Photovoltaic module snail trace detection method based on ensemble learning
Through an integrated learning method, features are extracted from the current signal of the photovoltaic system, combined with PCA and Isomap algorithms, integrated learning models are used to detect snail trajectory faults of photovoltaic modules, solving the detection difficulties in the existing technology and achieving fast and effective fault detection.
Patent Information
- Application Number
- CN202411648155.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-11-18
- Publication Date
- 2025-02-25
AI Technical Summary
The prior art is difficult to effectively detect snail trajectory failures in photovoltaic modules, especially in large-scale photovoltaic systems. Image analysis methods require specific conditions and large data sets, resulting in detection difficulties.
Using an integrated learning method, the current signal of the photovoltaic system is extracted, and the current signal related to the maximum power point is used, combined with principal component analysis (PCA) and isometric mapping (Isomap) algorithms for feature selection and dimensionality reduction. Finally, the data is input into the pre-trained integrated learning model, including support vector machine (SVM), K nearest neighbor (KNN) and decision tree (DT) models, and integrated classification is performed through most voting rules.
The rapid and effective detection of snail trajectory failure of photovoltaic modules is achieved, reducing dependence on specific conditions and large data sets, and improving the feasibility of batch detection in large-scale photovoltaic systems.
Smart Images
Figure CN119513585A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of information technology, and in particular to a method for detecting snail traces of photovoltaic components based on ensemble learning. Background Art
[0002] Some photovoltaic systems still have serious production continuity problems today, which are related to frequent failures of photovoltaic components (protection systems, boxes, inverters, photovoltaic generators, etc.). Due to the complexity of photovoltaic systems, power losses can be as high as 17.5% of the total power output.
[0003] A complete PV system includes several components that deliver power to the load or grid, such as switching converters, DC and AC fuses, all of which can fail. Nevertheless, research and actual outdoor PV system conditions show that most failures occur at the PV panel level, such as corrosion, film delamination, hot spot effects, etc.
[0004] Snail tracks in photovoltaic modules refer to a kind of defect phenomenon similar to dark lines or stripes that can be seen by the naked eye inside photovoltaic modules. These defects usually appear as dark, elongated, and sometimes curved traces that look like the traces left by a snail crawling, hence the name "snail tracks". Snail tracks indicate defects in the cells inside the photovoltaic module, reducing the absorption and conversion of light, thereby reducing power generation efficiency. In severe cases, snail tracks may increase the internal resistance of photovoltaic cells and even cause safety issues such as fire.
[0005] Faced with the above problems, many technologies use machine learning methods to detect snail tracks, but the proposed methods are mainly based on image analysis, which requires the use of drones to ensure image capture under the same training conditions under a variety of specific conditions, and requires a large number of data sets. Therefore, it becomes very difficult to perform such batch detection in large-scale photovoltaic systems. Summary of the invention
[0006] In view of the defects in the above-mentioned prior art, the present invention provides a method for detecting snail track faults in photovoltaic modules based on ensemble learning. The present invention proposes to extract features from the current signal of the photovoltaic system to identify such faults, and proposes an ensemble learning (EL) framework, using the current signal related to the maximum power point (MPP), using the feature selection and dimensionality reduction steps of the principal component analysis (PCA) and isometric mapping (Isomap) algorithms, and passing the input to the learning algorithm.
[0007] The technical solution adopted by the present invention to solve the technical problem is: a photovoltaic module snail track fault detection method based on ensemble learning, comprising the following steps:
[0008] S101: collecting current signals of power generation cycle;
[0009] S102: Decomposing and processing the collected current signal using a multi-resolution signal decomposition method;
[0010] S103: using principal component analysis (PCA) and isomap algorithms to calculate and process the original features;
[0011] S104: Input the data acquired in step S103 into a pre-trained integrated learning model, where the integrated learning model includes multiple machine learning models, and determine the health status of the photovoltaic component panel according to the output results of the integrated learning model.
[0012] In the above scheme, the ensemble learning model includes three models: support vector machine (SVM), K nearest neighbor (KNN) and decision tree (DT).
[0013] In the above scheme, the collected current signal is decomposed using multi-resolution signal decomposition (MSD), and wavelet decomposition is iteratively applied.
[0014] In the above scheme, based on the iterative wavelet decomposition of the current signal, for each photovoltaic panel, is the number of photovoltaic panels; the MSD based on the discrete wavelet decomposition DWT divides the input signal into a range of variable frequencies, wherein the decomposition of the signal is determined according to the mode or mother wavelet selected for the decomposition, and each mother wavelet is associated with or captures a specific frequency band.
