Method for detecting faults in a photovoltaic installation

Abstract

The invention relates to a method for detecting defects in a photovoltaic installation comprising several strings of photovoltaic panels. The method comprises the following steps: - collection of: - structural information relating to the photovoltaic installation, - meteorological data; and - operating data for each string of photovoltaic panels; - from the structural information and meteorological data, determination of a fitted predictive production model for the photovoltaic installation; - identification of the presence of defects on each of the strings of the photovoltaic installation by submitting the fitted predictive production model, the meteorological data, and the operating data to at least one pre-trained machine learning algorithm so as to identify each string of panels containing defects. Figure for the abstract: Fig. 3

Description

Title of the invention: Method for detecting faults in a photovoltaic installation. Technical field

[0001] The present invention relates to a method for detecting faults in photovoltaic installations, and devices for implementing such a detection method. Previous technique

[0002] The detection and diagnosis of faults are essential to ensure the normal operation of a photovoltaic installation throughout its lifetime, classically around 25 years, while minimizing significant production losses.

[0003] Furthermore, these approaches are essential to avoid destructive consequences and potential hazards to personnel coming into contact with the installations. In particular, fault detection and diagnosis are necessary to prevent human risks and / or energy losses that could cause fires.

[0004] For these various reasons of safety and productivity, detection and diagnostic methods have been developed to determine, with the greatest possible certainty, that faults have occurred in a photovoltaic installation and that the installation is therefore no longer operating within its optimal range. Conventional methods typically implemented rely on prior knowledge, estimations, and field measurements, and utilize visual methods, image processing methods, methods based on detection at the level of the installation's electrical operation, techniques based on the implementation of protective devices, or techniques for detecting defects in antireflective coatings. Several machine learning approaches also exist for detecting certain faults in photovoltaic installations.However, existing machine learning approaches are proving insufficient, failing to detect or differentiate certain defects, such as snail trails. In particular, these approaches are sensitive and dependent on the specific photovoltaic installation and the sunlight conditions under which the installation diagnosis is performed. For example, these approaches do not necessarily distinguish between a drop in production due to a change in sunlight conditions and a drop in production due to a fault in the installation.

[0005] However, these conventional methods of fault detection and diagnosis are limited. In particular, these conventional methods reach their own limits when fault conditions occur in several elements of the installation (inverters, string, batteries, connectors, etc.), or when multiple faults with different levels of impact are present, or when faults have expressions in different domains (frequency or time) or with signatures that do not come from a single type of data such as electrical and / or thermal signatures.

[0006] Conventional methods also have difficulty detecting the occurrence of several simultaneous faults or primary faults that generate more serious secondary faults. In particular, there are faults or combinations of faults that produce an electrical response almost identical to the electrical response of a non-faulty installation, and these conventional methods cannot identify them.

[0007] Indeed, the methods conventionally implemented analyze an overall response of the photovoltaic installation and do not allow for the differentiation of the various signatures. They require checking each panel by visual analysis or image analysis, which is time-consuming and expensive.

[0008] Among the defects that are difficult to detect are corrosion, snail trails, microcracks, delamination, yellowing, failures in bypass diodes, panel back failures, and hot spots. This list is not exhaustive.

[0009] Finally, some of these conventional methods are time-consuming to implement and / or require a shutdown of all or part of the photovoltaic installation during the diagnostic period, thus implying significant regular production losses. Description of the invention

[0010] There is therefore a need for an improved method enabling the detection of defects in a photovoltaic installation.

[0011] In particular, there is a need for a method enabling the detection of all types of defects, and in particular defects not detectable by conventional methods, such as the "Snail Trail".

[0012] There is also a need for a method enabling the detection of faults for any type of photovoltaic installation, regardless of the age of the installation or the technology used in the installation.

[0013] In general, there is a need for an improved fault detection method that can be easily implemented for any photovoltaic installation and that does not involve a loss of production.

[0014] The invention aims to meet all or part of these needs. Summary of the invention

[0015] According to a first aspect of the invention, the invention relates to a method for detecting defects in a photovoltaic installation comprising several strings of photovoltaic panels, the method comprising the following steps: - collection: - structural information relating to the structure of the photovoltaic installation; - meteorological data; - operating data, for example voltage and current, for each string of photovoltaic panels in the photovoltaic installation; - from structural information and meteorological data, determination of an adjusted predictive production model of the photovoltaic installation, modeling an optimal theoretical electricity production of the photovoltaic installation with regard to meteorological data, the adjusted predictive production model being determined assuming that the photovoltaic installation has no defects; - identification of the presence of defects on each of the strings of the photovoltaic installation by submitting the adjusted predictive production model, meteorological data and operating data to at least one pre-trained machine learning algorithm so as to identify each string of panels containing defects, called a "defective string", and preferably the types of defects.

[0016] In a particular embodiment, the adjusted predictive production model is determined by: - a prior generation of a predictive reference model determined based on the structural information of the photovoltaic installation and then - an adjustment of the reference predictive model based on meteorological data, the reference predictive model modeling a theoretical electricity production profile for the photovoltaic installation considering the absence of defects on all photovoltaic panel strings and independently of meteorological data.

