al. Renewable Energy 171 (2021) 246e253
Fig. 2. Measurement of the temperature sensor in the alternator of plant B. Anomalous values were detected by our system (start warning), timely actions were taken and the
correct functioning was restored (end warning).
Fig. 3. Plant A: KPI as a function of time. SOM-based results (top) are compared with t 2 -based results (bottom).
Fig. 4. Plant B: KPI as a function of time. SOM-based results (top) are compared with t 2 -based results (bottom).
252
A. Betti, E. Crisostomi, G. Paolinelli et al. Renewable Energy 171 (2021) 246e253
4.3. Theoretical comparison of the models Conceptualization, Methodology, Software, Validation, Formal
analysis, Investigation, Writing e original draft, Writing e review &
A comparison of the performance of learning methods in arti- editing.
ficial intelligence is commonly performed through numerical
(simulated) experiments, or, as in our case, by comparing their Declaration of competing interest
behavior in real applications. In our case, we applied both ap-
proaches using the same real data from the same hydropower The authors declare that they have no known competing
plants, and observed that SOMs outperform the Hotelling method: financial interests or personal relationships that could have
in case study 1 Hotelling approach shows false positives that are appeared to influence the work reported in this paper.
not present in our approach, while in case study 2 Hotelling
approach detects the anomaly with 20 days of delay with respect to Acknowledgment
our approach. While in principle one cannot guarantee that SOMs
will outperform Hotelling method for every single fault (as many This work was carried out in the framework of the project
possible faults have not obviously occurred in the hydropower “AUTENS” (Sustainable Energy Autarky) funded by the University of
plants during the period of observation), still SOMs have consis- Pisa (PRA 2020 program).
tently outperformed Hotelling charts for every single analyzed case.
We believe that there is a theoretical reason for the better perfor- References
mance of SOMs: while both SOMs and Hotelling are powerful
methods that generate a model of the data, Hotelling’s approach [1] A. Helseth, M. Fodstad, B. Mo, Optimal medium-term hydropower scheduling
considering energy and reserve capacity markets, IEEE.Trans.Sustain. Energy 7
models the data using the mean vector of a sample of points, and its (3) (2016) 934e942.
covariance matrix, hence using a total number of 2nþ nðn 1Þ= 2 [2] M.N. Hjelmeland, J. Zou, A. Helseth, S. Ahmed, Nonconvex medium-term hy-
parameters, where n is the number of variables. Conversely, the dropower scheduling by stochastic dual dynamic integer programming,
IEEE.Trans.Sustain. Energy 10 (1) (2019) 481e490.
SOM uses the distribution of D centroids and their relative positions [3] P. Heller, E.F. Bollaert, A.J. Schleiss, Comprehensive system analysis of a
in the output space, hence Dðn þ1Þ þ DðD 1Þ=2 parameters. Since multipurpose run-of-river power plant with holistic qualitative assessment,
in general D > n the SOM uses more parameters to model the data, Int. J. River Basin Manag. 8 (2010) 295e304, 1 3-4.
[4] D. Jarry-Bolduc, E. Co ^ te
, Hydro energy generation and instrumentation &
and usually allows one to obtain a more accurate model.
measurement: hydropower plant efficiency testing, IEEE Instrum. Meas. Mag.
17 (2) (2014) 10e14.
5. Conclusion [5] International Hydropower Association, Hydropower Status Report, Sector
Trend and Insights, 2018, 2018.
[6] W. Jiang, Research on predictive maintenance for hydropower plant based on
Driven by rapidly evolving enabling technologies, most notably MAS and NN, in: Third IEEE International Conference on Pervasive Computing
Internet-of-Thing sensors and communication tools, together with and Applications, 2008, pp. 604e609.
more powerful artificial intelligent algorithms, condition moni- [7] Z.-H. Li, X.-B. Yang, S.-Q. Niu, O.P. Malik, Maintenance-oriented information
digitalization of hydro turbine generator sets, in: IEEE Power & Energy Society
toring, early diagnostics and predictive maintenance methodolo- General Meeting, 2009, pp. 1e7.
gies and tools are becoming some of the most interesting areas of [8] L.C. Ribeiro, E.L. Bonaldi, L.E.L. de Oliveira, L.E. Borges da Silva, C.P. Salomon,
research in the power community. While some preliminary ex- W.C. Santana, J.G. Borges da Silva, G. Lambert-Torres, Equipment for predictive
maintenance in hydrogenerators, in: AASRI Procedia vol. 7, 2014, pp. 75e80.
