1. Introduction
In this period of ecological and environmental awareness, renewable and natural energy sources have experienced an important generalization and development in the last 20 years. Territories with a vast potential for solar energy boost their photovoltaic production and expand the number of solar plants. The stakes are even higher for non-interconnected island territories, such as the overseas territories of France (Corsica, etc.).
Favorably located in an intertropical zone, La Réunion increased its PV energy production from 0 to 206.3 MW in 15 years [
1]. The La Réunion grid has the particularity to be non-interconnected with electrical production that is mainly imported fossil fuel-based. The growing proportion of intermittent energy such as PV sources increases the risk of a strong variation in energy production. In order to ensure the stability of this insular power grid, the need for maintenance and energy production forecast increases with the size of the PV park. In addition to the normal aging of the oldest plants, humid tropical climate conditions induce an increase in faults and premature aging of the PV systems [
2]. Moreover, La Reunion’s PV park includes plants in a wide range of sizes, with a large proportion of small installations, for example, 62.66% of the number of the PV plants are below 9 kVA, and are unevenly equipped for FDD.
PV systems can be prone to degradation faults and electrical faults with different occurrences and severity [
3,
4,
5]. The meteorological conditions highly impact the PV aging velocity, as a hot-humid climate and thermal cycling have a negative effect on PV modules [
6,
7]. The conventional detection and protection methods available in the industry are already used in PV plants and allow to avoid electrical faults such as line–ground (L-G) faults, line–line (L-L) faults, arc series faults and open-circuit (OC) faults [
8,
9]. They are electrical fuses, circuit breakers, surge protectors [
10]. Arc fault detectors are not mandatory on EU installations, and their installation is expensive.
Degradation faults are hotspots, cell cracks, discoloration, delamination, connections sectioning, bypass diode dysfunction, mismatch, corrosion, shading or potential induced degradation (PID) [
2,
11,
12]. PV degradation faults detection requires the setting-up of methods such as electroluminescence, infrared thermography, UV fluorescence and I–V tracers. These methods have an important cost [
10], particularly in the case of small installations, and have the disadvantage of being on demand at the moment that a drop in production has already been noticeable. Furthermore, they involve the deployment of human and technical resources and are complex to implement in the case of rooftop plants. Some of them require module disconnection and therefore a production loss. Nevertheless, technics are newly developed to allow a rapid module analysis as an ultra-rapid and low-cost I–V curve [
13], containing an additional acquisition system with a high sampling frequency.
To solve these issues, it is important to enable the deployment of low-cost, on-line, easy-to-implement FDD solutions that target all production plant dimensions. The DETECT (Diagnosis onlinE of sTate of health of EleCTric systems) research project takes root in this context, with the objective to develop an on-line and low-cost FDD with no additional sensors or costly equipment. PV FDD is crucial and has taken a lot of effort in the last 10 years, and a considerable amount of research articles have been published. Reviews list and demonstrate the extent of the great diversity of the developed methods [
11,
14,
15,
16,
17,
18,
19].
In addition to the usual components of protection [
9] and visual solutions [
20], advanced fault detection methods are developed to enhance FDD and increase the PV system reliability. Four types of advanced FDD can be distinguished:
A model-based diagnostics method;
Real-time measurements;
An Output Signal Analysis (OSA);
A machine learning-based diagnosis.
According to [
18], the Output Signal Analysis (OSA) assumes that any fault has an effect on the PV system output. The output current or voltage can be affected by a drop in values or a change in signal dynamics. A fitted analysis of these perturbations allows the extracting diagnostics behavior of each fault. An OSA takes the advantages to require a limited number of sensors which make it low-cost. With the accuracy of the detection and the low complexity of the installation, an OSA is a prime candidate for the on-line and real-time diagnosis. An OSA-based diagnosis remains noise-sensitive, and the recent and ongoing research aims to reduce the impact of noise on the detection capability using a novel approach [
21].
Then, different methods coming from cross-domain applications have been implemented for PV system fault diagnosis. Classification tools have already widely been tested to detect PV faults, such as Decision Tree [
22], the Probabilistic Neural Network classifier [
23], Random Forest [
24], the Artificial Neural Network classifier [
25], a One-class Support Vector Machine [
26] and Machine Learning Based on Gaussian Process Regression [
27]. Machine learning has been widely experienced these last years. Review papers list and compare the different machine learning technic in the case of PV fault diagnosis [
28,
29].
The method limitations are the learning data and preliminary analysis needs, which limits the FDD robustness and scalability.
Decomposition techniques have already been tested in PV FDD, particularly Wavelet Transform. Undecimated Wavelet Transform is used to detect PV power plant swag and swell [
21]. Wavelet Transform associated to Radial Basis Function Networks (RBFNs) has been applied to early failure detection and fault classification on a real PV plant [
30]. Multiscale wavelet decomposition is applied to achieve a non-sensitive-to-noise short-circuit, open-circuit and shading FDD [
31]. A statistical analysis such as Principal Component Analysis was performed for a shading diagnosis [
32]. These decomposition algorithms require complex functions, which limits the implementation convenience and increases the computational cost. Novel decomposition algorithms have been developed, as mentioned below. Empirical Mode Decomposition (EMD) is a recursive algorithm that decomposes a signal into several modes of separate spectral bands, named Intrinsic Mode Functions (IMFs). This signal processing tool was introduced in 1998 [
33] and is widely used in the diagnosis domain. In the case of PV faults, EMD is applied to transform a non-stationary time series into a stable time series that allows a comparison between a model and a real performance in order to establish a state of health index [
34]. The output current is decomposed by EMD, and a classification tool based on machine learning is added to localize the islanding and tripping of the distributed generation [
35]. EMD limitations are also well-known, which are the ending effect, mixing mode in non-stationary signals and sensitivity to noise and sampling. To overcome these limitations, variational mode decomposition (VMD) has been developed [
36]. Power quality disturbances in a Grid-Connected Distributed Generation System are detected using VMD [
37]. In two cases of EMD and VMD [
35,
37], fault identification is carried out by Decision Tree classification.
