Intelligent grid-connected photovoltaic array fault diagnosis method and device

Abstract

The invention belongs to the field of power equipment fault diagnosis, and particularly discloses an intelligent grid-connected photovoltaic array fault diagnosis method and device, and the method comprises the steps: collecting a first voltage, a second voltage, a first current and a second current of a DC output side of a photovoltaic array; generating a sample data set according to the first path of voltage, the second path of voltage, the first path of current and the first path of current at fixed time intervals; dividing the sample data set into a training data set and a test data set, and training the initial fault diagnosis model by using the training data set to obtain a trained fault diagnosis model; and identifying a test data set by using the trained fault diagnosis model to obtain a fault type. According to the invention, the accuracy and real-time performance of fault diagnosis can be improved.

Description

Technical Field

[0001] This application belongs to the field of power equipment fault diagnosis, and more specifically, relates to a fault diagnosis method and device for an intelligent grid-connected photovoltaic array. Background Art

[0002] Currently, the fault diagnosis of photovoltaic arrays in the prior art mainly relies on two methods: visual imaging and electrical characteristic parameters, but both have significant defects. Visual imaging methods (such as infrared thermal imaging, electroluminescence / photoluminescence) require special equipment to collect images, and there are problems such as poor real-time performance, strict environmental requirements, scarce samples, etc. Moreover, they may interfere with the performance of components and it is difficult to meet the continuous monitoring requirements of large power stations.

[0003] Electrical parameter methods (such as I-V curves, mathematical models) need to interrupt the system operation to obtain data, resulting in power generation losses, and the model accuracy is insufficient under complex working conditions; the circuit structure method has too high cost due to a large number of sensor deployments, and its engineering applicability is limited.

[0004] Although existing machine learning methods can improve the diagnosis efficiency, the data sources they rely on are still limited by the above technical bottlenecks and cannot fuse time-series electrical characteristics for real-time online analysis. Therefore, there is an urgent need for a fault diagnosis technology that does not require shutdown, adapts to complex environments, and can efficiently utilize operation data to solve the problems of real-time monitoring and accurate classification. Summary of the Invention

[0005] Aiming at the defects of the prior art, the purpose of this application is to provide a fault diagnosis method for an intelligent grid-connected photovoltaic array, aiming to solve the problems of low real-time performance and accuracy in the fault diagnosis of grid-connected photovoltaic power generation systems in the prior art.

[0006] To achieve the above purpose, in the first aspect, this application provides a fault diagnosis method for an intelligent grid-connected photovoltaic array, including: Collect the first voltage, second voltage, first current, and first current on the DC output side of the photovoltaic array; Generate a sample data set at fixed time intervals according to the first voltage, second voltage, first current, and first current; Divide the sample data set into a training data set and a test data set, and use the training data set to train an initial fault diagnosis model to obtain a trained fault diagnosis model; Use the trained fault diagnosis model to identify the test data set to obtain the fault type; The fault diagnosis model is composed of a time-channel separation convolution module, and the time-channel separation convolution module includes a time convolution module in the single-variable time dimension and a space convolution module in the cross-variable space dimension.

[0007] Optionally, the training method of the fault diagnosis model includes: Input the training data set into the initial model, and use the time-channel separation convolution module of the initial model to perform time convolution and spatial convolution on the training data set to obtain convolution features; Reduce the feature dimension through max pooling and global average pooling, input the pooled features into the fully connected layer, and output the classification result through softmax; Use the cross-entropy loss function to predict the loss value between the classification result and the true label, and adjust the parameters through backpropagation until the loss value is minimized to obtain a trained fault diagnosis model.

[0008] Optionally, the step of using the time-channel separation convolution module of the initial model to perform time convolution and spatial convolution on the training data set to obtain convolution features includes: Use the time convolution module to perform time convolution on the multivariate time series of the training data set to extract the time pattern features of the training data set; Perform batch normalization on the time pattern features, and use the exponential linear unit to linearize the time pattern features; Use the spatial convolution module to perform spatial convolution with cross-channel interaction on the time pattern features after feature processing to obtain convolution features.

[0009] Optionally, the time convolution module includes one-dimensional convolution kernels of size t×1, and the spatial convolution module includes one-dimensional convolution kernels of size ×d; where and are positive integers, t is the time range, and d is the number of input channels; The total number of parameters of the time convolution module and the spatial convolution module satisfies the following formula: .

[0010] Optionally, the step of performing batch normalization on the time pattern features includes: Determine the size of the batch data for normalization, and obtain the batch data mean according to the batch data size and the training data set; Obtain the batch data variance according to the data mean, batch data size, and training data set; Determine the normalized data according to the batch data variance, data mean, training data set, and a very small positive constant.

[0011] Optionally, it further includes: Determine that the normalized data satisfies a normal distribution with a mean of 0 and a variance of 1; A scaling parameter and a translation parameter are introduced to perform scaling transformation and translation transformation on the normalized data, so that the fault diagnosis model can learn and use the parameters by itself.

[0012] Optionally, the sample data set includes m samples, the time series length of each sample is N; the training label set includes m labels, and each label includes c fault categories; the fault categories include no fault, open circuit fault of the first path, short circuit fault of the second path, and shadow occlusion fault; where m, N, and c are all positive integers.

[0013] In a second aspect, the present application further provides an intelligent grid-connected photovoltaic array fault diagnosis device, including: An acquisition unit, configured to acquire the voltage of the first path, the voltage of the second path, the current of the first path, and the current of the first path on the DC output side of the photovoltaic array; A data set generation unit, configured to generate a sample data set at fixed time intervals according to the voltage of the first path, the voltage of the second path, the current of the first path, and the current of the first path; A training unit, configured to divide the sample data set into a training data set and a test data set, and use the training data set to train an initial fault diagnosis model to obtain a trained fault diagnosis model; An identification unit, configured to use the trained fault diagnosis model to identify the test data set to obtain a fault type; The fault diagnosis model is composed of a time-channel separation convolution module, and the time-channel separation convolution module includes a time convolution module in the single-variable time dimension and a spatial convolution module in the cross-variable spatial dimension.