[0015] In the above scheme, the ensemble learning process adopts the majority voting rule to integrate the three classifiers (K nearest neighbor KNN, decision tree DT and support vector machine SVM) using majority voting (MV), taking KNN and DT as unstable / random classifiers and SVM as a stable learner.
[0016] In the above scheme, each stable learner and unstable / random learner is regarded as a voter, and the prediction results or outputs of the three classifiers are compared to determine the final class assignment. The voting methods include weighted majority voting, relative majority voting and absolute majority voting.
[0017] The scheme of this application uses a data collection module and a natural language processing module, so the system can quickly collect and analyze a large amount of case data and legal documents, greatly improving the efficiency of information processing. This enables judges to obtain key information of the case in a short period of time and make decisions more quickly. The natural language processing module uses advanced semantic understanding technology, which can deeply analyze the content of legal documents and extract implicit semantic information. This deep understanding capability far exceeds traditional keyword matching methods and provides a more reliable foundation for subsequent legal analysis. The large model machine learning module integrates multiple professional models, including a legal document key information extraction model, an answer content key information extraction model, a bill key information extraction model, and a similar case recommendation model. This multi-model collaborative working method can analyze cases from multiple angles and provide comprehensive and in-depth legal insights. BRIEF DESCRIPTION OF THE DRAWINGS
[0018] Figure 1 This is a flow chart of a photovoltaic module snail track fault detection method based on ensemble learning in an embodiment of the present application;
[0019] Figure 2 This is an example diagram of using KNN to classify new samples in an embodiment of the present application;
[0020] Figure 3 This is an example diagram of a classification hyperplane representation using SVM in an embodiment of the present application;
[0021] Figure 4 This is a schematic diagram of the structure of a decision tree classifier in an embodiment of the present application;
[0022] Figure 5 This is a schematic diagram of the fault classification results of a photovoltaic panel with a snail track using EL based on majority voting in an embodiment of the present application. DETAILED DESCRIPTION
[0023] In order to make the purpose, features and advantages of the present invention more obvious and easy to understand, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the embodiments described below are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention. The principles and features of the present invention are described below in conjunction with the drawings. The examples are only used to explain the present invention and are not used to limit the scope of the present invention.
[0024] The terms "including" and other expressions with similar meanings in the specification or claims of the present invention and the above-mentioned drawings are intended to cover non-exclusive inclusions, such as a process, method or system, or device comprising a series of steps or units is not limited to the listed steps or units.
[0025] Embodiment: A photovoltaic module snail track fault detection method based on ensemble learning comprises the following steps:
[0026] S101: Collecting current signals of a power generation cycle from sensors of photovoltaic modules.
[0027] S102: Decompose the collected current signal using a multi-resolution signal decomposition method; wherein multi-resolution signal decomposition (MSD) iteratively applies wavelet decomposition. This stage is based on iterative wavelet decomposition of the current signal, called multi-resolution signal decomposition I i , for each photovoltaic panel i=1,…,n p , n p is the number of photovoltaic panels. The MSD based on discrete wavelet decomposition DWT divides the input signal into a variable frequency range. The decomposition of the signal directly depends on the mode or mother wavelet selected for decomposition. Each mother wavelet is associated with or captures a specific frequency band. In addition, these mother wavelets have different computational speeds and decomposition qualities. The mother wavelet is defined by the following formula (1):
[0028]
[0029] Where c and d represent the scale and offset factors, respectively, t is the timestamp of the panel current signal, and Φ is the mother wavelet;
[0030] This embodiment tests all wavelet families, and selects the mother wavelet Daubechies38 (db38) due to its computational speed and good decomposition results;
[0031] Statistical features are extracted, including mean value, power spectrum density, skewness, entropy and kurtosis, and used for photovoltaic system fault detection. For the signal S, its mean value μ is calculated using the following formula (2):
[0032]
[0033] Where n S is the number of time points, S t represents the signal at time t.
[0034] The power spectral density PSD represents the functional relationship between the power content of the signal and the frequency. The amplitude is normalized to unit frequency and calculated using formula (3):
[0035]
[0036] The skewness data is calculated according to the following formula (4), where skewness represents the asymmetry of the data about μ;
[0037]
[0038] According to the following formula (5), entropy data is obtained. Entropy is widely used in information theory to evaluate the uncertainty of signals;
[0039]
[0040] Where p(S t ) is the signal S t probability.
[0041] The maximum value and skewness of the probability distribution of the kurtosis measurement data are calculated according to the following formula (6);
[0042]
[0043] Using the above statistical features, the feature matrix The size of the matrix is n P ×((L+1)×n F ):
[0044]
[0045] where *∈{morning, noon, afternoon, and evening}, L is the number of decomposition levels, n F is the number of statistical features, and are the vectors of approximation coefficients and detail coefficients respectively;
[0046] S103: Use principal component analysis PCA and isomap algorithm to transform the original features and convert the matrix Transform to M F,* , the matrix size is reduced to n P ×U:
[0047]
[0048] Here, U is the number of features obtained by PCA or Isomap.
