Photovoltaic array fault detection method based on improved artificial neural network

Abstract

The invention discloses a photovoltaic array fault detection method based on an improved artificial neural network. A photovoltaic array fault diagnosis model constructed by the photovoltaic array fault diagnosis method comprises an input layer, a gating convolution layer, a flattening layer, an output layer and one or more fully connected layers connected in sequence. By adding a gating convolution layer and a plurality of full connection layers, the problem of gradient disappearance existing in existing photovoltaic array fault diagnosis is solved; according to the method, the adaptive Softmax activation function is used as the activation function of the output layer, so that the accuracy and the stability of the model are improved; in addition, the photovoltaic array is constructed through analogue simulation, and a large number of working state parameters of the photovoltaic array can be obtained according to module parameters of the photovoltaic array under the condition of less data; according to the method, complex tasks with big data sets can be processed, rapid prediction can be provided after training, and the method has good generalization ability and extremely high feasibility and application prospects.

Description

Technical Field

[0001] The invention belongs to the technical field of solar photovoltaic fault detection, and in particular relates to a photovoltaic array fault detection method based on an improved artificial neural network. Background Art

[0002] In recent years, among all major energy types, the number of solar photovoltaic installations has experienced rapid growth. According to the latest data from the International Energy Agency (IEA), the global photovoltaic installed capacity will reach 420GW in 2023, a significant increase of 85% from 228GW in 2022. However, solar photovoltaic systems are installed in complex outdoor environments. As an important part of photovoltaic power stations, the photovoltaic array has a high failure rate in the entire photovoltaic system. Harsh external factors make the array more prone to various failures such as shadowing, aging, short circuit and open circuit. These failures have a significant impact on the reliability and performance of solar photovoltaic systems. The power loss caused by failures in solar photovoltaic systems is about 18.9% each year. Therefore, it is of great significance to monitor and supervise the control of solar photovoltaic arrays to detect, classify and diagnose possible failures to reduce losses and ensure the normal operation of the system.

[0003] In recent years, a large number of scholars and experts have proposed various advanced artificial intelligence algorithms such as decision trees, extreme learning machines and neural networks for solar photovoltaic array fault diagnosis, which has improved the stability of photovoltaic systems. Although different machine learning methods and deep learning methods are gradually replacing traditional fault diagnosis methods such as manual detection, each method has both advantages and disadvantages. Siva Ramakrishna Madeti et al. proposed a photovoltaic fault diagnosis model based on the k-nearest neighbor (kNN) algorithm, which has a high accuracy rate. However, as the dimension of the data set increases, the increase in the computational complexity of the kNN algorithm will lead to a slow speed, and the algorithm accuracy is greatly affected by noise. Liu Xingxing et al. used a fault diagnosis model based on extreme gradient boosting (XGBoost) to diagnose photovoltaic arrays, and used grid search and cross-validation to optimize the hyperparameters of the diagnosis model. The obtained model has good generalization and high accuracy. However, the spatial complexity of XGBoost in the training process is too high, which requires a large amount of computing resources and may cause problems such as insufficient memory. Appiah AY et al. proposed a photovoltaic array fault diagnosis technology based on long short-term memory network (LSTM), which can automatically extract the features of the original data of photovoltaic array fault classification for classification. Experiments show that the proposed method has high accuracy, but the LSTM has high computational complexity and may have insufficient generalization ability when there is less data.

[0004] Artificial neural network (ANN) is a mainstream method of machine learning. It can capture feature information in a large amount of data and establish an effective fault diagnosis model in the field of DC side fault diagnosis of photovoltaic arrays, and is therefore widely used. Hussain verified that ANN has advantages over other methods in the field of photovoltaic fault diagnosis by using t-test technology, because the ANN model can determine the differences in power and irradiance levels of photovoltaic arrays and has the potential to distinguish changes in these parameters, including shade and non-shade conditions. Cherifa proposed using ANN to conduct photovoltaic fault diagnosis research on small grid-connected photovoltaic generators, and experiments showed that this method has a high accuracy rate. Adel conducted comparative experiments on fault detection and classification of photovoltaic arrays by applying machine learning methods such as ANN, KNN, and XGBoost, and concluded that ANN is more accurate and has extremely high feasibility in classification. Similarly, Baojie Li et al. used eight fault diagnosis methods to conduct comparative experiments on photovoltaic fault diagnosis, and concluded that ANN is the best fault diagnosis model among them. Compared with other machine learning methods, ANN can be used to solve nonlinear models, can handle complex tasks with large data sets, and provide fast predictions after training, and also has good generalization ability. However, the neurons in the hidden layer of ANN will be compressed when propagating forward, causing the weights close to the output layer to be close to the optimal solution, while the weights close to the input are still in a random solution state, resulting in the gradient vanishing phenomenon. When the gradient vanishing phenomenon is mild, it will affect the training efficiency and cause a decrease in accuracy. When it is severe, it will cause the model to be unable to learn and the fault diagnosis results will be almost random guesses. Summary of the invention