[0014] In a third aspect, the present application provides an electronic device, including: at least one memory for storing a program; at least one processor for executing the program stored in the memory, and when the program stored in the memory is executed, the processor is used to execute the method described in the first aspect or any possible implementation manner of the first aspect.

[0015] In a fourth aspect, the present application provides a computer-readable storage medium, and the computer-readable storage medium stores a computer program, and when the computer program runs on a processor, the processor is caused to execute the method described in the first aspect or any possible implementation manner of the first aspect.

[0016] In a fifth aspect, the present application provides a computer program product, and when the computer program product runs on a processor, the processor is caused to execute the method described in the first aspect or any possible implementation manner of the first aspect.

[0017] It can be understood that the beneficial effects of the above second aspect to fifth aspect can refer to the relevant descriptions in the above first aspect, and will not be elaborated here.

[0018] Generally speaking, compared with the prior art, the above technical solutions conceived by this application have the following beneficial effects: (1) The fault diagnosis model of this application based on the Disjoint-CNN (Disjoint Convolutional Neural Network) significantly improves the accuracy and real-time performance of fault detection by regularly collecting the DC output voltage and current data of the photovoltaic array. The disjoint convolutional module combines the temporal convolution in the single-variable time dimension and the spatial convolution in the cross-variable spatial dimension, which can effectively extract the temporal features and spatial features in the data, ensure the accurate identification of the model for complex faults, and improve the accuracy of fault diagnosis; and this application uses the trained model to quickly identify the fault type to ensure the real-time performance of the diagnosis, enabling the system to take timely measures to reduce the downtime and economic losses.

[0019] (3) The design of the Disjoint-CNN of this application enables the model to have higher flexibility and adaptability when processing multi-variable signals, enhancing the diagnostic ability for potential faults in the photovoltaic system. The method of combining deep learning technology with the data analysis of photovoltaic equipment not only optimizes the fault detection process but also improves the reliability and maintenance efficiency of the entire system. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 is one of the schematic flowcharts of the intelligent grid-connected photovoltaic array fault diagnosis method provided by the embodiment of this application; Figure 2 is the schematic diagram of the training process of the fault diagnosis model provided by the embodiment of this application; Figure 3 is the schematic diagram of one-dimensional convolution of single-variable and multi-variable time series data provided by the embodiment of this application; Figure 4 is the comparison schematic diagram between the 1D convolution kernel and the 1+1D convolution module provided by the embodiment of this application; Figure 5 is the schematic diagram of the activation function provided by the embodiment of this application; Figure 6 is the schematic diagram of the structure of the photovoltaic array provided by the embodiment of this application; Figure 7 is the schematic diagram of the confusion matrix of photovoltaic array fault identification based on Disjoint-CNN provided by the embodiment of this application; Figure 8 is the comparison schematic diagram of the confusion matrices of the comparison model algorithms FCN, MC-DCNN, ResNet and MLSTM-FCN provided by the embodiment of this application; Figure 9 is one of the comparison schematic diagrams of the accuracy rates of each algorithm provided by the embodiment of this application; Figure 10 It is the second schematic diagram of the accuracy comparison of each algorithm provided by the embodiments of the present application; Figure 11 It is the structural schematic diagram of the intelligent grid-connected photovoltaic array fault diagnosis device provided by the embodiments of the present application; Figure 12 It is the structural schematic diagram of the electronic device provided by the embodiments of the present application. Detailed implementation manners

[0021] In order to make the objectives, technical solutions and advantages of the present application clearer, the present application will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application, and are not used to limit the present application.

[0022] The term "and / or" in this article is an association relationship describing associated objects, indicating that three relationships may exist. For example, A and / or B may represent: A exists alone, A and B exist simultaneously, and B exists alone. The symbol " / " in this article represents an "or" relationship between associated objects. For example, A / B represents A or B.

[0023] The terms "first" and "second" in the description and claims of this application are used to distinguish different objects, rather than to describe a specific order of objects. For example, the first response message and the second response message are used to distinguish different response messages, rather than to describe the specific order of the response messages.

[0024] In the embodiments of the present application, words such as "exemplary" or "for example" are used to represent examples, illustrations or explanations. Any embodiment or design solution described as "exemplary" or "for example" in the embodiments of the present application should not be construed as being more preferred or having more advantages than other embodiments or design solutions. Rather, the use of words such as "exemplary" or "for example" is intended to present relevant concepts in a specific manner.

[0025] In the description of the embodiments of the present application, unless otherwise specified, the meaning of "a plurality" refers to two or more. For example, a plurality of processing units refers to two or more processing units; a plurality of elements refers to two or more elements.

[0026] The embodiments of the present application will be described below with reference to the accompanying drawings in the embodiments of the present application.

[0027] The intelligent grid-connected photovoltaic array fault diagnosis method provided by the present application includes: Collect the first voltage, the second voltage, the first current and the first current on the DC output side of the photovoltaic array; Generate a sample data set based on the first-channel voltage, second-channel voltage, first-channel current, and first-channel current at fixed time intervals; Divide the sample data set into a training data set and a test data set, and use the training data set to train an initial fault diagnosis model to obtain a trained fault diagnosis model; Use the trained fault diagnosis model to identify the test data set to obtain the fault type; The fault diagnosis model is composed of a time-channel separation convolution module, and the time-channel separation convolution module includes a time convolution module in the single-variable time dimension and a space convolution module in the cross-variable space dimension.

