Photovoltaic fault diagnosis method combining kernel extreme learning machine and model-independent meta learning framework

Abstract

The invention discloses a photovoltaic fault diagnosis method combining a kernel extreme learning machine and a model-independent meta learning framework, and belongs to the technical field of photovoltaic system safety, and the method comprises the following steps: S1, constructing a photovoltaic fault diagnosis model based on the kernel extreme learning machine; s2, constructing a photovoltaic fault collaborative diagnosis model combining a kernel extreme learning machine and a model-independent meta learning framework, wherein the model is used for performing photovoltaic fault diagnosis; s3, constructing an evaluation strategy for a photovoltaic fault diagnosis result of the diagnosis model; according to the method, the fault diagnosis accuracy is guaranteed, the response speed is high, the faults of the photovoltaic system can be effectively diagnosed, and stable operation of the photovoltaic system is guaranteed.

Description

Technical Field

[0001] The present invention relates to a photovoltaic fault diagnosis method, in particular to a photovoltaic fault diagnosis method combining a kernel extreme learning machine (KELM) and a model-independent meta-learning framework (MAML), and belongs to the technical field of photovoltaic system safety. Background Art

[0002] Photovoltaic power generation has the advantages of being free from geographical restrictions, flexible in scale application, and pollution-free. In the past decade, the capacity of photovoltaic power generation has grown exponentially and it is one of the most promising renewable energy technologies. Fault diagnosis of photovoltaic arrays is an important research field in photovoltaic power generation. Its purpose is to detect and solve these faults as early as possible to ensure the stable operation of the system and the continuous supply of energy. So far, many fault diagnosis methods have been studied, but in practical applications, the characteristic parameters of different faults are not obviously different, and it is not easy to effectively obtain labels for multiple samples. Coupled with the nonlinear output characteristics of photovoltaic arrays and the complex working environment, fault diagnosis becomes extremely difficult. Therefore, it is necessary to develop more effective photovoltaic fault diagnosis methods so that faults in photovoltaic systems can be accurately and quickly identified and located.

[0003] Among many learning algorithms, extreme learning machine (ELM), as a new training framework in feedforward neural network learning, has been widely used in behavior recognition, emotion recognition and fault diagnosis. However, for photovoltaic data features, the performance of ELM is highly dependent on randomly generated weights and biases, as well as the number of hidden layer nodes. In addition, ELM is less effective when dealing with nonlinear problems. Kernel-based extreme learning machine (KELM) is a variant of extreme learning machine. It introduces kernel technology to address the limitations of the original extreme learning machine in dealing with highly nonlinear problems. KELM is mainly used for classification and regression tasks, which can effectively improve the performance when dealing with complex data sets and is more conducive to application in photovoltaic fault diagnosis problems. Although KELM has significant advantages such as simple implementation, extremely fast learning speed and less human intervention, it also has defects such as poor generalization and parameter sensitivity.

[0004] Meta-learning (ML), often referred to as "learning to learn", aims to enable machine learning algorithms to learn from past experience how to better learn new tasks. The model-agnostic meta-learning framework (MAML) aims to find a good model parameter initialization through training so that these parameters can quickly adapt to new tasks through a very small amount of gradient updates. MAML reduces its dependence on a large amount of task-specific data by optimizing the initial state of the model parameters, thereby improving the flexibility and practicality of the model. However, current meta-learning research rarely pays attention to the selection and optimization of the network, and often has problems such as high model complexity, high consumption of computing resources, and time-consuming algorithms. In addition, the meta-learning process sometimes encounters problems such as difficulty in meta-task selection and overfitting. Therefore, it is of great significance to propose a control method that can ensure the rapidity of fault diagnosis and have good generalization ability and accuracy for the stable operation of photovoltaic systems and continuous energy supply. Summary of the invention

[0005] In order to solve the above technical problems, the present invention provides a photovoltaic fault diagnosis method combining a kernel extreme learning machine with a model-independent meta-learning framework, which has a faster response speed while ensuring the accuracy of fault diagnosis, can effectively diagnose faults suffered by the photovoltaic system, and ensure the stable operation of the photovoltaic system.