[0005] In view of the gradient vanishing problem existing in the existing ANN model in the field of photovoltaic array fault diagnosis, the present invention provides a photovoltaic array fault detection method based on an improved artificial neural network to achieve accurate prediction and diagnosis of solar photovoltaic faults.

[0006] The present invention proposes a photovoltaic array fault detection method based on an improved artificial neural network, which comprises the following steps: Step 1: Obtain electrical parameter data of the photovoltaic array under different working states as a sample set, and divide the sample set into training samples and test samples; the working state of the photovoltaic array includes a normal state and a fault state.

[0007] Step 2, construct a photovoltaic array fault diagnosis model; the photovoltaic array fault diagnosis model includes an input layer, a gated convolution layer, a flattening layer, an output layer, and one or more fully connected layers connected in sequence; the gated convolution layer includes a feature extraction unit and a gated selection unit; after the electrical parameters are input into the input layer, the features are extracted by the feature extraction unit, and the gated selection unit is used to filter the noise; the flattening layer is used to flatten the two-dimensional array output by the gated convolution layer into a one-dimensional array and then input into the fully connected layer to complete the classification task, and then input into the output layer.

[0008] Step 3: Use the training samples to train the photovoltaic array fault diagnosis model, and use the model with the highest accuracy during the training process as the optimal photovoltaic array fault diagnosis model.

[0009] Step 4: Use the test samples to test the optimal photovoltaic array fault diagnosis model. If the diagnosis accuracy of the optimal photovoltaic array fault diagnosis model is lower than the preset value, rebuild the photovoltaic array fault diagnosis model, and use the training samples to train the reconstructed photovoltaic array fault diagnosis model to update the optimal photovoltaic array fault diagnosis model; repeat the above process until the accuracy of the optimal photovoltaic array fault diagnosis model is higher than the preset value.

[0010] Step 5: Use the optimal photovoltaic array fault diagnosis model to perform fault diagnosis according to the electrical parameters of the tested solar photovoltaic array, obtain the diagnosis results, and process the photovoltaic array according to the diagnosis results.

[0011] Preferably, in the step 1, the method for acquiring the electrical parameter data is as follows: a photovoltaic array simulation model is established, and the electrical parameters of the photovoltaic array in normal state and fault state are simulated, and the electrical parameters in the same working state are labeled with the same state.

[0012] Preferably, in the photovoltaic array simulation model, the irradiance range is set to 100W / m 2 ~1000W / m 2 , the change step is 50W / m 2 ; The temperature range is set to 10℃~48℃, and the change step is 2℃.

[0013] Preferably, in step 2, the number of fully connected layers in the photovoltaic array fault diagnosis model is two.

[0014] Preferably, the fault states include single-line faults and parallel faults; single-line faults include short circuit, partial open circuit, aging, shadow shielding and dirt shielding; parallel faults include open circuit shadow faults and open circuit dirt faults.

[0015] Preferably, the electrical parameters include irradiance, temperature, open circuit voltage, short circuit current, operating voltage and operating current.

[0016] Preferably, the activation function in the output layer adopts an adaptive Softmax activation function; the activation function in the fully connected layer adopts a Relu activation function.

[0017] Preferably, the method for reconstructing the photovoltaic array fault diagnosis model is: resetting the number of filters and neurons and the size of the convolution kernel.