[0017] Optionally, the method includes, after the step of identifying the presence of defects, a step of presenting a report indicating the defective chain(s), preferably the types of defects, and preferably one or several maintenance actions, the maintenance actions being determined from the identification of the presence of defects on the defective chain(s), and possibly from the adjusted predictive production model and operating data relating to each defective chain, preferably by means of a statistical analysis, preferably by means of a residual-based method.

[0018] One of the main advantages of the invention is that it enables universal detection. In other words, the method allows for the detection of defects on any type of photovoltaic installation, regardless of the installation's topology, technology, age, or even the weather conditions to which it is subjected. To achieve this, the invention is based on the joint analysis, preferably using a supervised machine learning algorithm and an unsupervised machine learning algorithm, of structural information relating to the installation's structure, operating data from the installation, and meteorological data.

[0019] The combined use of an unsupervised machine learning algorithm and a supervised machine learning algorithm, in particular in cascade, accelerates the identification of faults that seriously impair the production of the photovoltaic installation by limiting the need for computing resources, in particular thanks to the unsupervised machine learning algorithm, while allowing the fine detection of faults and preventive diagnosis of faults, in particular thanks to the supervised learning algorithm.

[0020] In particular, the invention makes it possible not only to detect that a photovoltaic installation has one or more defects but also to locate the string or strings of photovoltaic panels having these defects and to identify what type of defects it is.

[0021] Preferably, the meteorological data may include at least irradiance and ambient temperature. In one embodiment, the meteorological data include at least irradiance, ambient temperature, and wind speed. This list is not exhaustive.

[0022] Structural information may include: the brand and / or age of the photovoltaic installation and / or the material of the photovoltaic panels and / or the technology of the photovoltaic panels and / or the tilt of the photovoltaic panels and / or the topology of the photovoltaic installation and / or information relating to the replacement of photovoltaic panels and / or the number of panels per string and / or the number of strings in the photovoltaic installation. This list is not exhaustive.

[0023] Preferably, meteorological data and / or operating data, preferably meteorological data and operating data are acquired at a frequency greater than or equal to 1 Hz, preferably greater than or equal to 10 Hz, preferably greater than or equal to 100 Hz, preferably greater than or equal to 500 Hz, preferably even greater than or equal to 1 kHz.

[0024] The collection of operating data is preferably carried out using at least one analog sensor.

[0025] The collection of structural information can be carried out using a graphical interface and / or a database. For example, the collection of structural information is performed by an operator. All or part of the structural information can be extracted from a database.

[0026] The invention notably enables the detection of defects in photovoltaic panel strings. Such detection in a photovoltaic panel string allows for more precise and rapid location of defects.

[0027] In a particular embodiment, the operating data are divided into time windows, the identification of the presence of faults being carried out for each of the time windows, at least one machine learning algorithm being pre-trained for each of said time windows.

[0028] The number of time windows in a day can be greater than or equal to 2 and / or less than or equal to 50.

[0029] The operating data is, for example, divided into four time windows. In particular, it can be divided into a window corresponding to the morning, a window corresponding to midday, a window corresponding to the afternoon, and a window corresponding to the night. This division is given by way of example.

[0030] Preferably, the time windows are such that in normal operation, and under constant weather conditions, the production of the installation over the entire duration of the time window is substantially constant.

[0031] Preferably, time windows allow for the isolation of specific operating cycles of the photovoltaic system. In particular, a photovoltaic system does not produce power in the same way at the end of the day as during the day. Fault analysis by time window thus enables the detection of faults that are only detectable when the photovoltaic system is producing little power or, conversely, when it is operating at maximum capacity.

[0032] The operating and meteorological data can be normalized.

[0033] The operating and meteorological data can be pre-processed so to extract attributes, the identification of the presence of defects being carried out by submitting these attributes to at least one pre-trained machine learning algorithm so as to identify each defective string, and preferably the types of defects, from a training database.

[0034] The attributes can be determined by decomposing the operational and meteorological data, for example via wavelet decomposition, and / or can be determined by means of an attribute selection algorithm, for example by analysis of attribute correlations and / or by analysis of variance.

[0035] Advantageously the invention includes an analysis of detected, located and identified defects, so as to prioritize maintenance and repair tasks to be carried out.

[0036] The invention thus provides a decision support method enabling optimization of repairs and therefore optimization of the operation of the photovoltaic installation.

[0037] The prioritization of maintenance and repair tasks to be carried out may depend on the type or types of defect identified, for example based on the presence of snail marks, broken glass, delamination, corrosion, yellowing, microcracks in solar cells, cracks in solar cells, potential-induced degradation, a defective junction box, cracks in a backsheet.