[0006] In order to solve the above technical problems, the technical solution adopted by the present invention is:

[0007] A photovoltaic fault diagnosis method combining a kernel extreme learning machine with a model-independent meta-learning framework comprises the following steps:

[0008] Step S1: constructing a photovoltaic fault diagnosis model based on a nuclear extreme learning machine;

[0009] Step S2: constructing a photovoltaic fault collaborative diagnosis model combining a nuclear extreme learning machine and a model-independent meta-learning framework for photovoltaic fault diagnosis;

[0010] Step S3: construct an evaluation strategy for the photovoltaic fault diagnosis results of the diagnosis model.

[0011] A further improvement of the technical solution of the present invention is that step S1 includes the following steps:

[0012] Step S1.1: Construct a kernel matrix based on Gaussian radial basis function RBF. Among them, x i and x j is the sample in the photovoltaic fault training data set, σ is the kernel function parameter;

[0013] Step S1.2: Output the weight β by solving a regularized least squares problem. The objective function of the fault diagnosis model is:

[0014]

[0015] The output weight formula is:

[0016]

[0017] Where T is an N×L matrix containing the target values ​​of all training samples, L is the number of output nodes, C is the regularization parameter, C=1 / λ, which is used to control the complexity of the model, and I is the unit matrix.

[0018] The output function of fault diagnosis is:

[0019] f(x)=K(x,X)β,

[0020] Among them, K(x,X) represents the kernel function value of the new input x and all training samples X.

[0021] A further improvement of the technical solution of the present invention is that step S2 is implemented through a two-level optimization process of an inner loop and an outer loop, wherein the inner loop is a fast adaptation for a single task, and the outer loop is an update of initial parameters across tasks, specifically including the following steps:

[0022] Step S2.1: Inner loop process, the specific operations are:

[0023] Initialize the kernel extreme learning machine parameters, set the regularization parameter and kernel function parameter of the first subtask to C0 and σ0 respectively, then the initial kernel matrix and output weights are:

[0024]

[0025] Based on the model-independent meta-learning framework, the regularization parameters and kernel function parameters after the first round of update are expressed as follows:

[0026]

[0027] Among them, α is the inner loop learning rate, and Gaussian radial basis function RBF is selected as the kernel function K ij =K(x i ,x j ); then use the same steps to train each subtask to obtain the corresponding parameter value. The calculation process is:

[0028]

[0029] Step S2.2: External circulation process, the specific operations are as follows:

[0030] After completing the inner loop on all selected tasks, perform a meta-update, using each task T i The test set computes the loss in the inner loop and then updates the global parameters:

[0031]

[0032] Among them, β is the outer loop learning rate, C best and σ best Output for the top-level kernel extreme learning machine model.

[0033] A further improvement of the technical solution of the present invention is that after the outer loop is completed, the model is adaptively improved and the central difference method is used as the gradient calculation method:

[0034] For a univariate function f(x), the central difference formula is used to approximate its derivative at point x. The calculation formula is:

[0035]

[0036] Among them, h is a small step length;

[0037] For a multivariate function f(x), where x=(x1,x2,…,x n ), by i Use the central difference to calculate its gradient f(x), the i-th component of the gradient of function f at point x can be expressed as:

[0038]

[0039] Among them, the i-th variable x i In x i +h and x i -h, and keep other variables unchanged.

[0040] A further improvement of the technical solution of the present invention is that step S3 includes the following steps:

[0041] Step S3.1: Calculate the precision rate, Where TP is the true positive, i.e. the number of correctly predicted positive examples, and FP is the false positive, i.e. the number of negative examples that are incorrectly predicted as positive examples;

[0042] Step S3.2: Calculate the recall rate, Where FN is the false negative example, that is, the number of positive examples that are incorrectly predicted as negative examples;

[0043] Step S3.3: Calculate the F1 score,

[0044] The further improvement of the technical solution of the present invention is that the specific operation of step S3 is:

[0045] Calculate the accuracy, Accuracy = Correct / N, where Correct is the number of correctly classified samples and N is the total number of samples.