[0028] Refer to Figure 1 The intelligent grid-connected photovoltaic array fault diagnosis method provided by the embodiments of the present application specifically includes the following steps: Start; The DC output side of the photovoltaic array under different working conditions U A 、 U B 、 I A 、 I B Data acquisition; Generate a sample set every 60s, and divide it into a training data set and a test sample data set; Input the training data set and perform training on the Disjoint-CNN photovoltaic array fault diagnosis model; The test sample data set is input into the Disjoint-CNN photovoltaic array fault diagnosis model for fault type identification.

[0029] Specifically, the present application first collects electrical parameters on the DC output side of the photovoltaic array to obtain the first-channel voltage, second-channel voltage, and first-channel current and second-channel current, that is, the voltages of the A and B channels on the DC output side U A 、 U B and the branch currents I A 、 I B The accurate measurement of these parameters provides a basis for subsequent analysis, ensuring the reliability and effectiveness of the data. High-frequency data sampling can capture the instantaneous state of the system during operation, record the electrical characteristics under different environmental and load conditions, and thus provide comprehensive data support for fault diagnosis.

[0030] Next, based on the first-channel voltage, second-channel voltage, and first- and second-channel currents collected, the system will generate a sample data set at fixed time intervals (preferably 60 s in this embodiment).

[0031] In this embodiment, a Disjoint-CNN network model is used to diagnose faults in the photovoltaic array, and the voltages of the A and B channels on the DC output side of the photovoltaic array are collected U A 、 U B and the branch currents I A 、 I B to form a four-channel multivariate time series data. One time point is collected every 1 s, and every 60 s is packed into a group of samples.

[0032] During the generation of the sample data set, it is necessary to ensure the integrity and consistency of the data, handle missing values and outliers to improve the data quality. At the same time, necessary preprocessing steps can also be completed at this stage, which will help improve the efficiency and accuracy of subsequent model training. The complete sample data set provides rich information for the training of the fault diagnosis model and helps to establish a more accurate fault identification system.

[0033] Optionally, the sample data set includes m samples, and the time series length of each sample is N; the training label set includes m labels, and each label includes c fault categories. The fault categories include no fault, open circuit fault of the first channel (A channel), short circuit fault of the second channel (B channel), and shadow occlusion fault.

[0034] Specifically, this application gives a time series data set with m samples X , and each sample in the data set is d a time series of dimension N , X is the length of the time series , as shown in Equation (1) is a N × d set of real numbers.

[0035]

[0036]

[0037] (1) The set of labels corresponding to the input time series data set X is Y , YAny one of the elements is represented by as shown in Equation (2), c where

[0039]

[0040] is the total number of categories. The goal of multivariate time series classification is to train a neural network classifier to map the set X to the set Y . The probability of the label C is calculated using the classifier as shown in Equation (3).

[0042] is the probability vector mapped by the specific input vector where is the specific input vector mapped, C is the mapped category. f is the function of the neural network. is all the parameters used in the model, including all weights and biases.

[0044] In the embodiments of the present application, the generated sample data set will subsequently be divided into a training data set and a test data set. Usually, the training data set accounts for 70% to 80% of the total data set, and the remaining part is used for model verification. This division method ensures the independence of the data for the model in the training and test phases and can effectively evaluate the generalization ability of the model.

[0045] During the training process, the initial fault diagnosis model is iteratively trained using the training data set. By continuously adjusting the model parameters, it can identify the patterns in the input data. This process involves selecting suitable loss functions and optimization algorithms to improve the accuracy and efficiency of the model in fault type recognition. Finally, through an appropriate evaluation mechanism, it is ensured that the trained fault diagnosis model has good prediction performance.

[0046] Finally, after the training is completed, the trained fault diagnosis model is used to identify the test data set to determine the fault type of the system. The key in this stage is to input the test data into the model and obtain the corresponding fault diagnosis results through the inference ability of the model. The use of the test data set can effectively verify the ability of the model to handle new data and ensure the efficiency and correctness of fault identification.

[0047] The proposed fault diagnosis model is based on the Disjoint-CNN module, which diagnoses different types of faults by dynamically extracting temporal and spatial features. The temporal convolution module can capture the changing patterns of time series, while the spatial convolution module focuses on the relationships between different electrical parameters. The application of this model not only improves the accuracy of fault identification but also can quickly respond to system anomalies and take maintenance measures in a timely manner, thereby optimizing the operation and maintenance efficiency of the photovoltaic array.

[0048] Optionally, the training method of the fault diagnosis model includes: Input the training data set into the initial model, and use the temporal and spatial convolution of the Disjoint-CNN module of the initial model to perform temporal convolution and spatial convolution on the training data set to obtain convolution features; Reduce the feature dimension through max pooling and global average pooling, input the pooled features into the fully connected layer, and output the classification result through softmax; Use the cross-entropy loss function to predict the loss value between the classification result and the true label, and adjust the parameters through backpropagation until the loss value is minimized to obtain the trained fault diagnosis model.

[0049] Specifically, referring to Figure 2 , the input of the Disjoint-CNN network of this application is a multivariate time series, where l represents the length of the input sequence and d represents the number of input channels. As the first step of the "1+1D" convolution module, apply a temporal convolution kernel to the input multivariate time series. In this step, the model extracts the temporal patterns that appear in the input signal, and then apply batch normalization (BN) and exponential linear units (ELU) activation functions. Feed the output of the temporal convolution into the spatial convolution kernel to finally obtain the output of the "1+1D" convolution module.