[0018] Preferably, the output of each neuron in the fully connected layer for: in, is the bias of the i-th neuron in the l-th layer; For the The number of neurons in the layer; For the The weight learned between the jth neuron in layer 1 and the i-th neuron in layer 2; For the The output of the jth neuron in the layer after the activation function.

[0019] Preferably, the gated convolutional layer is defined as follows: Among them, Conv1 and Conv2 are the outputs of the gate selection unit and the feature extraction unit respectively; W g and W f They are convolution filters for calculating gated values ​​and eigenvalues ​​respectively; I is the channel value obtained by the convolution kernel at the feature position; O is the output of the gated convolution layer.

[0020] Preferably, in step 3, the training process is: a. Set the cross entropy function as the loss function Loss; b. Using the training samples to train the initial photovoltaic array fault diagnosis model until the number of training Epochs is greater than the maximum number of training Epochs Max_Epochs, a trained photovoltaic array fault diagnosis model is obtained; in this embodiment, the maximum number of training Epochs is 100.

[0021] c. Using training samples ω train The validation set separated from the training set calculates the error and accuracy of the model, and saves the network parameters at all levels.

[0022] d. Repeat the above process until the number of iterations Iterator_Times is greater than the set maximum number of iterations T.

[0023] e. Compare the photovoltaic array fault diagnosis models obtained in multiple iterations, and select the model with the highest accuracy as the optimal photovoltaic array fault diagnosis model.

[0024] The present invention has the following beneficial effects: 1. The present invention solves the gradient vanishing problem existing in the existing photovoltaic array fault diagnosis by adding a gated convolution layer and multiple fully connected layers to the photovoltaic array fault diagnosis model; at the same time, the present invention uses an adaptive Softmax activation function as the activation function of the output layer, which improves the accuracy and stability of the model.

[0025] 2. The present invention constructs a photovoltaic array through simulation, and can obtain a large number of working status parameters of the photovoltaic array according to the module parameters of the photovoltaic array when there is less data; at the same time, the present invention can handle complex tasks with large data sets and provide fast predictions after training, and has good generalization ability and extremely high feasibility and application prospects. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 It is the overall flow chart of the present invention.

[0027] Figure 2 It is a schematic diagram of the photovoltaic array simulation model in the present invention.

[0028] Figure 3 It is a schematic diagram of the structure of the photovoltaic array fault diagnosis model in the present invention.

[0029] Figure 4 Schematic diagrams of confusion matrices of the present invention, XGBoost, KNN, ANN and LSTM; wherein, (a) is a schematic diagram of the confusion matrix of the present invention; (b) is a schematic diagram of the confusion matrix of XGBoost; (c) is a schematic diagram of the confusion matrix of KNN; (d) is a schematic diagram of the confusion matrix of ANN; and (e) is a schematic diagram of the confusion matrix of LSTM. DETAILED DESCRIPTION

[0030] The present invention will be further described below in conjunction with the accompanying drawings.

[0031] like Figure 1 As shown, a photovoltaic array fault detection method based on an improved artificial neural network includes the following steps: Step 1: Figure 2As shown in the figure, a photovoltaic module of model lSoltech lSTH-350-WH is selected in MATLAB / Simulink software to establish a 4×3 array photovoltaic array simulation model; the module parameters of the photovoltaic array are set as follows: maximum power point voltage is 43V, maximum power point current is 8.13A, open circuit voltage is 51.5V, and short circuit current is 9.4A. In order to cover a variety of operating environments, the input irradiance range is set to 100W / m through the constant module in MATLAB / Simulink software. 2 ~1000W / m 2 , the change step is 50W / m 2 ; The temperature range is set to 10℃~48℃, and the step size is 2℃. The established photovoltaic array simulation model simulates the actual single-row faults and parallel faults of the photovoltaic array; single-row faults include short circuit, partial open circuit, aging, shadow shading and dirt shading; parallel faults include open circuit shadow faults and open circuit dirt faults. In the simulation experiment, multiple sets of data are obtained through the Goto, From and Scope modules. Each set of data includes irradiance, temperature, open circuit voltage, short circuit current, working voltage and working current, and each set of data is labeled. A complete data set is formed through multiple sets of different experiments. The labeled multiple sets of data are used as the sample set ω. The data in the sample set ω are shown in Table 1.