[0046] Due to the adoption of the above technical solution, the technical progress achieved by the present invention is:

[0047] Compared with traditional photovoltaic fault diagnosis methods, the present invention combines the kernel extreme learning machine with a model-independent meta-learning framework, utilizes the kernel method of the kernel extreme learning machine to process the data of each task, and uses the model-independent meta-learning framework to optimize the initialization of model parameters, making full use of the high efficiency and stability of the kernel extreme learning machine and the rapid cross-task adaptability of the model-independent meta-learning framework, so that the two can complement each other and can accurately and quickly identify and locate faults in photovoltaic systems; this method can process large amounts of data, including data collected from various components of the photovoltaic system and data on environmental factors. In addition, it can also handle various types of faults, including common and rare faults.

[0048] At the same time, in terms of model method verification, the present invention adopts multiple standards for joint evaluation, fully combining accuracy, recall rate, precision and F1 score, ensuring that the present invention will not be biased, avoiding contingency, and the evaluation is more comprehensive and accurate, which better verifies the feasibility of the method proposed in the field of photovoltaic faults and the accuracy of diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 It is the structure diagram of the extreme learning machine;

[0050] Figure 2 It is a flow chart of the algorithm combining the kernel extreme learning machine with the model-independent meta-learning framework structure;

[0051] Figure 3 This is a simulation model diagram of a photovoltaic array under normal operating conditions;

[0052] Figure 4 It is the IV curve diagram of the photovoltaic array simulation model under different operating conditions;

[0053] Figure 5 It is the PV curve diagram under different operating conditions of the photovoltaic array simulation model;

[0054] Figure 6 It is the confusion matrix diagram of fault diagnosis classification of nuclear extreme learning machine;

[0055] Figure 7 It is the confusion matrix diagram of fault diagnosis classification of kernel extreme learning machine combined with model-independent meta-learning framework;

[0056] Figure 8It is the confusion matrix diagram of fault diagnosis classification using kernel extreme learning machine combined with bee colony algorithm;

[0057] Fig. 9 This is a graph of the accuracy of fault diagnosis using various algorithms;

[0058] Fig.10 It is the F1 score (harmonic mean) result graph of fault diagnosis of various algorithms;

[0059] Fig.11 It is a conceptual diagram before experimental data processing;

[0060] Fig.12 It is a conceptual map after the experimental data is processed;

[0061] Fig.13 It is the fault diagnosis training and validation loss curves of the kernel extreme learning machine combined with the model-independent meta-learning framework;

[0062] Fig.14 It is the training and validation loss curve of convolutional neural network fault diagnosis;

[0063] Fig.15 It is the training and validation accuracy curve of MA-KELM fault diagnosis combining KELM with MAML;

[0064] Fig.16 It is a graph of the training and validation accuracy of convolutional neural network fault diagnosis. DETAILED DESCRIPTION

[0065] The present invention is further described in detail below with reference to the accompanying drawings and embodiments:

[0066] like Figures 1 to 16 As shown, a photovoltaic fault diagnosis method combining a kernel extreme learning machine with a model-independent meta-learning framework comprises the following steps:

[0067] S1: Construct a photovoltaic fault diagnosis model based on nuclear extreme learning machine. The specific process is as follows:

[0068] S1.1: The structure of the extreme learning machine is as follows Figure 1 As shown, given a photovoltaic fault training set where x i is the input fault feature vector, t i is the corresponding target output. Assuming that the hidden layer has L nodes, the extreme learning machine first randomly initializes the weights w and biases b of the hidden layer. These weights and biases will not be updated during the training process. The output matrix of the hidden layer is shown in formula (1):

[0069]

[0070] Among them, g(x) is the activation function. Common activation functions include sigmoid function, tanh function, etc. H is an N×L matrix, and each row corresponds to the output of a photovoltaic fault training sample in the hidden layer. The corresponding objective function and output weight are:

[0071]

[0072] The solution of formula (2) can be obtained by solving the linear equation system H T Hβ=H T T is obtained. In actual operation, in order to minimize the output error, the linear least squares problem is solved and the Moore-Penrose pseudo-inverse is used to calculate β. In formula (3) represents the Moore-Penrose pseudoinverse of the matrix H.