[0049] Principal Component Analysis (PCA) is a powerful multivariate statistical technique that identifies and extracts uncorrelated features (called latent components) from a multidimensional raw feature space. That is, the PCA algorithm constructs new features using linear combinations of the original features while maintaining maximum variance information. PCA represents two subspaces, the first subspace is named the principal subspace or "representation" subspace, and the second subspace is called the complementary subspace or residual subspace, where noise and outliers are eliminated. Let M be the original features and their covariance matrix. The first principal component spanning the principal subspace is defined by the first q dominant eigenvectors of the data covariance matrix, and the last unretained eigenvector Mq is defined as the residual subspace. Projecting the data onto the dominant eigenvectors provides the principal component scores. The percentage of variance contained in each principal component is represented by its corresponding eigenvalue. Each principal component is aligned along the direction corresponding to the maximum variance in the data, starting with the first principal component. Therefore, the principal components range from the highest variance or most dynamic variance associated with the highest eigenvalue to the lowest eigenvalue.
[0050] Isomap solves the dimensionality reduction problem in graph theory. In graph theory, the distance between two vertices on a graph corresponds to the number of edges in the shortest path connecting them. This distance is also called geodesic distance, and Isomap aims to preserve the pairwise geodesic distances between conformations in a graph, usually measured using the standard Euclidean distance. Using Isomap for dimensionality reduction or feature selection is based on the following steps:
[0051] S111: Isomap estimates the neighborhood graph by determining the neighbors of each input point in the manifold. In the input space, data point X is based on the distance d X (i, j) Between the two input points, the neighbors are drawn into a weighted graph G through K nearest neighbors, and the edge connecting the two adjacent points is a weighted edge d X (i,j).
[0052] S112: Calculate the shortest path given the neighborhood graph. For each pair of data points, calculate their shortest path lengths on the neighborhood graph (i.e., geodesic distances). These distances reflect the actual distances on the manifold, not the straight-line distances in Euclidean space.
[0053] S113: Construct low-dimensional embeddings and apply classic MDS to the graph distance matrix D G ={d G (i, j)}, construct an embedding of the data in d-dimensional Euclidean space. Using the calculated shortest paths as a similarity matrix, multidimensional scaling (MDS) is used to find a low-dimensional embedding that tries to keep the distances of these shortest paths unchanged.
[0054] S104: using ensemble learning for classification, inputting the data obtained in step S103 into a pre-trained ensemble learning model, wherein the ensemble learning model includes multiple machine learning models, and determining the health status of the photovoltaic module panel according to the output results of the ensemble learning model; if the final output result of the model is 0, it indicates that the panel is in a healthy state; if the output result is 1, it indicates that the panel has a snail track fault;
[0055] K Nearest Neighbor (KNN), KNN is one of the most commonly used classification models. For a given sample n F is the number of features, KNN will find its nearest neighbor in the classified samples according to a distance metric and calculate the Euclidean distance according to the following formula (7):
[0056]
[0057] KNN assigns a sample to the most common class among its k nearest neighbors, such as Figure 2 As shown in Figure 2. The correct choice of parameter k has a significant impact on the performance of the model. As k increases, the model can tolerate more noise because the impact of variance caused by random errors decreases, but there is a risk of missing small but important patterns in the data. Striking a balance between overfitting and underfitting is the key to choosing a suitable parameter k.
[0058] Support Vector Machine (SVM),SVM classifier searches for one or a set of optimal separating hyperplanes, such as Figure 3 , thereby maximizing the margin between classes. When the classes are not linearly separable, the SVM will map the input space to a higher-dimensional feature space through a kernel function, and then apply a linear SVM in that space. It uses a set of data points (called support vectors) that are close to the hyperplane and affect its position and direction. In soft-margin SVM, some samples may break through the margin, and the nonlinear decision boundary F can be achieved by projecting the data into a higher-dimensional space through a nonlinear function Φ(x). Although the data may not be linearly separable in its original space, the points can be mapped to the feature space F, where the hyperplane can separate them. To avoid overfitting noisy data, the variable ξ can allow some data points to be within the margin. Adjust the balance between the classification error of the training data and the maximization of the margin. The objective function of the SVM classifier can be minimized, as shown in the following calculation formula (8):
[0059]
[0060] Where C>0,w T is the weight matrix and b is the bias matrix.