[0050] First, the processes of temporal convolution and spatial convolution will be described in detail: This application inputs the constructed training data set into the initial fault diagnosis model. The core of this model is the Disjoint-CNN module, which is designed to consider feature extraction in both the temporal and spatial dimensions simultaneously, which is very important for processing time series data (such as the voltage and current of a photovoltaic array). In this step, the temporal convolution module will first perform a convolution operation on the input time series data to capture the temporal context information and dynamic changes, helping the model identify patterns and trends.

[0051] Subsequently, the features completed by the temporal convolution module will be transmitted to the spatial convolution module, which performs cross-variable convolution analysis to extract the spatial features between different electrical parameters. Through this separated convolution design, the model can better understand the complex relationships between multiple inputs. The convolution features jointly produced by this process will serve as the basis for subsequent feature processing and provide rich information for the final classification task.

[0052] It should be noted that the CNN architecture is very suitable for learning complex patterns in image or time series data. These types of data usually have a hierarchical structure. CNN can learn local non-linear patterns through convolutional kernels and activation functions, and represent higher-level patterns as combinations of lower-level patterns. In addition, by applying pooling layers, CNN can also create a coarser intermediate feature representation, thereby learning position-invariant features in the input data. Therefore, this application selects CNN as the underlying architecture model for the multivariate time series classification algorithm, carefully studies the one-dimensional (1D) filters for univariate and multivariate time series inputs, and then uses this new 1+1D convolution module to improve the classification performance.

[0053] (1) 1D Convolution: Refer to Figure 3 , Figure 3 represents the one-dimensional convolution of univariate and multivariate time series data. As shown in Figure 3 (a), when the input is a univariate time series, the size of the convolutional kernel is one-dimensional . Using this type of convolution is to apply the univariate CNN model to multivariate time series input data, treating it as multiple independent univariate sequences. Since this convolution only extracts temporal information, it is called temporal convolution.

[0054] However, when the input is a multivariate time series, the size of the convolutional kernel is , where d is the input dimension, t is the length of the convolutional kernel set by the user. Simply put, when defining a one-dimensional convolutional kernel for a two-dimensional input, only the length of the convolutional kernel needs to be set, and the width of the convolutional kernel is always set to the number of input dimensions. As shown in Figure 3 (b), for a 10-dimensional time series, the size of the one-dimensional convolutional kernel is .

[0055] (2) 1+1D Convolution Module: The comparison between the 1D convolutional kernel and the 1+1D convolution module is shown in Figure 4 . This figure shows the simplified setting of inputting l × d data into the convolutional layer and generating l × 1 output, where l represents the length of the input sequence, d represents the number of input channels. As shown inFigure 4 As shown in (a), the application of 1D convolution t × d convolution kernel, where t represents the time range ( t < l ). The 1+1D convolution module is as shown in Figure 4 (b), replacing the 1D convolution kernel with a 1+1D block composed of disjoint temporal and spatial convolution kernels, setting the number of spatial convolutions to M i , which is the dimension of the intermediate subspace for projecting the input signal between the temporal and spatial convolution kernels. Replace N i one-dimensional convolution kernels of size t × d with a 1+1D convolution module composed of M i temporal convolution kernels of size t ×1 and N i spatial convolution kernels of size M i ×d .

[0056] To make the number of parameters of the 1+1D convolution consistent with that of the typical 1D convolution, Equation (4) needs to hold.

[0057] (4) Therefore,[[]] M i is calculated as shown in Equation (5).

[0058] (5) 1+1D is closely related to the pseudo-3D network (P3D) and spatio-temporal convolution, which decompose 3D convolution into 2D and 1D convolutions respectively and have achieved good results in 3D data of computer vision. In addition, the 1+1D convolution module emphasizes the interaction between input channels, decomposing 1D convolution into two non-mixing consecutive operations, namely 1D temporal convolution and 1D spatial convolution for each channel, and learning the interaction between channels through the features extracted from the temporal convolution.

[0059] Compared with 1D convolution, the 1+1D decomposition used in this application has two advantages. First, an additional non-linear activation function is used between the temporal convolution and the spatial convolution in each 1+1D convolution module, providing additional non-linear representations for the model. Using an additional non-linear activation function doubles the number of non-linear functions in the model, enabling the model to extract more complex functions with fewer parameters while keeping the number of parameters roughly the same as that of a typical one-dimensional convolution operation. Second, decomposing the one-dimensional convolution into separate temporal and spatial components makes the optimization process of the deep neural network model easier. In summary, compared with the typical 1D convolution kernel in which the temporal and spatial convolution kernels are intertwined, the 1+1D convolution module is easier to optimize, and lower losses in the test set and training set are proven in practice.

[0060] Furthermore, the method for obtaining convolution features further includes: Performing temporal convolution on the multivariate time series of the training data set using the temporal convolution module to extract the temporal pattern features of the training data set; Performing batch normalization on the temporal pattern features and linearizing the temporal pattern features using the exponential linear unit; Performing spatial convolution with cross-channel interaction on the temporally patterned features after feature processing using the spatial convolution module to obtain convolution features.

[0061] The following provides a detailed description of the batch normalization of data: A batch is a set randomly selected from a training sample set, and batch normalization is to perform normalization in units of batches, and the mathematical expression is shown in Equation (6).

[0062]

[0063]

[0064] (6) In the formula, m represents the size of the batch, is the mean of the batch data, is the variance of the batch data, is a very small positive constant.

[0066] Moreover, batch normalization is performed on the intermediate output values of the network before the activation function. After transforming the data into a normal distribution with a mean of 0 and a variance of 1, the BN layer also performs scaling and translation transformations on the normalized data, allowing the neural network to learn and modify the scaling parameter g and the translation parameter b , and the mathematical expression is shown in Equation (7).