[0032] Table 1. Running status, quantity and labels of simulation datasets Running status describe quantity Label normal Trouble-free operation 400 0 Short Circuit Two levels, one or two photovoltaic modules are short-circuited, 400 sets of data each 800 1 Partial open circuit Two levels, one branch open and two branches open, 400 sets of data each 800 2 aging Connect the same branch to three different levels of resistance in sequence and obtain 400 sets of data respectively. 1200 3 Shadow Occlusion Reduce the irradiance of the entire photovoltaic array 400 4 Dirt blocking Two levels of dirt shielding, reducing the irradiance of one and two photovoltaic modules in a branch, 400 sets of data each 800 5 Open circuit shadow Open circuit and shadow faults 400 6 Open Circuit Dirt Open circuit and dirt faults 400 7 Since the sample set ω has different types of data with large numerical differences, which leads to a decrease in the model fault diagnosis accuracy and generalization ability, in order to better train the improved artificial neural network model, the above-obtained data set is normalized. The normalization formula is as follows: Among them, X is the normalized data; x is the original data of the sample set ω; x max and x min are the maximum and minimum values ​​in the original data respectively.

[0033] Use one-hot encoding to divide the sample set ω into categories with relatively independent characteristics to complete fault classification. Divide the processed sample set ω into training samples ω at a ratio of 7:3 train and the test sample ω test , the number of cross-validation iterations is set to 10.

[0034] Step 2: Build a photovoltaic array fault diagnosis model like Figure 3As shown in the figure, the photovoltaic array fault diagnosis model includes an input layer, a gated convolution layer, a Flatten layer, two fully connected layers and an output layer connected in sequence. The gated convolution layer includes a feature extraction unit and a gated selection unit, which combines the advantages of the gated recurrent unit (GRU) and the convolutional neural network. The feature extraction unit generates new features from the input data, improving the feature extraction capability; the gated selection unit retains the update gate and reset gate in the GRU in order to simplify the model structure and adapt to the characteristics of the convolution operation. Compared with LSTM, GRU, as a variant of RNN, has been simplified in structure, so that GRU can more easily capture the timing information of the input signal and will not produce the advantage of gradient disappearance over time. The update gate and reset gate in the gated selection unit can screen features, selectively retain useful information, filter out unimportant or noisy information, and select features that mainly act on the generation of results for backward transmission, thereby improving the model's expressiveness and generalization performance. The reset process, update process and output propagation process of GRU are as follows:

[0035] r t = σ( W r · [ ℎ t−1 , x t ]+ b r ) z t = σ( W z · [ ℎ t−1 , x t ]+ b z ) Among them, r t and z t are the outputs of the reset gate and the update gate, 0 <r t ,z t <1; σ is the sigmoid function; W r and W z are the weight matrices corresponding to the reset gate and the update gate respectively; h t-1 is the hidden state of the previous time step; x t is the input of the current time step; b r and bz is the bias term for the reset gate and update gate; is a candidate hidden state.

[0036] The definition of the gated convolutional layer is as follows: Among them, Conv1 and Conv2 are the outputs of the gate selection unit and the feature extraction unit respectively; W g and W f are the convolution filters for calculating the gated value and the eigenvalue respectively; I is the channel value obtained by the convolution kernel at the feature position; O is the output of the gated convolution layer; is the product at the pixel level.

[0037] In this embodiment, the convolution kernel size in the gated convolution layer is 3, and the number of filters is 16.

[0038] Since the output of the gated convolutional layer is a two-dimensional array, but the fully connected layer usually requires a one-dimensional input array, directly using the fully connected layer to process high-dimensional input data will require a very large number of parameters, resulting in a waste of computing resources and an increased risk of overfitting. Therefore, the dimensionality reduction process from two dimensions to one dimension is particularly important.

[0039] The two-dimensional array output by the gated convolutional layer is mapped to a one-dimensional array through the flattening layer, which greatly reduces the redundant parameters of the network structure. The calculation formula of the flattening layer is: Among them, x ij is the data in the i-th row and j-th column in the two-dimensional array; array is an array.