amples can be found in many fields, solar and wind plants being [9] L. Selak, P. Butala, A. Sluga, Condition monitoring and fault diagnostics for
one of them, fewer applications can be found in the field of hy- hydropower plants, Comput. Ind. 65 (2014) 924e936.
dropower plants. In this context, this paper is one of the first ex- [10] I. Buaphan, S. Premrudeepreechacharn, Development of expert system for
fault diagnosis of an 8-MW bulb turbine downstream irrigation hydro power
amples, at least up to the knowledge of the authors, that results of a plant, in: 6th IEEE International Youth Conference on Energy, IYCE), 2017.
newly proposed KPI are validated over a 1-year running test field in [11] S.D. Mili c, A.D. Zigi
c, M.M. Ponjavi c, Online temperature monitoring, fault
two hydropower plants. detection, and a novel heat run test of a water-cooled rotor of a hydro-
generator, IEEE Trans. Energy Convers. 28 (3) (2013) 698e706.
This paper provided very promising preliminary results, that
[12] C. Fu, L. Ye, Y. Liu, R. Yu, B. Iung, Y. Cheng, Y. Zeng, Predictive maintenance in
encourage further research on this topic. In particular the proposed intelligent-control-maintenance-management system for hydroelectric
procedure can not be implemented in a fully unsupervised fashion generating unit, IEEE Trans. Energy Convers. 19 (1) (2004) 179e186.
yet, but some iterations with plant operators still take place when [13] E. Khalilzadeh, M. Fotuhi-Firuzabad, F. Aminifar, A. Ghaedi, Reliability
modeling of run-of-the-river power plants in power system Adequacy studies,
alarm signals are alerted. Also, the proposed condition monitoring IEEE.Trans.Sustain. Energy 5 (4) (2014) 1278e1286.
strategy is a first step towards fully automatic predictive mainte- [14] Huanhuan Li, et al., Performance evaluation in enabling safety for a hydro-
nance schemes, where faults are not just observed but are actually power generation system, Renew. Energy 10 (2019) 1628e1643.
[15] N. Rouge, B. Valluy, A. Bircher, O. Bernard, B. Ge hant, Performance-based
predicted ahead of time, possibly when they are only at an incipient maintenance to protect the profitibility of hydropower assets during difficult
stage. In the opinion of the authors this is a very promising area of market conditions, in: Proc. Hydro 2016 - Achievements, Opportunities and
research and there is a general interest towards the development of Challenges, 2016. Montreux, Switzerland.
[16] Y. Chunzhen, et al., Real-time condition monitoring and fault detection of
such predictive strategies. components based on machine-learning reconstruction model, Renew. En-
ergy 133 (2019) 433e441.
CRediT authorship contribution statement [17] G. Cowan, Statistical Data Analysis, Oxford University Press, USA, 1998.
[18] T. Kohonen, The self-organizing map, Proc. IEEE 78 (9) (1990) 1464e1480.
[19] M. Tucci, M. Raugi, Adaptive FIR neural model for centroid learning in self-
Alessandro Betti: Supervision, Conceptualization. Emanuele organizing maps, IEEE Trans. Neural Network. 21 (6) (2010) 948e960.
Crisostomi: Writing e original draft, Writing e review & editing, [20] H. Hotelling, Multivariate Quality Control, Techniques of Statistical Analysis,
McGraw-Hill, New York, 1947.
Visualization, Validation, Investigation. Gianluca Paolinelli:
[21] H. Fouzi, et al., Reliable fault detection and diagnosis of photovoltaic systems
Conceptualization, Methodology, Software, Validation. Antonio based on statistical monitoring approaches, Renew. Energy 116 (2018) 22e37.
̆
Piazzi: Software, Validation, Formal analysis, Investigation, Re- [22] R. De Maesschalck, D. Jouan-Rimbaud, D.L. Massart, The Mahalanobis distance,
sources, Data curation, Writing e original draft. Fabrizio Ruffini: Chemiometrics. Intell. Lab. Syst. 50 (1) (2000) 1e18.
[23] F. Aparisi, M.A. de Luna, The design and performance of the multivariate
Software, Validation, Formal analysis, Investigation, Resources, synthetic-T2 control chart, Commun. Stat. Theor. Methods 38 (2) (2009)
Data curation, Writing e original draft. Mauro Tucci: Supervision, 173e192.
253