[0067] (7) After BN processing, the input value distributions of each layer in the deep neural network during training are similar. By keeping the input of the hidden layer closer to the normal distribution during training to promote optimization, it is easier to learn the patterns between input data. Therefore, a larger learning rate can be used to improve the convergence speed of the neural network model. Different parameter initialization methods can also be tried to speed up the training of the network, alleviating the problem of gradient dispersion in deep networks to a certain extent. In addition, the BN layer introduces the mean and variance, which are different for each batch, as parameters into the network, equivalent to randomly adding noise to the training process, having a certain regularization effect and preventing the model from overfitting.

[0068] The following provides a detailed description of feature pooling and classification: After obtaining the convolutional features, the system will apply max pooling and global average pooling to reduce the dimension of the features. The role of max pooling is to select the maximum value in the feature map, effectively suppressing noise and retaining the most prominent features; while global average pooling will calculate the average of each channel in the feature map to summarize the overall features. This combined pooling strategy helps to reduce the number of parameters, lower the computational complexity of the model, and also avoid the risk of overfitting.

[0069] The pooled features will be input into the fully connected layer for further non-linear transformation, and the final decision interface will be constructed through a multi-layer perceptron (MLP) architecture. The fully connected layer will integrate the convolutional features from a high level and achieve probability output through the softmax function, finally giving the prediction probabilities of each fault category. The design of this step enables the model to select from multiple classes to a certain extent, making the fault recognition result more accurate.

[0070] It should be noted that the ELU activation function in the embodiments of this application combines the features of the sigmoid activation function and the ReLU activation function. It has soft saturation when x < 0, enhancing the robustness of ELU to input changes and noise; the linear part when x > 0 makes ELU not troubled by gradient explosion or disappearance. Different from ReLU, ELU does not have the problem of neuron death, that is, when abnormal input occurs, a large gradient will be generated during the backpropagation process, resulting in neuron death and gradient disappearance, because the gradient of ELU is non-zero for all negative values. The ELU activation function has been proven to be superior to ReLU and its variants, such as Leaky-ReLU and Parameterized-ReLU. Compared with ReLU and its variants, since the output mean of ELU is close to zero, the convergence speed is faster. Using the ELU activation function can shorten the training time and improve the accuracy in the neural network.

[0071] Before applying the activation function, batch normalization is performed on the intermediate temporal and spatial convolution outputs to make these representations closer to a normal distribution and simplify gradient optimization. Finally, before the fully connected layer, max pooling (MaxPool) and global average pooling (GAP) are applied to the output of the last ELU activation function, resulting in a more translation-invariant model.

[0072] Finally, the process of loss calculation and backpropagation is described in detail.

[0073] The cross-entropy loss function is used to evaluate the difference between the model classification results and the true labels. The cross-entropy loss function is particularly suitable for multi-class classification tasks, can quantify the inconsistency between the true labels and the prediction results, and provides important information about the model's performance. By calculating the loss value between the probability distribution output by the model and the true label distribution, the performance of the current model in the classification task can be clearly identified, which is crucial for further optimizing the model.

[0074] After that, the backpropagation algorithm is used to adjust the parameters of the model. By minimizing the loss function, the model adjusts the weights of each layer according to the gradient information, gradually reducing the loss value. In each iteration, the model updates the parameters until the preset convergence condition is reached. This continuous learning process ensures that the model can fully learn the features in the training data, thus forming a trained fault diagnosis model that can efficiently identify fault types and be applied to actual scenarios.

[0075] After multiple iterations and parameter optimizations, the finally obtained trained fault diagnosis model will have strong generalization ability and can effectively identify faults in newly input data.

[0076] The following describes the present application in detail in combination with specific implementation cases: The training set samples of the multivariate time series of the output voltage and current of each branch of the grid-connected photovoltaic array during normal operation are shown in Table 1, the training set samples during an open circuit fault in Circuit A are shown in Table 2, the training set samples during a short circuit fault in Circuit B are shown in Table 3, and the training set samples during shadow occlusion in Circuit B are shown in Table 4.

[0077] Table 1: Example of training set samples of the multivariate time series output by the grid-connected photovoltaic array during normal operation

[0078] Table 2: Example of training set samples of the multivariate time series output by the grid-connected photovoltaic array during an open circuit fault

[0079] Table 3: Sample Example of Multivariate Time Series Training Set Output by Grid-Connected Photovoltaic Array during Short-Circuit Fault

[0080] Table 4: Sample Example of Multivariate Time Series Training Set Output by Grid-Connected Photovoltaic Array during Shadow Occlusion Fault

[0081] In this embodiment, a total of 4463 groups of multivariate time series sample data are collected in the grid-connected photovoltaic array fault simulation test. The 4463 groups of data are randomly divided into 3259 groups of training sets and 1204 groups of test sets. The number of samples in normal operation is much larger than the number of samples with faults, which conforms to the probability of fault occurrence in real engineering. In addition, it is necessary to mark the corresponding labels of the grid-connected photovoltaic array operating states on each sample. For the convenience of the Disjoint-CNN network to classify different working conditions of the grid-connected photovoltaic array, the correspondence between the labels of the sample set and the operating states of the photovoltaic array is shown in Table 5.