[0038] The prioritization of maintenance and repair tasks may depend, in particular, on the production or safety benefits that the tasks would provide, the cost incurred by the tasks, the time required for the repair, and whether or not it is necessary to shut down the photovoltaic installation to perform a task, considered individually or in combination. Specifically, prioritization may depend on a ratio between a production or safety benefit and a cost, and / or a ratio between a production or safety benefit and the time required for the repair.

[0039] Preferably, with each implementation of the detection method, the collected data and the identification of defects are stored in such a way as to allow the machine learning algorithms to be updated. Optionally, the identification of defects is validated by a domain expert, for example by verifying that the actual defect corresponds to the identified defect.

[0040] The invention also relates, according to another of its aspects, to a computer program comprising instructions which, when the program is executed by a computer, lead the latter to implement a method of detecting defects according to the invention.

[0041] Another aspect of the invention relates to a measuring device for collecting meteorological and operational data in a photovoltaic installation comprising several strings of photovoltaic panels connected in parallel to each other at a junction box, the device comprising: - a weather station for acquiring meteorological data, and - a module for measuring operating data, including voltage and current, the measuring module being configured to acquire data at a frequency greater than or equal to 1 Hz, preferably greater than or equal to 10 Hz, preferably greater than or equal to 100 Hz, preferably greater than or equal to 500 Hz, preferably even greater than or equal to 1 kHz, and to be connected in series between at least some of the panel strings of the photovoltaic installation, preferably all the panel strings, and the junction box, the measuring module being configured to acquire operating data for each panel string connected to the measuring device.

[0042] The weather station can acquire meteorological data within a radius of approximately 100 m around the measuring device, preferably within a radius of approximately 10 m.

[0043] The weather station can be configured to acquire meteorological data at the same frequency as the acquisition frequency of the measurement module.

[0044] The measurement module may advantageously include at least one analog sensor. Analog sensors allow, among other things, the acquisition of data at a high sampling frequency, in particular above 1 kHz and below 1 MHz.

[0045] The measuring module and the weather station can be connected by wired communication means and / or by wireless communication means.

[0046] The measurement module and / or the weather station can be configured to allow the storage and / or processing of data. In a particular embodiment, only the measurement module or the weather station is configured to allow the storage and / or processing of operational and meteorological data.

[0047] In one embodiment, the device is configured so as to be connected to at least one string, preferably at least 8, preferably at least 10, preferably at least 15, even better at least 20, even better at least 27 and / or less than 500, or less than 100 strings of photovoltaic panels.

[0048] In one embodiment, the device is configured to be connected to at least 25% of the photovoltaic panel strings of a photovoltaic installation, preferably at least 50%, and even better at least 75%. Preferably, the device is configured to be connected to all the photovoltaic panel strings of a photovoltaic installation.

[0049] The connection of the device to the panel strings can advantageously be made at the junction box of a photovoltaic installation, the box junction being used to allow the parallel connection of the strings of panels in a photovoltaic installation.

[0050] According to yet another aspect of the invention, the invention relates to a fault detection device in a photovoltaic installation comprising several strings of photovoltaic panels connected in parallel to each other at a junction box, the device comprising: - a measuring device for collecting meteorological and operational data according to the invention, and - a computer program comprising instructions which, when the program is executed by a computer, lead the computer to implement a method of defect detection according to the invention.

[0051] The fault detection device may include, in particular, means for acquiring structural information of the photovoltaic installation, preferably means for acquiring information comprising a graphical interface.

[0052] The invention also relates to a photovoltaic installation comprising several strings of photovoltaic panels in parallel connected to a junction box, in which a fault detection device according to the invention is connected in series between all or part of the strings of panels and the junction box. Definitions

[0053] Structural information defines the intrinsic characteristics of the photovoltaic installation, and in particular the characteristics relating to the technology of the photovoltaic panels and their arrangement. Among the structural information, one can cite, among other things, the technology of the photovoltaic panels, the number of photovoltaic panels in series for each string of photovoltaic panels, the age of the photovoltaic installation, the age of the photovoltaic panels, information relating to replacements carried out in the photovoltaic installation, the tilt and / or orientation of the photovoltaic panels, the average, minimum and / or maximum height of the photovoltaic installation or of the photovoltaic panels relative to the ground.

[0054] A photovoltaic panel technology depends on the type of cells used to form the panel. Depending on the cells used, a photovoltaic panel does not produce energy in the same way and does not have the same profitability.

[0055] The technology of a photovoltaic panel can be based on panels made of monocrystalline silicon cells, polycrystalline, i.e. based on mixed silicon crystals, amorphous, i.e. based on cells formed by vaporizing a layer of silicon on an amorphous surface, such as glass, steel or polypropylene, or hybrid, i.e. based on heterojunction or tandem cells.

[0056] An adjusted predictive production model provides a prediction of the production of a string of panels or a photovoltaic installation under a theoretical ideal situation for given weather conditions; in other words, under a situation in which the string of panels or the photovoltaic installation, respectively, is free of defects and operates at its maximum capacity under the given weather conditions. An adjusted predictive production model is said to be "adjusted" because it is adapted to the given weather conditions.