[0061] Decision Tree (DT), a tree-based machine learning technique is one of the most widely used nonlinear models in many applications. DT models use two types of nodes, decision nodes and leaf nodes. Decision nodes have multiple branches and are used to make decisions, while leaf nodes are the results of these decisions. The structure of a decision tree classifier is as follows: Figure 4 As shown. There are many variants of DT, and the present invention adopts the C4.5 variant. C4.5 uses the concept of information gain of reference entropy to build a decision tree from a set of training data. Information gain calculates the change in entropy of a data set before and after splitting the data according to the value of a given feature. Further, it is the information gain IG of the split split Equal to the original entropy H minus the weighted sum of the sub-entropies H j , the weight is equal to the proportion of data samples moved to the sub-dataset given by the equation, as shown in the following calculation formula (9):
[0062]
[0063] Where |D| is the number of samples D in the original data set, |D j | is the jth sub-dataset. C4.5 tries to reduce the deviation of information gain IG split , a normalization term called intrinsic information is introduced to study the predictor variables of high degree of branching II split , as shown in the following calculation formula (10):
[0064]
[0065] C4.5 The final gain rate GR split , as shown in the following calculation formula (10):
[0066]
[0067] The feature with the highest gain ratio will be selected for splitting.
[0068] Majority voting rule, the three classifiers (KNN, DT and SVM) are integrated using majority voting (MV), KNN and DT are called unstable / random classifiers, and SVM is called a stable learner, the purpose is to obtain better detection results. In this case, a stable learner refers to a classifier that can consistently make accurate predictions, while an unstable / random learner usually exhibits less consistent prediction capabilities. The present invention utilizes the majority voting principle and regards each stable learner and unstable / random learner as a voter. In this method, the prediction results or outputs of the three classifiers are compared to determine the final class assignment. Voting methods include weighted majority voting, relative majority voting, and absolute majority voting. Assume a set of n v is a weak classifier or voter. For each sample x, each classifier returns a single output set Oi ,i=1,…,v includes a label value l j ,j=1,…,n l The voting output set is contained in the multiset Then the majority vote is returned in The label l that appears most frequently v , as shown in the following calculation formula (11):
[0069]
[0070] Where #l j Indicates the label l j frequency.
[0071] If the confidence placed in each weak classifier is different, a weighted variant of relative voting can be used. In this case, the final label is obtained as shown in the following calculation formula (12):
[0072]
[0073] in Indicates the label l j In the collection O i The frequency of occurrence in .
[0074] Snail track detection using current signal classification Figure 5 shown.
[0075] The technical means disclosed in the scheme of the present invention are not limited to the technical means disclosed in the above-mentioned implementation mode, but also include technical schemes composed of any combination of the above-mentioned technical features. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications are also regarded as the protection scope of the present invention.
Claims
1. A photovoltaic module snail track fault detection method based on ensemble learning, characterized in that: The steps include: S101: collecting current signals of power generation cycle; S102: Decomposing and processing the collected current signal using a multi-resolution signal decomposition method; S103: using principal component analysis and isometric mapping algorithm to calculate and process the original features; S104: Input the data acquired in step S103 into a pre-trained integrated learning model, where the integrated learning model includes multiple machine learning models, and determine the health status of the photovoltaic component panel according to the output results of the integrated learning model.
2. The photovoltaic module snail track fault detection method based on ensemble learning according to claim 1 is characterized in that: The ensemble learning model includes three models: support vector machine, K nearest neighbor and decision tree.
3. The photovoltaic module snail track fault detection method based on ensemble learning according to claim 1 is characterized in that: The acquired current signal is decomposed using a multi-resolution signal decomposition method, and wavelet decomposition is iteratively applied.
4. The photovoltaic module snail track fault detection method based on ensemble learning according to claim 3 is characterized in that: Based on the current signal I i The iterative wavelet decomposition of i For each photovoltaic panel i=1,…,n p The corresponding signal, n p is the number of photovoltaic panels; MSD based on discrete wavelet decomposition DWT divides the input signal into a range of variable frequencies, wherein the decomposition of the signal directly depends on the selected mode or the choice of the mother wavelet for decomposition, and each mother wavelet is associated with or captures a specific frequency band.
5. The photovoltaic module snail track fault detection method based on ensemble learning according to claim 1 is characterized in that: In the process of ensemble learning, the majority voting rule is adopted to integrate the three classifiers using majority voting, KNN and DT are regarded as unstable / random classifiers, and SVM is regarded as a stable learner.
6. The photovoltaic module snail track fault detection method based on ensemble learning according to claim 5 is characterized in that: Each stable learner and unstable / random learner is regarded as a voter, and the prediction results or outputs of the three classifiers are compared to determine the final class assignment. The voting methods include weighted majority voting, relative majority voting, and absolute majority voting.