Multi-Feature Fusion Photovoltaic Fault Diagnosis Method Based on Bagging Algorithm

Abstract

The present invention discloses a multi-feature fusion photovoltaic fault diagnosis method based on the Bagging algorithm, which includes the following steps: sample data acquisition; extracting transient features and steady-state features from the sample dataset, and respectively associating them with non-electrical data to obtain fusion features; constructing a photovoltaic fault diagnosis model based on the Bagging algorithm, and performing training and testing; collecting electrical data and non-electrical data of the distributed photovoltaic system to be diagnosed, extracting fusion features and performing photovoltaic fault diagnosis through the training of the photovoltaic fault diagnosis model. By constructing the steady-state features and transient features of photovoltaic faults, combining the electrical data and non-electrical data of photovoltaic faults, the present invention constructs a fusion feature library, and further constructs a photovoltaic fault prediction model based on the Bagging algorithm. After completing the training and testing of the model, fault diagnosis is carried out through the fusion features of the electrical data and non-electrical data of the photovoltaic system, and the diagnosis accuracy is significantly improved.

Description

Technical Field

[0001] The present invention relates to a multi - feature fusion photovoltaic fault diagnosis method based on the Bagging algorithm, belonging to the technical fields of intelligent power and power grid. Background Art

[0002] While the photovoltaic power generation is developing rapidly, its operation safety problems are becoming increasingly prominent. Photovoltaic power stations are generally built outdoors with abundant sunlight and are greatly affected by natural conditions. When operating in harsh natural environments, photovoltaic power stations are prone to problems such as component aging, partial shading, open circuits, and short circuits. These problems seriously affect the health of the photovoltaic system, the overall operation of the photovoltaic power station, and the power generation efficiency. If the faults are not handled in a timely manner, it is extremely easy to cause the photovoltaic panels and busbar boxes to burn out, resulting in economic losses and threatening the safety of the personnel in the station. Therefore, quickly detecting whether a fault occurs in the photovoltaic system and accurately diagnosing the fault type are crucial for the safe and reliable operation of the photovoltaic power supply system.

[0003] Existing photovoltaic fault diagnosis methods generally include infrared image diagnosis method, mathematical model diagnosis method, and artificial intelligence diagnosis method. Based on the infrared imaging principle, a method for automatically detecting photovoltaic hot spots is proposed, which improves the automation level of hot spot fault diagnosis. However, such methods overly rely on the precision of infrared devices, are also easily affected by the external environment, have a high cost, and can only diagnose a single fault type. Based on the mathematical model diagnosis method, the fault diagnosis of photovoltaic power stations is realized. This method only needs to provide photovoltaic operation data and does not rely on expensive infrared detection devices. However, the diagnosis effect depends on the accuracy of the model and it is difficult to be applied online.

[0004] Therefore, a multi - feature fusion photovoltaic fault diagnosis method based on the Bagging algorithm is needed to solve the above problems. Summary of the Invention

[0005] Object of the Invention: Aiming at the problems existing in the prior art, the present invention provides a multi - feature fusion photovoltaic fault diagnosis method based on the Bagging algorithm.

[0006] A multi - feature fusion photovoltaic fault diagnosis method based on the Bagging algorithm,

[0007] comprises the following steps:

[0008] Step 1, sample data acquisition: Collect the electrical data and non - electrical data of the distributed photovoltaic system in normal and fault states to obtain a sample data set;

[0009] Step 2, construct a fused feature sample data set: Extract the transient features and steady - state features of the distributed photovoltaic system from the sample data set in Step 1, and associate them with the corresponding non - electrical data to obtain fused features;

[0010] Specifically, each group of transient features, steady-state features, and corresponding non-electrical data are used as a fused feature sample data group to obtain a fused feature sample data set; each sample data in the fused feature sample data set is a data group.

[0011] Step 3: Construct a photovoltaic fault diagnosis model based on the Bagging algorithm. Divide the fused feature sample data set into a training set and a test set, and use the training set to train the photovoltaic fault diagnosis model.

[0012] Step 4: Use the test set to test the photovoltaic fault diagnosis model. When the accuracy rate reaches the target threshold P, the training is completed; otherwise, return to Step 3.

[0013] Step 5: Photovoltaic fault diagnosis: Collect the electrical data and non-electrical data of the distributed photovoltaic system to be diagnosed, extract the fused features, and perform photovoltaic fault diagnosis through the training of the photovoltaic fault diagnosis model.

[0014] Furthermore, in Step 1, the normal and fault states include normal, short circuit, open circuit, aging, and shadow occlusion. The electrical data includes the current I, voltage V, output power P, and photovoltaic cell temperature T of each photovoltaic string, and the non-electrical data includes the operating duration , ambient temperature , ambient humidity H, and irradiance E.

[0015] Furthermore, in Step 2, the steady-state features include output power P, output current I, voltage V, average output power , average output current , average voltage , absolute value of the power difference before and after the fault occurs , absolute value of the output current difference before and after the fault occurs and absolute value of the voltage difference before and after the fault occurs ;

[0016] In Step 2, the transient features include transient time , transient power change coefficient , transient current change coefficient and transient voltage change coefficient .