[0082] Table 5: Correspondence Table between Sample Set Labels and Actual Operating States of Photovoltaic Array

[0083] The training set is input into the Disjoint-CNN photovoltaic array fault diagnosis model for training. The training process of the training set refers to Figure 2 the overall framework of the Disjoint-CNN network model. Since the input to the Disjoint-CNN network model is four-channel multivariate time series data, that is, the branch voltages and branch currents of the two branches A and B of the grid-connected photovoltaic array, so Figure 2 d in

[0084] is equal to 4, and l is the length of the multivariate time series. The Disjoint-CNN photovoltaic array fault diagnosis model uses a total of 4 "1+1D" convolutional layers. The number of time and space convolutional kernels of 1+1D is set to 64, and the lengths of the time convolutional kernels in the 4 "1+1D" convolutional modules are set to 8, 5, 5, and 3. The width of the first spatial convolutional kernel is equal to the dimension of the input time series. For the remaining layers, the width of the spatial convolution depends on the number of output channels of the previous layer, that is, 64. The parameters of the time convolution and spatial convolution in the "1+1D" module of Disjoint-CNN are shown in Table 6.

[0085] Table 6: Parameters of "1+1D" Convolutional Kernel Module of Disjoint-CNN Network

[0086] The "1+1D" convolution module first performs temporal convolution on the input multivariate time series. In this step, the model extracts the temporal patterns of the input signal. After that, batch normalization (BN) and the ELU activation function are applied. The output of the temporal convolution is fed into the spatial convolution kernel, and finally, the output of the "1+1D" convolution module is obtained. After BN processing, the input values of each layer in the Disjoint-CNN network during the training process have a similar normal distribution, making it easier to learn the patterns between the input multivariate time series data and achieving the effect of simplifying gradient optimization. The ELU activation function combines the characteristics of the sigmoid activation function and the ReLU activation function, without the problems of gradient explosion and gradient vanishing and can shorten the training time. The output of each layer of the convolutional network's "1+1D" convolution module also needs to go through another batch normalization operation and the ELU activation function. Before inputting into the fully connected layer, max pooling (MaxPool) and global average pooling (GAP) are applied to the output of the last ELU activation function to compress the length of the time series, leaving only the time series segments with the most multivariate time series characteristics, resulting in a model with stronger translational invariance. Finally, the Softmax function is applied to obtain the classification prediction results, which are compared with the labels input in the training set to update the parameters of the network model and train the model.

[0087] The test set samples of the multivariate time series of the output voltage and current of each branch of the grid-connected photovoltaic array under normal operation are shown in Table 7. The test set samples when an open circuit fault occurs in Branch A are shown in Table 8. The test set samples when a short circuit fault occurs in Branch B are shown in Table 9. The test set samples when there is a shadow occlusion in Branch B are shown in Table 10. The test set is input into the trained Disjoint-CNN model suitable for photovoltaic array fault diagnosis, and finally, the fault classification results of the grid-connected photovoltaic array are obtained.

[0088] Table 7: Example of test set samples of the multivariate time series output by the grid-connected photovoltaic array under normal operation

[0089] Table 8: Example of test set samples of the multivariate time series output by the grid-connected photovoltaic array during an open circuit fault

[0090] Table 9: Example of test set samples of the multivariate time series output by the grid-connected photovoltaic array during a short circuit fault

[0091] Table 10: Example of test set samples of the multivariate time series output by the grid-connected photovoltaic array during a shadow occlusion fault

[0092] To comparatively analyze the effectiveness and accuracy of the Disjoint-CNN model used in this technology, taking the same 3259 groups of training sets and 1204 groups of test sets as data samples, the following four mainstream classification algorithms are respectively used for multivariate time series classification work: Multi-channel Deep Convolutional Neural (MC-DCNN), which only uses temporal convolution to extract features from the input sequence; Fully Convolutional Neural network (FCN); Residual Network (ResNet); Multivariate long-short term Memory FCN (MLSTM-FCN), as a comparison model for the Disjoint-CNN network.

[0093] The 1204 groups of test set samples are composed of 1049 normally operating samples, 55 open circuit fault samples, 58 short circuit fault samples, and 42 shadow occlusion fault samples. The confusion matrix of photovoltaic array fault identification based on Disjoint-CNN is as Figure 7 shown. Among the 1204 groups of test set samples, only 6 samples are misclassified, and the overall classification accuracy is as high as 99.5%.

[0094] It can be seen from the confusion matrix in the figure that the photovoltaic array fault diagnosis model based on Disjoint-CNN does not classify the abnormal state of the photovoltaic array as the normal operating state, and the accuracy of abnormal diagnosis is 100%, which fully plays the role of fault warning. However, there is still a small part of confusion in the distinction between short circuit faults and shadow occlusion.

[0095] The confusion matrices of the four mainstream comparison model algorithms FCN, MC-DCNN, ResNet, and MLSTM-FCN are as Figure 8 shown.

[0096] From Figure 8 the confusion matrices of each algorithm, it can be seen that other improved algorithms based on CNN have the situation of classifying the abnormal state of the photovoltaic array as the normal operating state, and the diagnosis of the abnormal state of the photovoltaic array is not accurate enough to fully play the role of warning. Generally, each comparison model is extremely likely to identify the shadow occlusion situation as normal, and the recognition accuracy of shadow occlusion faults has not reached the passing line, which will be extremely unfavorable to the cleaning work on the surface of photovoltaic modules. The FCN algorithm and the MC-DCNN algorithm are prone to misidentifying short circuit faults as normal operating states, posing a major hidden danger to the safe operation of the photovoltaic array.

[0097] The accuracy rates of Disjoint-CNN and the above methods in the fault diagnosis of photovoltaic arrays are compared, and the results are shown in Table 11 Figure 9 and Figure 10 as shown below.