[0057] A reference predictive model provides a prediction of the production of a string of panels or a photovoltaic installation in an ideal theoretical situation, independent of weather conditions, in other words in a situation for which the string of panels or respectively the photovoltaic installation has no defects and operates at its maximum capacity regardless of the given weather conditions.

[0058] The reference and adjusted predictive models simulate an optimal theoretical electricity production for a string of panels or for a photovoltaic installation.

[0059] A reference predictive model or an adjusted production predictive model may in particular take the form of a power curve. Brief description of the drawings

[0060] The invention will be better understood upon reading the detailed description that follows, the non-limiting examples of its implementation, and upon examination of the accompanying drawing, on which: - [Fig. 1] [Fig. 1] represents the steps of a defect detection method according to the invention, - [Fig.2] [Fig.2] is a block diagram representing the operation of a computer program for implementing a fault detection method according to the invention, - [Fig.3] [Fig.3] represents a photovoltaic installation comprising a measuring device according to the invention, - [Fig.4] Figure 4 schematically illustrates a fault detection device according to the invention, - [Fig.5] Figure 5 illustrates the training of a machine learning algorithm and the enrichment of a training database, - [Fig.6] [Fig.6] is an example of fault detection according to the invention, - [Fig.7] [Fig.7] illustrates the acquisition of operating data from a panel, - [Fig.8] [Fig.8] illustrates the acquisition of meteorological data, - [Fig.9] [Fig.9] illustrates adjusted predictive models representative of the production of two determined healthy panels compared with the actual production values ​​of healthy panels measured using a sensor.

[0061] In the following description, identical elements or elements with identical functions bear the same reference numeral. Their description is not repeated opposite each figure; only the main differences between the embodiments are mentioned. Detailed description Detection method

[0062] Fig. 1 illustrates the steps of a method 10 for detecting faults in a photovoltaic installation according to the invention.

[0063] A method 10 of fault detection according to the invention comprises - the collection 100 of structural information 300 relating to the structure of the photovoltaic installation, meteorological data 220 and operating data 200; - the determination of an adjusted predictive model 102; - the identification of the presence of defects 104, and possibly types of defects, and - optionally, the generation of a report 106. Data and information collection

[0064] The collection includes the collection of structural information relating to the structure of the photovoltaic installation 300, the collection of meteorological data 220, and the collection of operating data of the photovoltaic installation 200.

[0065] The structural information 300 may include: a brand and / or age of the photovoltaic installation and / or a material of the photovoltaic panels and / or a technology of the photovoltaic panels and / or an inclination of the photovoltaic panels and / or a topology of the photovoltaic installation and / or information relating to a replacement of photovoltaic panels and / or a number of panels per string of panels and / or a number of strings in the photovoltaic installation.

[0066] Preferably, the structural information 300 is entered by an operator, for example via an interface 30b, and / or extracted from a database 30a.

[0067] In particular, structural information can be extracted from field reports carried out for the photovoltaic installation and / or from technical data sheets of the photovoltaic installation.

[0068] Preferably, the meteorological data 220 include at least one irradiance and one ambient temperature in the region of the photovoltaic installation. The Meteorological data may include at least irradiation, ambient temperature, and wind speed.

[0069] Fig. 8 illustrates an example of meteorological data 220 consisting of temperature acquisition 220i, irradiation acquisition 2202 and wind speed acquisition 2203, within a radius of 100 m around a photovoltaic installation.

[0070] Preferably, meteorological data are acquired using a weather station 22.

[0071] The meteorological data 220 are preferably acquired at a frequency greater than or equal to 1 Hz, preferably greater than or equal to 1 kHz.

[0072] The operating data 200 preferably include the voltage and / or current flowing in one or more strings of panels of the photovoltaic installation for which the presence of faults is to be detected.

[0073] In general, the operating data 200 are chosen so as to allow the determination of the electrical production of the photovoltaic installation over time.

[0074] Fig. 7 illustrates an example of operating data 200 composed of acquisitions of current intensities 200i, 2002, 2003, 2004 and a voltage 2005 representative of the operation of a string of photovoltaic panels.

[0075] The operating data 200 are preferably acquired at a frequency greater than or equal to 1 Hz, preferably greater than or equal to 1 kHz.

[0076] In one embodiment, the operating data 200 and the meteorological data 220 are acquired at the same frequency.

[0077] The operating data 200 are preferably acquired by means of at least one analog sensor. The analog sensor can advantageously be configured for acquiring operating data at a frequency greater than or equal to 1 Hz, preferably greater than or equal to 1 kHz, and / or less than or equal to 1 MHz.

[0078] The collection of operational and / or meteorological data can be carried out over a specified period of time, preferably continuously, exceeding 24 hours.

[0079] The collection of operational and / or meteorological data can be carried out over a period of time, preferably continuously, of less than one week, or even less than 72 hours, or even less than 48 hours.

[0080] The collection of operational and / or meteorological data can be carried out over a period of time, possibly discontinuous, including at least one sunrise and / or sunset.