[0040] The one-dimensional array output by the flattening layer is input into the fully connected layer; two layers of fully connected layers play a classification role and have stronger classification ability than one layer of fully connected layers. Although more fully connected layers will further improve the classification ability of the model, they will also increase the complexity and computational cost of the model. In order to improve the performance of the model as much as possible, two layers of fully connected layers are used in the present invention. Each layer of the fully connected layer contains a large number of neurons, each neuron is connected to the neurons of the previous layer, receives the output of the neurons of the previous layer, and outputs it to the next layer. The fully connected layer is essentially a linear conversion from one feature space to another feature vector. For the i-th neuron in the l-th layer, its output z i The expression of (l) is:

[0041] in, is the bias of the i-th neuron in the l-th layer; For the The number of neurons in the layer; For the The weight learned between the jth neuron in layer 1 and the i-th neuron in layer 2; For the The output of the jth neuron in the layer after the activation function.

[0042] Each fully connected layer is connected to an activation function; since the Relu activation function is nonlinear, it does not have the problem of gradient explosion compared to other activation functions. And using nonlinear activation functions, the neural network can not only perform linear transformations, but also process and represent nonlinear relationships in data, so the Relu activation function is selected as the activation function of the fully connected layer.

[0043] The activation function of the output layer uses Adaptive Softmax; Adaptive Softmax is sufficiently universal, and using Adaptive Softmax as the classifier of the improved ANN fault diagnosis model can effectively reduce the impact of increased computing time due to the large size of the data set without affecting the accuracy. Step 3: Use training sample ω train The photovoltaic array fault diagnosis model is trained; during the training process, the error between the expected utility distribution and the target distribution is evaluated using the cross entropy loss function, and then the Adam optimization algorithm is used to optimize the loss function value. The specific process of the Adam optimization algorithm to optimize the loss function value is as follows:

[0044] 3-1. Set the cross entropy function as the loss function Loss; 3-2. Using training samples ω train The initial photovoltaic array fault diagnosis model is trained until the number of training Epochs is greater than the maximum number of training Epochs Max_Epochs, thereby obtaining a trained photovoltaic array fault diagnosis model. In this embodiment, the maximum number of training Epochs is 100.

[0045] 3-3. Using training samples ω train The validation set separated from the training set calculates the error and accuracy of the model, and saves the network parameters at all levels.

[0046] 3-4. Repeat the above process until the number of iterations Iterator_Times is greater than the set maximum number of iterations T.

[0047] 3-5. Compare the photovoltaic array fault diagnosis models obtained in multiple iterations, and select the model with the highest accuracy as the optimal photovoltaic array fault diagnosis model.

[0048] Step 4: Test sample ω test Input the optimal photovoltaic array fault diagnosis model, output the fault diagnosis results of the test samples, and verify the effectiveness of the model established in step 4; if the accuracy is lower than the preset value, reset the number of filters and neurons and the size of the convolution kernel, and retrain the updated photovoltaic array fault diagnosis model to update the optimal photovoltaic array fault diagnosis model; repeat the above process until the accuracy of the optimal photovoltaic array fault diagnosis model is higher than the preset value.

[0049] Step 5: Model Evaluation The model of the present invention, XGBoost, KNN, ANN and LSTM were tested respectively, and the results output by the expected utility module were obtained as the diagnostic basis for fault diagnosis. The final model training confusion matrix effect is as follows: Figure 4 The accuracy comparison of the five models is shown in Table 2. From Table 2 and Figure 4 It can be seen that the present invention is superior to the other three models in terms of single-type fault diagnosis accuracy and overall fault diagnosis accuracy, and the diagnostic accuracy of its model can eventually reach 98.6%, and the model establishment is completed.

[0050] Table 2 Accuracy of five diagnostic models in the dataset Model Name The present invention XGBoost KNN ANN LSTM Accuracy 98.6% 88.4% 88.0% 96.0% 92.0% Step 6: Input the parameters of the solar photovoltaic array to be tested into the photovoltaic array fault diagnosis model to diagnose the solar photovoltaic fault to be tested.