[0098] Table 11 Fault diagnosis accuracy rates of grid-connected photovoltaic arrays under different algorithms

[0099] As can be seen from Table 11, compared with other improved algorithms based on CNN such as FCN, MC-DCNN, ResNet, and MLSTM-FCN, the Disjoint-CNN algorithm has the highest accuracy rate in the fault diagnosis of grid-connected photovoltaic arrays. Especially in the case of partial shadow occlusion, the diagnostic accuracy rate is significantly higher than that of other algorithms, and it is more suitable for multi-variable time series classification tasks. This shows that as an improved algorithm of CNN, Disjoint-CNN decomposes the original one-dimensional convolution kernel into disjoint time components and spatial components, which can significantly improve the accuracy with almost no additional computational cost, has a better classification effect, can distinguish the subtle differences between short-circuit faults and partial shadow occlusion, and shows better applicability and accuracy in the fault diagnosis of grid-connected photovoltaic arrays.

[0100] Referring to Figure 11 , this application also provides an intelligent grid-connected photovoltaic array fault diagnosis device, including: A collection unit 111 for collecting the first voltage, the second voltage, the first current, and the first current on the DC output side of the photovoltaic array; A data set generation unit 112 for generating a sample data set at fixed time intervals according to the first voltage, the second voltage, the first current, and the first current; A training unit 113 for dividing the sample data set into a training data set and a test data set, and training an initial fault diagnosis model using the training data set to obtain a trained fault diagnosis model; An identification unit 114 for identifying the test data set using the trained fault diagnosis model to obtain the fault type; The fault diagnosis model is composed of a time-channel separation convolution module, and the time-channel separation convolution module includes a time convolution module in the single-variable time dimension and a spatial convolution module in the cross-variable spatial dimension.

[0101] Optionally, the training method of the fault diagnosis model includes: Inputting the training data set into the initial model, and performing time convolution and spatial convolution on the training data set using the time-channel separation convolution module of the initial model to obtain convolution features; Reduce the feature dimension through max pooling and global average pooling, input the pooled features into the fully connected layer, and output the classification result through softmax; Use the cross-entropy loss function to predict the loss value between the classification result and the true label, and adjust the parameters through backpropagation until the loss value is minimized to obtain a trained fault diagnosis model.

[0102] Optionally, using the time-channel separation convolution module of the initial model to perform time convolution and spatial convolution on the training dataset to obtain convolution features, including: Use the time convolution module to perform time convolution on the multivariate time series of the training dataset to extract the time pattern features of the training dataset; Perform batch normalization on the time pattern features, and use the exponential linear unit to linearize the time pattern features; Use the spatial convolution module to perform spatial convolution with cross-channel interaction on the time pattern features after feature processing to obtain convolution features.

[0103] Optionally, the time convolution module includes one-dimensional convolution kernels of size t×1, and the spatial convolution module includes one-dimensional convolution kernels of size ×d; where and are positive integers, t is the time range, and d is the number of input channels; The total number of parameters of the time convolution module and the spatial convolution module satisfies the following formula: .

[0104] Optionally, performing batch normalization on the time pattern features includes: Determine the batch data size for normalization, and obtain the batch data mean according to the batch data size and the training dataset; Obtain the batch data variance according to the data mean, batch data size, and training dataset; Determine the normalized data according to the batch data variance, data mean, training dataset, and a very small positive constant.

[0105] Optionally, it further includes: Determine that the normalized data follows a normal distribution with a mean of 0 and a variance of 1; Introduce scaling parameters and translation parameters to perform scaling transformation and translation transformation on the normalized data, so that the fault diagnosis model can learn and use the parameters by itself.

[0106] Optionally, the sample data set includes m samples, and the time series length of each sample is N; the training label set includes m labels, and each label includes c fault categories; the fault categories include no fault, open circuit fault of the first path, short circuit fault of the second path, and shadow occlusion fault; where m, N, and c are all positive integers. It can be understood that for the detailed function implementation of each of the above units / modules, reference can be made to the introduction in the foregoing method embodiments, and details are not described herein again.

[0107] It should be understood that the above device is used to execute the method in the above embodiment. For the corresponding program modules in the device, their implementation principles and technical effects are similar to those described in the above method. The working process of the device can refer to the corresponding process in the above method, and details are not described herein again.

[0108] Referring to Figure 12 , based on the method in the above embodiment, an embodiment of the present application provides an electronic device, which may include: a processor (Processor) 121, a communication interface (Communications Interface) 122, a memory (Memory) 123, and a communication bus 124. Among them, the processor 121, the communication interface 122, and the memory 123 complete mutual communication through the communication bus 124. The processor 121 can call the logical instructions in the memory 123 to execute the method in the above embodiment.

[0109] In addition, when the logical instructions in the above memory 123 are implemented in the form of software functional units and sold or used as an independent product, they can be stored in a computer-readable storage medium. Based on such an understanding, the technical solution of the present application, in essence, or the part that contributes to the prior art, or a part of this technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present application.

[0110] Based on the method in the above embodiment, an embodiment of the present application provides a computer-readable storage medium. The computer-readable storage medium stores a computer program. When the computer program runs on a processor, the processor is caused to execute the method in the above embodiment.

[0111] Based on the method in the above embodiment, an embodiment of the present application provides a computer program product. When the computer program product runs on a processor, the processor is caused to execute the method in the above embodiment.

[0112] It can be understood that the processor in the embodiments of the present application may be a central processing unit (CPU), or may also be other general-purpose processors, digital signal processors (DSPs), application specific integrated circuits (ASICs), field programmable gate arrays (FPGAs), or other programmable logic devices, transistor logic devices, hardware components, or any combination thereof. The general-purpose processor may be a microprocessor or any conventional processor.

[0113] The method steps in the embodiments of the present application may be implemented in a hardware manner or by a processor executing software instructions. The software instructions may be composed of corresponding software modules, and the software modules may be stored in a random access memory (RAM), flash memory, read-only memory (ROM), programmable ROM (PROM), erasable PROM (EPROM), electrically erasable PROM (EEPROM), registers, hard disks, removable hard disks, CD-ROMs, or any other form of storage medium well known in the art. An exemplary storage medium is coupled to the processor so that the processor can read information from the storage medium and write information to the storage medium. Of course, the storage medium may also be a component of the processor. The processor and the storage medium may be located in the ASIC.