[0081] The collection of operational and / or meteorological data can be carried out over a period of time, possibly discontinuous, during which it is not raining.

[0082] Preferably, at the end of the collection, a set of normalization equations is applied to the operational and / or meteorological data, preferably to the operational and meteorological data.

[0083] Normalization is a step classically carried out in the field of data processing to obtain data conforming to a particular format without loss of information.

[0084] For example, for a set of measures in the set of real numbers, a normalization may consist of applying a set of equations so that the measures ultimately belong to the interval [0; 1].

[0085] The normalization equations may depend on minimum and maximum threshold values ​​of the data; these threshold values ​​may be a function of the operating ranges of the sensors. For example, for a power density sensor operating over a range of values ​​[0; 1500 W / m²], the minimum and maximum threshold values ​​considered for the normalization equations may be defined as 0 W / m² and 1350 W / m², respectively. Such a normalization method makes it possible to obtain a good distribution of the normalized data in the interval [0; 1].

[0086] This not only improves the visualization of data relating to the photovoltaic installation, but also improves the accuracy of fault detection algorithms. Determination of the adjusted predictive model

[0087] The determination 102 of an adjusted predictive model includes the determination of an adjusted predictive production model 320 of the photovoltaic installation, from structural information 300 and meteorological data 220.

[0088] The objective of this step is to enable the analysis of any type of photovoltaic installation, regardless of the number of panel strings it contains, the technology used, its orientation, its age, etc., and regardless of the environment in which it is located, by a modeling capable of predicting the production of the photovoltaic installation whose defects we wish to detect.

[0089] For this, one method consists of determining a production of a theoretical installation of the same type as the photovoltaic installation whose defects we wish to detect, that is to say whose structural information is similar, preferably identical, the theoretical installation corresponding to an installation in which all the strings of photovoltaic panels are sound and therefore operate at their maximum capacity.

[0090] Production can be expressed as a power.

[0091] The adjusted predictive model can take the form of a power curve, as illustrated in Fig. 9b of [Fig.9].

[0092] The determination of the adjusted predictive model 320 can be carried out via the exploitation of knowledge-based models, for example by means of generative models such as described in Theis, et al. "A note on the evaluation of generative models." arXiv preprint arXiv: 1511.01844 (2015), by means of domain adaptation learning models considering the invariant domain such as described in Baktashmotlagh, M. et al. "Unsupervised Domain Adaptation by Domain Invariant Projection" 2013 IEEE International Conference on Computer Vision, pages 769-776, 2013, or preferably by means of equations simulating the behavior of a photovoltaic installation such as described in Herteleer, B., et al. "Normalised efficiency of photovoltaic Systems: Going beyond the performance ratio." Solar Energy, vol. 157, pages 408-418, 2017.

[0093] The approach using predetermined equations simulating the operation of a photovoltaic installation allows in particular a high accuracy of prediction of the behavior of a photovoltaic installation and is also easily adaptable to real data measured on any type of photovoltaic installation.

[0094] In one embodiment of the invention, a reference predictive model is determined from the structural information of the photovoltaic installation, representing the ideal operation of a photovoltaic installation comprising healthy panels and having the same intrinsic characteristics as those collected in the structural information relating to the photovoltaic installation, then from the reference predictive model, an adjusted predictive model is determined from meteorological data, determining an ideal production profile of the photovoltaic installation under meteorological conditions as collected.

[0095] Preferably, the adjusted predictive model allows for modeling the ideal operation of each of the photovoltaic system's chains. Ideal operation means fault-free operation of the system.

[0096] In the absence of defects, the adjusted predictive model therefore corresponds substantially to the actual production.

[0097] Figure 9 illustrates a comparison between fitted predictive models (Fig. 9b) and operating data of defect-free photovoltaic panel strings (Fig. 9a). The fitted predictive models model the current intensity of the photovoltaic panel strings over time, and the operating data are data acquired during the actual operation of the photovoltaic panel strings. Figure 9c is a superposition of the fitted predictive models and the operating data. In the example in Figure 9, the panel strings do not showing no defects, the current intensity curves representing the predictive models and the actual operating data are virtually identical.

[0098] The determination of the adjusted predictive model is implemented by computer means 32, the computer means preferably exploiting methods based on knowledge-based models.

[0099] The computer means comprise at least one computer program including lines of code which, when implemented by a computer, allow the determination of the adjusted predictive model. Fault identification

[0100] The identification of defects involves determining the presence of defects on the string(s) of the photovoltaic installation by submitting the adjusted predictive production model, meteorological data and operating data to at least one pre-trained machine learning algorithm so as to identify one or more strings of panels containing defects, referred to as "defective string(s)".

[0101] By submission of the adjusted predictive production model, meteorological data and operational data, it is necessary to understand submission of raw, pre-processed data and / or of attributes extracted and / or selected from the raw or pre-processed data.

[0102] Preferably, the identification of defects also includes the determination of the types of defects detected.