Claims

1. A photovoltaic array fault detection method based on an improved artificial neural network, characterized in that: The following steps are involved: Step 1: Obtain electrical parameter data of the photovoltaic array under different working states as a sample set, and divide the sample set into training samples and test samples; the working state of the photovoltaic array includes a normal state and a fault state; Step 2: Construct a photovoltaic array fault diagnosis model; the photovoltaic array fault diagnosis model includes an input layer, a gated convolution layer, a flattening layer, an output layer, and one or more fully connected layers connected in sequence; the gated convolution layer includes a feature extraction unit and a gated selection unit; after the electrical parameters are input into the input layer, the features are extracted by the feature extraction unit, and the gated selection unit is used to filter the noise; the flattening layer flattens the two-dimensional array output by the gated convolution layer into a one-dimensional array and then inputs it into the fully connected layer to complete the classification task, and then inputs it into the output layer; Step 3: Use the training samples to train the photovoltaic array fault diagnosis model, and use the model with the highest accuracy in the training process as the optimal photovoltaic array fault diagnosis model; Step 4: Use the test samples to test the optimal photovoltaic array fault diagnosis model. If the diagnosis accuracy of the optimal photovoltaic array fault diagnosis model is lower than the preset value, rebuild the photovoltaic array fault diagnosis model, and use the training samples to train the reconstructed photovoltaic array fault diagnosis model to update the optimal photovoltaic array fault diagnosis model; repeat the above process until the accuracy of the optimal photovoltaic array fault diagnosis model is higher than the preset value; Step 5: Use the optimal photovoltaic array fault diagnosis model to perform fault diagnosis according to the electrical parameters of the tested solar photovoltaic array to obtain the diagnosis results.

2. A photovoltaic array fault detection method based on an improved artificial neural network according to claim 1, characterized in that: In the step 1, the method for acquiring the electrical parameter data is as follows: a photovoltaic array simulation model is established, and the electrical parameters of the photovoltaic array in normal state and fault state are simulated, and the electrical parameters in the same working state are labeled with the same state.

3. A photovoltaic array fault detection method based on an improved artificial neural network according to claim 2, characterized in that: In the photovoltaic array simulation model, the irradiance range is set to 100W / m 2 ~1000W / m 2 , the change step is 50W / m 2 ; The temperature range is set to 10℃~48℃, and the change step is 2℃.

4. The photovoltaic array fault detection method based on improved artificial neural network according to claim 1, characterized in that: In the step 2, the number of fully connected layers in the photovoltaic array fault diagnosis model is two.

5. The photovoltaic array fault detection method based on improved artificial neural network according to claim 1, characterized in that: The fault states include single-line faults and parallel faults; single-line faults include short circuits, partial open circuits, aging, shadowing and dirt blocking; parallel faults include open circuit shadow faults and open circuit dirt faults.

6. A photovoltaic array fault detection method based on an improved artificial neural network according to claim 1, characterized in that: The electrical parameters include irradiance, temperature, open circuit voltage, short circuit current, operating voltage and operating current.

7. A photovoltaic array fault detection method based on an improved artificial neural network according to claim 1, characterized in that: The activation function in the output layer adopts the adaptive Softmax activation function; the activation function in the fully connected layer adopts the Relu activation function.

8. A photovoltaic array fault detection method based on an improved artificial neural network according to claim 1, characterized in that: The method for reconstructing the photovoltaic array fault diagnosis model is: resetting the number of filters and neurons and the size of the convolution kernel.

9. A photovoltaic array fault detection method based on an improved artificial neural network according to claim 1, characterized in that: The output z of each neuron in the fully connected layer is i (l) is: in, is the bias of the i-th neuron in the l-th layer; For the The number of neurons in the layer; For the The weight learned between the jth neuron in layer 1 and the i-th neuron in layer 2; For the The output of the jth neuron in the layer after the activation function.

10. A photovoltaic array fault detection method based on an improved artificial neural network according to claim 1, characterized in that: In the step 3, the training process is: a. Set the cross entropy function as the loss function Loss; b. Use the training samples to train the initial photovoltaic array fault diagnosis model until the number of training Epochs is greater than the maximum number of training Max_Epochs, and obtain the trained photovoltaic array fault diagnosis model c. Use the validation set separated from the training sample to calculate the model's error and accuracy, and save the network parameters at all levels; d. Repeat the above process until the number of iterations Iterator_Times is greater than the set maximum number of iterations T; e. Compare the photovoltaic array fault diagnosis models obtained in multiple iterations, and select the model with the highest accuracy as the optimal photovoltaic array fault diagnosis model.

Images