Among the important diversity of the methods, information theory offers innovative and accurate tools. Already present in other fields of research, entropy is widely used in medical applications and commonly used in other diagnosis applications. [
38] initiates the use of entropy in PV fault detection, to the specific case of arc fault detection. Sample entropy (SE) is performed to detect L-L, L-G, OC, weather disturbances, high impedance, blocking diode and partial shading faults [
39]. However, the faults are not isolated and localized. These works encourage drawing inspiration from other FDD application domains, taking advantage of entropy efficiency. Multiscale entropy is increasingly used in diagnosis research fields such as in biomedical time series, electroseismic time series, rotary machine vibrations and financial time series [
40,
41]. A multiscale approach allows evaluating entropy of different phenomena with different time scales, and the computation of global entropy is the sum of each time scales’ entropy. In order to analyze a biomedical time series involving different time scales, multivariate and multiscale dispersion entropy is developed [
42].
The superposition of VMD and different types of entropy computation are studied, such as multiscale dispersion entropy (MDE) to diagnose partial discharge [
43], Modified Multiscale Symbolic Dynamic Entropy or Generalized Composite Multiscale Symbolic Dynamic Entropy for planetary gearboxes fault diagnosis [
44] and, more recently, VMD is associated to Improved Multiscale Fuzzy Entropy to PV arc fault detection [
45]. The developed methods, even with real-time applicability, require a high sampling frequency and thus an addition of a sample device on a standard PV plant. Furthermore, the methods cannot be extended to all time scale faults.
On this basis, the presented work takes advantage of the techniques named above in order to detect the faults presented on a PV installation. VMD is applied to extract information-carrying signals, removing the non-informative noise with an IMF’s selection criterion. In addition, an MDE analysis is performed in order to identify PV faults. The proposed method tackles the machine learning limitations with no learning dataset, for the purpose of good flexibility and robustness. The proposed decomposition algorithm has the advantage of an ease of implementation, conversely to other decomposition tools such as WT. The multiscale approach will be used, allowing detecting faults with different time scales.
The proposed FDD requires only an additional software component. Using the electrical data commonly collected by the inverters, the FDD is easy to implement in new or existing PV plants. The algorithm has been tested on a real rooftop PV plant of 4 kW. The advantages of the proposed method are as follows:
Model identification is not necessary;
There is no dependence on the PV plant characteristics;
It is insensitive to weather variations;
It has a low computation cost.
The present work aims to synthetically present the used method and to clearly and methodically group all of the elements allowing rapid application and repeatability. Furthermore, the first results of the proposed method applied to the PV faults diagnosis are shown. This is, as far as we know, the first application of this method to shading faults.
The paper is organized as follows:
Section 2 introduces VMD, the MDE concept description and the proposed fault diagnosis strategy.
Section 3 presents the experimental setup and the data used to develop and validate the proposed method.
Section 4 is dedicated to the experimental results obtained using the proposed diagnosis method. The conclusion is provided in
Section 5.
5. Conclusions
In this paper, an innovative online Fault Detection and Diagnosis (FDD) method has been developed. This method relies on variational mode decomposition (VMD) and multiscale dispersion entropy (MDE). The values of entropy at specific scales have been chosen as quantitative indicators to determine the state of health of the system. The VMD-MDE-based diagnosis does not require any additional equipment or sensor, can easily be implemented on-line and has a very low computation cost.
The VMD-MDE-based diagnosis capability and accuracy have been investigated on a real aging PV plant, without additional equipment, and based on the inverter output current. The first results exhibit the promising perspectives of multiscale dispersion entropy, combined with variational mode decomposition. Variational mode decomposition exposes its ability to pre-process PV data, in spite of very low frequency sampling and poor precision data. The most informative IMF has been selected and a multiscale entropy analysis has provided distinct curves trends in accordance with the experimental conditions.
The present method exclusively includes the inverter output current as the input. The information regarding the temperature, humidity, irradiance, plant structure or panel characteristics is not necessary, which ensures method robustness and scalability. The algorithms are easy to implement and require low computing capacity. Moreover, although developed under MatLab©, the method can be easily transposed into other programming languages, notably in Python, which make the proposed method highly adaptive.
The first promising results can be supplemented in order to produce a complete automated fault detection. Advanced classification approaches can be investigated. Moreover, in order to assess the method applicability on other faults, it will be tested on diode and hot-spot faults. The effectiveness of the method on simultaneous faults could be studied. On the assumption that the method is blind to meteorological conditions and PV plants technical characteristics, tests on different PV plant scales and technologies will be processed to confirm its robustness and scalability.