[0073] S1.2: Kernel Extreme Learning Machine introduces kernel technology to solve the limitations of the original extreme learning machine in processing nonlinear photovoltaic fault data. The construction of the kernel matrix K is based on the selected kernel function K(x i ,x j ), where x i and x j is a sample in the photovoltaic fault training data set. The kernel matrix K is expressed as shown in formula (4):

[0074] K ij =K(x i ,x j ) (4)

[0075] The advantages and disadvantages of various kernel functions are shown in Table 1. After comparison, the present invention adopts the RBF kernel function which is more suitable for photovoltaic data, such as formula (5):

[0076]

[0077] Table 1 Different kernel functions of kernel extreme learning machine

[0078]

[0079] Among them, σ is the kernel function parameter, and the corresponding objective function and output weight are:

[0080]

[0081] Where T is an N×L matrix containing the target values ​​of all fault training samples; L is the number of output nodes; C is the regularization parameter (C=1 / λ), which is used to control the complexity of the model. When C is large, 1 / C is close to zero, which means that the model focuses more on fitting the data and reduces the impact of regularization. On the contrary, when C is small, the impact of regularization is stronger, which can help the model prevent overfitting and enhance its generalization ability; I is the unit matrix.

[0082] The use of C instead of λ in the present invention may be more intuitive in practical applications of photovoltaic fault diagnosis. This representation method is helpful to adjust and optimize the strength of regularization in practical applications, especially when performing parameter tuning and cross-validation.

[0083] The output function of multi-fault diagnosis is:

[0084] f(x)=K(x,X)β,

[0085] Among them, K(x,X) represents the kernel function value of the new input x and all training samples X.

[0086] S2: Establish a photovoltaic fault collaborative diagnosis model that combines the kernel extreme learning machine with the model-independent meta-learning framework for photovoltaic fault diagnosis. Considering the compatibility of the two algorithms, the combined algorithm needs to be improved. The algorithm structure flow chart is shown in the figure. Figure 2 As shown in the figure, the left side of the flowchart is the process of updating parameters of the meta-learning processing subtask, and the right side is the process of fault diagnosis and result evaluation of the top model. The specific steps include:

[0087] For the model-independent meta-learning framework, if the global update parameter is θ, the inner and outer loop processes of meta-learning are:

[0088]

[0089]

[0090] Among them, formula (8) is the inner loop, and formula (9) is the outer loop. α is the inner loop learning rate, which is usually small. is the gradient of the loss function, L Ti The mean square error loss function is used. β is the outer loop learning rate, and this gradient is the update parameter θ in each task i 'Calculated above.

[0091] The approach of combining kernel extreme learning machines with a model-independent meta-learning framework is also implemented through a two-level optimization process: an inner loop (fast adaptation for a single task) and an outer loop (updating initial parameters across tasks):

[0092] S2.1 Inner loop process: Initialize the kernel extreme learning machine parameters, set the regularization parameter and kernel function parameter of the first subtask to C0 and σ0 respectively, then the initial kernel matrix and output weight are:

[0093]

[0094] Based on the model-independent meta-learning framework, the first round of update parameters is expressed as follows:

[0095]

[0096] Among them, α is the inner loop learning rate, and the Gaussian radial basis function is selected: K ij =K(x i ,x j ). After this update, new regularization and kernel function parameters are generated, and they serve as the initial values ​​of the next subtask in the new round of updates. Finally, the same steps are used to train each subtask to obtain the corresponding parameter values. The calculation process is:

[0097]

[0098] S2.2 Outer loop process: After completing the inner loop on all selected tasks, perform meta-update. This step updates the global initial parameters and improves the model's ability to quickly adapt to new tasks. Use each task T i The test set computes the loss in the inner loop and then updates the global parameters to minimize the sum of these losses:

[0099]

[0100] Among them, β is the outer loop learning rate, C best and σ best Output for the top-level kernel extreme learning machine model.

[0101] S2.3: Adaptive improvement measures for the model

[0102] In the meta-learning process of the hybrid diagnosis model, corresponding to the characteristics of photovoltaic fault data, the central difference method is used as the gradient calculation strategy to improve the parameter sensitivity.

[0103] For a univariate function f(x), its derivative at point x can be approximated by the central difference formula, which is:

[0104]

[0105] Here, h is a small step length, and this formula is derived based on Taylor expansion.