[0103] The identification of the presence of defects, and possibly the identification of the types of defects, can be carried out using several machine learning algorithms. Preferably, at least one first machine learning algorithm identifies the presence of defects and then at least one second machine learning algorithm identifies the type of defect.

[0104] If the fault cannot be identified, a manual identification step, preferably performed by a qualified technician, may be carried out. Preferably, after such a manual identification step, the machine learning algorithm is enhanced to enable the future detection of such a fault.

[0105] Preferably, at least a second machine learning algorithm is a supervised machine learning algorithm.

[0106] A type of defect may be snail tracks, broken glass, delamination, corrosion, yellowing, microcracks in solar cells, cracks in solar cells, potential-induced degradation, a defective junction box, cracks in a backsheet.

[0107] In general, a machine learning algorithm is trained from a training database DB so as to learn to be able to identify output data from input data.

[0108] For training a supervised machine learning algorithm, the training database DB comprises an input dataset and an output dataset. For training an unsupervised machine learning algorithm, the training database DB comprises at least one input dataset.

[0109] Training a machine learning algorithm consists of providing at least a part of the input data set and possibly at least a part of the output data set from the training database DB, so that the algorithm learns to determine, from an input data, an output data set.

[0110] In the present case, an input data includes an adjusted predictive production model, meteorological data and operating data, or attributes extracted from said operating and meteorological data, raw or pre-processed, and an output data corresponds to the detection of the presence of defects on a chain and possibly the determination of the type of defect detected on the chain.

[0111] Preferably, the input data are pre-processed. Typically, the operational and / or meteorological data are at least normalized.

[0112] The input data can be decomposed, for example into time windows and / or attributes can be extracted and / or selected from the operating data and / or meteorological data.

[0113] In particular, attributes can be extracted from operational and / or meteorological data, for example by applying an attribute extraction algorithm, in particular via a multi-resolution decomposition of operational data followed by statistical attribute extraction.

[0114] In particular, a multi-resolution decomposition allows for a time and frequency analysis of the operating data.

[0115] The attributes can be specifically chosen from: - an asymmetry of the data compared to the average of said data; - a flattening measuring the peak of a probability distribution of the data; - a variance in the data; - a peak-to-peak distance, corresponding to the distance between the peak of highest amplitude and the valley of lowest amplitude; and - an energy calculated from the data.

[0116] This list is indicative only and is not exhaustive.

[0117] Preferably, when the data are decomposed into time windows, at least one machine learning algorithm is built for each of the time windows.

[0118] Such a decomposition makes it possible in particular to analyze the behavior of photovoltaic panels under different irradiation conditions of the day, in particular: morning, noon, afternoon and night.

[0119] An attribute extraction can be performed in each of these windows.

[0120] Machine learning can be performed using an algorithm employing k-Nearest Neighbors algorithms (kNN), Support Vector Machines (SVMs), and / or Decision Trees (DTs).

[0121] A combination of these algorithms with a relative majority vote can be carried out.

[0122] “k - Nearest -Neighbors”

[0123] The kNN is one of the most widely used models for classification due to its simplicity. For a given input sample x' = [x} x „F '] with attributes for classification, the kNN finds its nearest neighbors among the samples already classified on the basis of a distance metric, usually Euclidean distance.

[0124] A kNN assigns the input sample to the most common class among its k nearest neighbors.

[0125] As the value of k increases, the kNN model can tolerate more noise because the impact of variance caused by random error is reduced, but there is a risk of missing a small but important model in the data. The key to choosing an appropriate value of k is to find a balance between overfitting and underfitting.

[0126] “Vector Machines Support”

[0127] Selective Verification Machines (SVMs) are among the most powerful and complex classification algorithms. The SVM classifier searches for one or a set of optimal separation hyperplanes, maximizing the margin between classes constituting the outputs. The SVM is based on the idea that it is possible to map an input space into a higher-dimensional attribute space via a kernel function, and then apply a linear SVM to this space, thus making it possible to distinguish classes that were not linearly separable.

[0128] “Decision Trees”

[0129] The DT model is characterized by decision nodes with multiple branches, and leaf nodes corresponding to the outputs of the decision nodes. DTs are created from a recursive splitting of the sample set based on a set of attribute-related splitting rules.

[0130] There are many variants of DT in the literature, such as Iterative Dichotomies 3 (ID3), ID3 Successor (C4.5), Chi-Square Automatic Interaction Detector (CHAID), Classification and Regression Tree (CART). This list is not exhaustive.

[0131] Majority vote

[0132] In a particular embodiment, the pre-trained machine learning algorithm for identifying one or more chains of panels containing defects includes the implementation of several classifiers, for example the three classifiers kNN, SVM and DT, combined using the majority voting principle.

[0133] Each of the classifiers can indeed be considered as a voter, the prediction results, i.e. the outputs, of each classifier being compared to determine a final prediction.

[0134] The principle of majority voting can be a weighted majority vote, a relative majority vote or an absolute majority vote.