[0114] In the above embodiments, it can be implemented in whole or in part by software, hardware, firmware, or any combination thereof. When implemented using software, it can be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, the processes or functions described in the embodiments of the present application are generated in whole or in part. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable devices. The computer instructions can be stored in a computer-readable storage medium or transmitted through the computer-readable storage medium. The computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center in a wired manner (such as coaxial cable, optical fiber, digital subscriber line (DSL)) or a wireless manner (such as infrared, wireless, microwave, etc.). The computer-readable storage medium can be any available medium that can be accessed by a computer or a data storage device such as a server or a data center that includes one or more integrated available media. The available medium can be a magnetic medium (such as a floppy disk, a hard disk, a magnetic tape), an optical medium (such as a DVD), or a semiconductor medium (such as a solid state disk (SSD)), etc.

[0115] It can be understood that the various numerical numbers involved in the embodiments of the present application are only for the convenience of description and are not used to limit the scope of the embodiments of the present application.

[0116] Those skilled in the art can easily understand that the above are only the preferred embodiments of the present application and are not intended to limit the present application. Any modifications, equivalent replacements, and improvements made within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. A method for fault diagnosis of an intelligent grid-connected photovoltaic array, characterized in that: include: Collecting a first voltage, a second voltage, a first current and a second current at a DC output side of a photovoltaic array; Generate a sample data set according to the first voltage, the second voltage, the first current and the first current at fixed time intervals; The sample data set is divided into a training data set and a test data set, and the initial fault diagnosis model is trained using the training data set to obtain a trained fault diagnosis model; Using the trained fault diagnosis model to identify the test data set to obtain the fault type; The fault diagnosis model is constructed based on a time channel separation convolution module, and the time channel separation convolution module includes a time convolution module of a univariate time dimension and a space convolution module of a cross-variable space dimension.

2. The intelligent grid-connected photovoltaic array fault diagnosis method according to claim 1, characterized in that: The training method of the fault diagnosis model includes: Inputting the training data set into the initial model, performing temporal convolution and spatial convolution on the training data set using the time channel separation convolution module of the initial model to obtain convolution features; Reduce feature dimensions through maximum pooling and global average pooling, input the pooled features into the fully connected layer, and output the classification results through softmax; The cross entropy loss function is used to predict the loss value of the classification result and the true label. The parameters are adjusted through back propagation until the loss value is minimized to obtain a trained fault diagnosis model.

3. The intelligent grid-connected photovoltaic array fault diagnosis method according to claim 2, characterized in that: The method of using the time channel separation convolution module of the initial model to perform time convolution and space convolution on the training data set to obtain convolution features includes: Using the time convolution module to perform time convolution on the multivariate time series of the training data set to extract the time pattern features of the training data set; performing batch normalization on the temporal pattern features, and performing linearization processing on the temporal pattern features using an exponential linear unit; The spatial convolution module is used to perform cross-channel interactive spatial convolution on the temporal pattern features after feature processing to obtain convolution features.

4. The intelligent grid-connected photovoltaic array fault diagnosis method according to claim 1, characterized in that: The temporal convolution module includes A one-dimensional convolution kernel of size t×1, the spatial convolution module includes The size is ×d one-dimensional convolution kernel; in, and is a positive integer, t is the time range, and d is the number of input channels; The total amount of parameters of the temporal convolution module and the spatial convolution module satisfies the following formula: 。 5. The intelligent grid-connected photovoltaic array fault diagnosis method according to claim 3, characterized in that: The batch normalizing the temporal pattern features comprises: Determine the batch size for normalization, and obtain the mean of the batch data based on the batch size and the training data set; Obtaining batch data variance according to the data mean, batch data size, and training data set; Normalized data is determined according to the batch data variance, the data mean, the training data set, and the minimum positive constant.

6. The intelligent grid-connected photovoltaic array fault diagnosis method according to claim 1, characterized in that: Also includes: Make sure the normalized data satisfies the normal distribution with mean 0 and variance 1; Scaling parameters and translation parameters are introduced to perform scaling and translation transformations on the normalized data, so that the fault diagnosis model can learn and use parameters by itself.

7. The intelligent grid-connected photovoltaic array fault diagnosis method according to claim 1, characterized in that: The sample data set includes m samples, and the time series length of each sample is N; the training label set includes m labels, and each label includes c fault categories; the fault categories include no fault, first open circuit fault, second short circuit fault and shadow obstruction fault; wherein m, N, and c are all positive integers.

8. An intelligent grid-connected photovoltaic array fault diagnosis device, characterized in that: include: A collection unit, used for collecting a first voltage, a second voltage, a first current and a second current at a DC output side of a photovoltaic array; A data set generating unit, configured to generate a sample data set according to the first voltage, the second voltage, the first current and the first current at fixed time intervals; A training unit, used to divide the sample data set into a training data set and a test data set, and train the initial fault diagnosis model using the training data set to obtain a trained fault diagnosis model; An identification unit, used to identify the test data set using the trained fault diagnosis model to obtain the fault type; The fault diagnosis model is constructed based on a time channel separation convolution module, and the time channel separation convolution module includes a time convolution module of a univariate time dimension and a space convolution module of a cross-variable space dimension.

9. An electronic device, characterized in that: include: at least one memory for storing a computer program; At least one processor is used to execute the program stored in the memory. When the program stored in the memory is executed, the processor is used to execute the method according to any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that: When the computer program runs on a processor, the processor is caused to execute the method according to any one of claims 1 to 7.

Images