[0106] For a multivariate function f(x), where x=(x1,x2,…,xn ), its gradient f(x) can be obtained by i Use the central difference to calculate respectively. Specifically, the i-th component of the gradient of function f at point x can be expressed as:

[0107]

[0108] Among them, the i-th variable x i In x i +h and x i -h, and keep other variables unchanged.

[0109] S3: Establish an evaluation strategy for the photovoltaic fault diagnosis results of the diagnostic model. The process is as follows:

[0110] In order to make the evaluation of fault diagnosis results more convincing, avoid accidental factors, and fully and accurately verify the feasibility of fault diagnosis methods, a method combining multiple evaluation criteria is used for fault diagnosis results. The specific strategy is as follows: The F1 score is the harmonic mean of precision and recall, so it takes into account both the accuracy and recovery ability of the model. The value range of the F1 score is from 0 to 1, where 1 represents the best performance and 0 represents the worst performance. The relevant algorithm is as follows:

[0111] S3.1: Accuracy is a commonly used evaluation indicator in classification problems, used to measure the overall performance of the classification model. It defines the ratio of the number of correctly classified samples to the total number of samples. The implementation formula is as follows:

[0112]

[0113] S3.2: The F1 score is the harmonic mean of precision and recall, so it takes into account both the accuracy and recall of the model. The F1 score ranges from 0 to 1, where 1 represents the best performance and 0 represents the worst performance. The relevant algorithm is as follows:

[0114] Precision refers to the proportion of samples predicted by the model to be positive that are actually positive, as shown in formula (21):

[0115]

[0116] Where TP is the true positives (i.e., the number of correctly predicted positive examples), and FP is the false positives (i.e., the number of negative examples that are incorrectly predicted as positive examples). Formula (22) is the recall rate, which refers to the proportion of all actual positive classes that are correctly predicted as positive by the model, and FN is the false negatives (i.e., the number of positive examples that are incorrectly predicted as negative examples). The F1 score is the harmonic mean of precision and recall, and its calculation formula is:

[0117]

[0118] The F1 score ensures that the PV diagnostic model does not tend to optimize only one of the two by calculating the harmonic mean of precision and recall. In practical applications, sometimes improving precision will reduce recall, and vice versa. This situation is called the precision-recall trade-off. The F1 score helps ensure that the model maintains a good balance between the two, which is very important for PV application scenarios that are equally sensitive to false positives and false negatives.

[0119] S4: In order to intuitively verify the effect of the control strategy proposed in the present invention, specific examples from simulation data and experimental data will be used to illustrate it, and it will be compared with other algorithms. The process is as follows:

[0120] S4.1: Simulation data verification, the specific operations are:

[0121] S4.1.1 Model construction: This experiment uses matlab / simulink software to build a 3×3 photovoltaic array simulation model. Normal light 1000W / m2, temperature 25℃; photovoltaic cell model JKM260P-60; the normal operation model is as follows Figure 3 As shown. Taking 4 typical faults as examples, the methods for generating each fault are as follows:

[0122] A. Short circuit fault: The essence is the reduction of the number of photovoltaic cells in the branch.

[0123] B. Open circuit fault: achieved by connecting an infinite resistance in series.

[0124] C. Degradation fault: Degradation fault can be achieved by increasing the series resistance Rs or reducing the parallel resistance. In the present invention, degradation faults of different degrees are achieved by setting the resistance value of the series resistance (1Ω, 2Ω and 3Ω).

[0125] D. Partially shaded condition: Partially shaded condition is when the irradiance is less than other normal PV modules, making the PV array receive different irradiance.

[0126] S4.1.2 Algorithm model parameters and results are as follows:

[0127] S4.1.2.1 Photovoltaic model IV and PV data characteristics: The IV and PV curves under normal and four fault operating conditions are as follows: Figure 4 , 5 As shown, assuming that we set the short-circuit current to Isc, the open-circuit voltage to Voc, and the maximum power point (MPP) current and voltage to Vm and Im, then compared with the normal operating state, the change characteristics of different fault parameters are shown in Table 2:

[0128]

[0129] Table 2 Different fault characteristic data

[0130] S4.1.2.2 For the obtained photovoltaic array simulation data, if it is to be applied to the algorithm of the present invention, data preprocessing is required. The data normalization method is selected in the present invention to scale the numerical data in different feature dimensions so that it falls within a specific small interval, such as the common range between 0 and 1. The maximum minimum normalization method (MMS) calculation formula is:

[0131]

[0132] For the photovoltaic fault diagnosis model proposed in the present invention, the relevant algorithm parameter configuration is shown in Table 3:

[0133]

[0134] Table 3. Fault diagnosis algorithm parameter configuration

[0135] S4.1.2.3 In order to further highlight the superiority of the present invention for photovoltaic fault diagnosis, the model algorithm proposed in the present invention is compared with other algorithms, and the obtained classification confusion matrix is ​​as follows: Figure 6-8 The obtained accuracy and F1 score (harmonic mean) results are Fig. 9 , 10 shown.

[0136] S4.2 Experimental data verification

[0137] This experiment uses the Python platform. In order to better verify the accuracy of the model algorithm, the comparison algorithm selected is a typical convolutional neural network. The selected experimental data set contains 4 types of fault features, a total of 13×920 fault data, and the last column is the label column. The algorithm model parameter configuration is the same as the simulation parameter configuration, except that the number of training rounds is 200.

[0138] The compared convolutional neural network contains 3 convolutional layers, 2 pooling layers, 1 linear layer and 1 fully connected layer. The conceptual diagram before and after the experimental data processing is shown in the figure. Fig.11 , Fig.12 The loss curves of the two algorithms for photovoltaic fault diagnosis are shown in Fig.13 , 14 The diagnostic accuracy curve is shown in Fig.15 , 16 shown.

[0139] Figure 2 The left side of the flowchart shows the process of updating parameters of the meta-learning subtasks, and the right side shows the process of fault diagnosis and result evaluation of the top model. Each sub-core extreme learning machine processes D train The training task on , optimizes the input parameters C and σ. The optimal parameters C generated best and σ best as initial parameters for the top model.

[0140] Figure 6-8 They are the single nuclear extreme learning machine, the photovoltaic fault diagnosis method of the nuclear extreme learning machine combined with the model-independent meta-learning framework of the present invention, and the fault diagnosis classification confusion matrix of the nuclear extreme learning machine + bee colony algorithm. For the confusion matrix, the smaller the false positive (FP) and false negative (FN) values, that is, the non-diagonal element values, the less the model misjudgment, and the more obviously the diagonal value is higher than the non-diagonal value, the better the diagnosis result. It is not difficult to see from the information in the figure that the photovoltaic fault diagnosis method of the proposed nuclear extreme learning machine combined with the model-independent meta-learning framework has better classification performance.

[0141] Figure 9-10 MAELM refers to an algorithm that combines an ordinary extreme learning machine with a model-independent meta-learning framework. By comparing the five algorithms, it can be seen that the present invention has higher fault diagnosis performance in terms of accuracy and F1 score, among which the algorithm combined with the model-independent meta-learning framework has better performance than the extreme learning machine, which further illustrates the advantages of the selected kernel extreme learning machine in dealing with nonlinear problems and is more suitable for photovoltaic fault data.

[0142] Figure 13-14 The loss curve, Figure 15-16 In the accuracy curve, by comparing with the convolutional neural network, the loss of the present invention tends to be stable and the accuracy is closer to 1. At the same time, it is not difficult to see from the two accuracy curves that when about 25 rounds are performed, the accuracy curves of the training and verification of the present invention are faster than those of the convolutional neural network to reach or even exceed 0.8, and reach a higher accuracy first. This is because the present invention introduces the nuclear extreme learning machine, which makes full use of the rapidity of the nuclear extreme learning machine, so that the entire photovoltaic fault diagnosis model has a fast response speed while ensuring accuracy.

[0143] In summary, the present invention proposes a control strategy for photovoltaic fault diagnosis that combines a nuclear extreme learning machine with an improved model-independent meta-learning framework. This strategy combines the rapidity of the nuclear extreme learning machine with the accuracy of meta-learning. It has been verified that the present invention is applicable to both simulation models (4.1) and actual experimental data (4.2). At the same time, by comparing with other algorithms, the superiority of the diagnostic method of the present invention is further highlighted. It can effectively diagnose faults suffered by photovoltaic systems, accurately and quickly identify the fault types of photovoltaic systems, and ensure the stable operation of photovoltaic systems.