[0135] Once trained, the algorithm can be used to predict an output, here a detection of the presence, and possibly the type, of defects present on one or more string(s) of photovoltaic panels of a photovoltaic installation, from operating data, meteorological data, and / or attributes extracted from this data, possibly for a time window, and the adjusted predictive model.

[0136] For each subsequent operation of the machine learning algorithm training, the input data E and the output data S can be recorded in the database DB, as shown schematically in [Fig.5].

[0137] The machine learning algorithm can be updated, thanks to the enrichment of the database DB.

[0138] Preferably, output data S is checked by an operator before being added to the database DB. Generating a report

[0139] The method preferably includes the optional step of presenting a 360 report indicating the defective chain(s), and preferably the types of defects.

[0140] In a preferred embodiment of the invention, the method comprises determining one or more maintenance actions, the maintenance actions being determined from the identification of the presence of defects on the defective chain(s), and optionally from the adjusted predictive production model and operating data relating to each defective chain, preferably to by means of a statistical analysis, preferably using a residual-based method.

[0141] For example, the determination of maintenance actions may include: - the determination of a power threshold value, the calculation of a centroid of the power of healthy panel strings or a centroid of a theoretical power for a panel string determined for example from the adjusted predictive model, healthy panel strings being the strings for which no fault has been identified; - for each defective panel string, the power of the defective panel string is compared with the power threshold value in order to determine the magnitude of the difference between the power of the defective string and the power threshold value; The maintenance priority of defective panel chains is determined by the magnitude of the difference (the greater the magnitude, the higher the maintenance priority).

[0142] The threshold value may be a centroid of the power of the healthy panel strings, the healthy panel strings being the panel strings for which no fault has been identified.

[0143] The threshold value can be a centroid of the theoretical powers calculated for each of the panel strings of the installation, the theoretical powers being able to be determined from the adjusted predictive model.

[0144] In one embodiment, the threshold value is a centroid of the theoretical powers and the powers of the healthy panel strings. Devices Computer program

[0145] Figure 2 schematically represents a computer program 1 comprising instructions which, when the program is executed by a computer, lead the computer to implement a fault detection method according to the invention.

[0146] The program includes a prediction module 32 which makes it possible to determine the adjusted predictive production model 320, from the collected meteorological data 220 and the collected structural information 300.

[0147] The program includes a detection module 34 that enables the identification of the presence of faults on the string(s) of the photovoltaic installation and preferably the fault types 340, based on the adjusted predictive production model 320, meteorological data 220, and operating data 200 and / or attributes extracted from said data 220;200. The detection module includes at least one pre-trained machine learning algorithm so as to identify a or chains of panels containing defects, referred to as a "defective chain", and preferably the types of defects detected.

[0148] Program 1 may include a module for generating a report 36 determining maintenance actions, based on the identification of the presence and possibly the types of defects 340, and possibly the adjusted predictive production model 320, and / or operating data 200.

[0149] The generation module 36 preferably allows the presentation of the 360 ​​report to an operator, for example, by sending the 360 ​​report to a graphical interface 38, such as a computer screen, a mobile phone, a tablet or any other well-known means of display. Photovoltaic installation

[0150] Figure 3 illustrates a photovoltaic installation 4 comprising strings 40 of photovoltaic panels, each formed of one or more photovoltaic panels 42. The strings 40 of photovoltaic panels are connected to a junction box 44.

[0151] The photovoltaic installation classically includes an inverter 46.

[0152] In [Fig.3], a measuring device 2 for collecting meteorological and operational data is positioned on the photovoltaic installation 4.

[0153] The invention also relates to a photovoltaic installation 4 comprising such a measuring device 2 according to the invention. Measuring device

[0154] A measuring device 2 according to the invention comprises at least: - a weather station 22 for the acquisition of meteorological data 220, and - an operating data measurement module 20, in particular of voltage and current, the measurement module being configured to acquire data at a frequency greater than or equal to 1 Hz, preferably greater than or equal to 1 kHz, and to be connected in series between at least some of the strings 40 of panels of the photovoltaic installation, preferably all the strings of panels, and the junction box 44 so as to acquire the operating data 200 for each string of panels connected to the measuring device.

[0155] In a preferred embodiment of the invention, the measuring module 20 comprises one or more analog sensor(s), in particular for measuring the voltage and current for each string of panels in the photovoltaic installation.

[0156] The measuring device advantageously includes means for connecting 24, preferably switchable, the measuring module 20 with the strings 40 of photovoltaic panels.

[0157] The connection means 24 may include means for connecting the measuring module 20 to the meteorological station 22, the connection means being able to include wireless and / or wired communication means.

[0158] The connection means 24 may include means for connecting the measuring module 20 and / or the weather station 22, with a battery 26.

[0159] Preferably, the measuring module 20 and the weather station 22 are configured so as to be electrically powered independently of each other.

[0160] The battery advantageously provides autonomy to the measuring device 2. Fault detection device

[0161] The [Fig.4] is a block diagram of a fault detection device 3 according to the invention.

[0162] A fault detection device 3 comprises: - a measuring device 2 according to the invention; - a computer program 1 for implementing a method of detecting defects 10 according to the invention.

[0163] A fault detection device may include one or more features of a computer program 1 for implementing a detection method as described above. Examples

[0164] Figure 6 illustrates an example of classification for determining the presence of the "snail tracks" type defect on a photovoltaic panel in a string of photovoltaic panels.

[0165] The operating data 200 are presented in the form of a time curve. The curve illustrates the current intensity, in amperes, of a panel during one day. The operating data 200 are divided into 4 time windows 200a, 200b, 200c and 200d.

[0166] For each time window, three classifiers are applied to the operating data from chains of panels that have been previously detected as defective.

[0167] Submitting the operating data to the three classifiers allows the types of faults to be determined.

[0168] In this case, a kNN, an SVM, and a DT are implemented to determine the presence of a "snail trail" type defect or the absence of a defect. In the example, if the classifier identifies the presence of a defect, its output is "snail trail," while in the absence of a defect, its output is "clean."

[0169] A majority vote is then implemented, based on the classifier outputs, for each of the time windows. This yields a prediction of the presence of a defect for each of the time windows.

[0170] Subsequently, an absolute majority vote is applied to a portion of the predictions obtained for each time window, determining whether the panel string for which the operating data was acquired is defective or not. In the example, at least one panel in the panel string has a snail-track defect.

[0171] As is clear, the invention allows for finer, faster and more precise detection than conventional methods.

[0172] Furthermore, the invention does not require a production stoppage, as the data can be acquired during the operation of the photovoltaic installation.

[0173] Furthermore, unlike many well-known methods that rely on the history of the photovoltaic installation to accurately model its aging, it is not necessary to have the complete history of the installation to implement the fault detection method according to the invention. Unlike these methods, which can only be applied with very precise knowledge of the photovoltaic installation, the proposed method makes it possible to diagnose faults on any type of installation, without the need for an exhaustive history.

[0174] Obviously, the invention is not limited to the embodiments described.

[0175] In particular, once a defect has been detected at the level of a string of panels, a person skilled in the art would have no difficulty adapting the method as described to apply it not to the photovoltaic installation to detect a defect at the level of a string of panels but to the string of panels to detect on which panel a defect is present.

[0176] Other types of machine learning algorithms could also be implemented, either in combination or alone.

Claims

Demands

1. A method for detecting faults (10) in a photovoltaic installation comprising several strings of photovoltaic panels, the method comprising the following steps: - collection of: - structural information (300) relating to the structure of the photovoltaic installation, - meteorological data (220); and - operating data (200) for each string of photovoltaic panels of the photovoltaic installation; - from the structural information and meteorological data, determination of an adjusted predictive production model of the photovoltaic installation, said adjusted predictive production model modeling an optimal theoretical electricity production of the photovoltaic installation with regard to the meteorological data, the adjusted predictive production model being determined assuming that the photovoltaic installation does not have any faults;- identification of the presence of defects on each of the strings of the photovoltaic installation by submitting the adjusted predictive production model, meteorological data and operating data to at least one pre-trained machine learning algorithm so as to identify each string of panels containing defects, referred to as a "defective string", and preferably the types of defects; - optionally, presentation of a report indicating the defective string(s), preferably the types of defects, and preferably one or more maintenance actions, the maintenance actions being determined from the identification of the presence of defects on the defective string(s), and possibly from the adjusted predictive production model and operating data relating to each defective string, preferably by means of a statistical analysis, preferably by means of a residual-based method.

2. A detection method according to claim 1, wherein the meteorological data includes at least one irradiation, and an ambient temperature.

3. A detection method according to any one of the preceding claims, wherein meteorological data and / or operating data, preferably meteorological and operating data, are acquired at a frequency greater than or equal to 1 Hz, preferably greater than or equal to 1 kHz.

4. A detection method according to any one of the preceding claims, wherein the structural information includes: a brand and / or an age of the photovoltaic installation and / or a material of the photovoltaic panels and / or a technology of the photovoltaic panels and / or an inclination of the photovoltaic panels and / or a topology of the photovoltaic installation and / or information relating to a replacement of photovoltaic panels and / or a number of panels per string of panels and / or a number of strings in the photovoltaic installation.

5. A detection method according to any one of the preceding claims, wherein the collection of operating data is carried out by means of at least one analog sensor.

6. A detection method according to any one of the preceding claims, wherein the collection of structural information is carried out by means of a graphical interface and / or by means of a database.

7. A detection method according to any one of the preceding claims, wherein the operating data are divided into time windows, the identification of the presence of defects being carried out for each of the time windows, at least one machine learning algorithm being pre-trained for each of said time windows.

8. A detection method according to the preceding claim, wherein the operating data are divided into time windows.

9. Computer program (1) comprising instructions which, when the program is executed by a computer, cause the computer to implement a fault detection method according to any one of the preceding claims.

Images