[0144] The above describes and illustrates the implementation mode of the present invention in detail in conjunction with the accompanying drawings, but is not limited to the above-mentioned methods. Various modifications or applications made according to the above-mentioned embodiments are within the protection scope of this technical solution.

Claims

1. A photovoltaic fault diagnosis method combining a nuclear extreme learning machine with a model-independent meta-learning framework, characterized in that: The following steps are involved: Step S1: constructing a photovoltaic fault diagnosis model based on a nuclear extreme learning machine; Step S2: constructing a photovoltaic fault collaborative diagnosis model combining a nuclear extreme learning machine and a model-independent meta-learning framework for photovoltaic fault diagnosis; Step S3: construct an evaluation strategy for the photovoltaic fault diagnosis results of the diagnosis model.

2. A photovoltaic fault diagnosis method combining a nuclear extreme learning machine and a model-independent meta-learning framework according to claim 1, characterized in that: Step S1 includes the following steps: Step S1.1: Construct a kernel matrix based on Gaussian radial basis function RBF. Among them, x i and x j is the sample in the photovoltaic fault training data set, σ is the kernel function parameter; Step S1.2: Output the weight β by solving a regularized least squares problem. The objective function of the fault diagnosis model is: The output weight formula is: Where T is an N×L matrix containing the target values ​​of all training samples, L is the number of output nodes, C is the regularization parameter, C=1 / λ, which is used to control the complexity of the model, and I is the unit matrix. The output function of multi-fault diagnosis is: f(x)=K(x,X)β, Among them, K(x,X) represents the kernel function value of the new input x and all training samples X.

3. A photovoltaic fault diagnosis method combining a nuclear extreme learning machine and a model-independent meta-learning framework according to claim 1, characterized in that: Step S2 is implemented through a two-level optimization process of an inner loop and an outer loop, wherein the inner loop is a fast adaptation for a single task, and the outer loop is an update of initial parameters across tasks, specifically including the following steps: Step S2.1: Inner loop process, the specific operations are: Initialize the kernel extreme learning machine parameters, set the regularization parameter and kernel function parameter of the first subtask to C0 and σ0 respectively, then the initial kernel matrix and output weights are: Based on the model-independent meta-learning framework, the regularization parameters and kernel function parameters after the first round of update are expressed as follows: Among them, α is the inner loop learning rate, and Gaussian radial basis function RBF is selected as the kernel function K ij =K(x i ,x j ); then use the same steps to train each subtask to obtain the corresponding parameter value. The calculation process is: Step S2.2: External circulation process, the specific operations are as follows: After completing the inner loop on all selected tasks, perform a meta-update, using each task T i The test set computes the loss in the inner loop and then updates the global parameters: Among them, β is the outer loop learning rate, C best and σ best Output for the top-level kernel extreme learning machine model.

4. A photovoltaic fault diagnosis method combining a nuclear extreme learning machine and a model-independent meta-learning framework according to claim 3, characterized in that: After the outer loop is completed, the model is adaptively improved and the central difference method is used as the gradient calculation method: For a univariate function f(x), the central difference formula is used to approximate its derivative at point x. The calculation formula is: Among them, h is a small step length; For a multivariate function f(x), where x=(x1,x2,…,x n ), by using the central difference to calculate the gradient f(x) for each component xi, the i-th component of the gradient of function f at point x can be expressed as: Among them, the i-th variable x i In x i +h and x i -h, and keep other variables unchanged.

5. The photovoltaic fault diagnosis method according to claim 1, characterized in that: Step S3 includes the following steps: Step S3.1: Calculate the precision rate, Where TP is the true positive, i.e. the number of correctly predicted positive examples, and FP is the false positive, i.e. the number of negative examples that are incorrectly predicted as positive examples; Step S3.2: Calculate the recall rate, Where FN is the false negative example, that is, the number of positive examples that are incorrectly predicted as negative examples; Step S3.3: Calculate the F1 score, 6. A photovoltaic fault diagnosis method combining a kernel extreme learning machine and a model-independent meta-learning framework according to claim 1, characterized in that: The specific operations of step S3 are: Calculate the accuracy, Accuracy = Correct / N, where Correct is the number of correctly classified samples and N is the total number of samples.

Images