[0017] Furthermore, the transient time , transient power change coefficient , transient current change coefficient and transient voltage change coefficient are expressed by the following formula:

[0018] ,

[0019] Among them, is the transient time, is the transient power change coefficient, is the transient current change coefficient, is the transient voltage change coefficient; is the transient start time, is the transient end time; is the maximum value of the output power change amount during the transient process; is the maximum value of the current change amount during the transient process; is the maximum value of the voltage change amount during the transient process; is the absolute value of the power difference before and after the fault occurs, is the absolute value of the output current difference before and after the fault occurs, is the absolute value of the voltage difference before and after the fault occurs.

[0020] Furthermore, the fused feature sample data set described in step three is divided into a training set and a test set according to a ratio of 7:3.

[0021] Furthermore, the fault diagnosis model described in step three includes multiple convolutional neural networks, and the multiple convolutional neural networks are aggregated by the Bagging algorithm.

[0022] Furthermore, the convolutional neural network includes a first basic residual block, a second basic residual block, a first bidirectional LSTM module, a second bidirectional LSTM module, a first conversion module, and a first fully connected layer module,

[0023] Both the first basic residual block and the second basic residual block include a convolutional layer, a layer normalization processing module, and an activation layer connected in sequence,

[0024] The first basic residual block and the second basic residual block are respectively connected to the first bidirectional LSTM module and the second bidirectional LSTM module through a reshaping module;

[0025] The first basic residual block, the second basic residual block, and the first conversion module are connected in sequence, and the first bidirectional LSTM module, the second bidirectional LSTM module, and the first conversion module are connected in sequence;

[0026] The first fully connected layer is connected to the first conversion module.

[0027] Furthermore, it includes a third basic residual block, a fourth basic residual block, a fifth basic residual block, a channel attention module, a second conversion module, and a second fully connected layer module connected in sequence. The third basic residual block, the fourth basic residual block, and the fifth basic residual block all include a convolutional layer, a layer normalization processing module, and an activation layer connected in sequence.

[0028] Furthermore, the parameters of the convolutional layer are k = 3 and c = 64, where k is the size of the convolutional kernel and c is the number of output channels of this convolutional layer; the activation function of the activation layer is the LeakyReLU function.

[0029] Furthermore, the target threshold P in step 4 is 90%.

[0030] Beneficial effects: The multi-feature fusion photovoltaic fault diagnosis method based on the Bagging algorithm of the present invention constructs the steady-state features and transient features of photovoltaic faults, combines the electrical data and non-electrical data of photovoltaic faults, constructs a fusion feature library, and further constructs a photovoltaic fault prediction model based on the Bagging algorithm. After the model is trained and tested, fault diagnosis is carried out through the fusion features of the electrical data and non-electrical data of the photovoltaic system, and the diagnosis accuracy is significantly improved. Description of the Drawings

[0031] Figure 1 is a schematic flow chart of the multi-feature fusion photovoltaic fault diagnosis method based on the Bagging algorithm of the present invention;

[0032] Figure 2 is a schematic structural diagram of the basic residual block;

[0033] Figure 3 is a schematic structural diagram of the convolutional neural network model 1 constructed based on the basic residual block;

[0034] Figure 4 is a schematic structural diagram of the convolutional neural network model 2 constructed based on the basic residual block;

[0035] Figure 5 is an implementation flow chart of the Bagging algorithm. Specific Embodiments

[0036] The following will describe the preferred embodiments of the present invention in conjunction with the drawings, and the technical solutions of the present invention will be described more clearly and completely.

[0037] Please refer to Figure 1 shown, the multi-feature fusion photovoltaic fault diagnosis method based on the Bagging algorithm of the present invention includes the following steps:

[0038] Step S1, sample data collection of normal and faulty states of the photovoltaic, specifically:

[0039] Collect 100 - 400 sets of sample data for normal and fault states of photovoltaic power generation respectively to form a sample data set. The normal and fault states of photovoltaic power generation include five states: normal, short - circuit, open - circuit, aging, and shadow occlusion. Each set of sample data includes electrical data and non - electrical data. The electrical data includes the current I, voltage V, output power P, and photovoltaic cell temperature T of each photovoltaic string. The non - electrical data includes the operation duration T s 、ambient temperature T h 、ambient humidity H, and irradiance E;

[0040] Step S2: Construct a fusion feature library. The fusion feature library includes steady - state features, transient features, and non - electrical features. Specifically:

[0041] Step S21: The steady - state features include output power P, output current I, voltage V, average output power P j 、average output current I j 、average voltage Vj, absolute value of power difference ΔP before and after fault occurrence, absolute value of output current difference ΔI before and after fault occurrence, and absolute value of voltage difference ΔV before and after fault occurrence;

[0042] Step S22: The transient features include transient time ΔT, transient power change coefficient k P 、transient current change coefficient k I 、and transient voltage change coefficient k V ;

[0043] ,

[0044] where T et is the transient start time, T st is the transient end time; ΔP peak is the maximum value of the output power change during the transient process; ΔI peak is the maximum value of the current change during the transient process; ΔV peak is the maximum value of the voltage change during the transient process;

[0045] Step S23: The non - electrical data includes operation duration T s 、ambient temperature T h 、ambient humidity H, and irradiance E;

[0046] Step S24: For the five states of normal, short - circuit, open - circuit, aging, and shadow occlusion of the photovoltaic, perform feature selection respectively based on the method of mutual information and recursive feature elimination, and obtain the optimal feature combination that can maximize the model performance by adding or removing specific feature variables;

[0047] ,

[0048] In the formula, represents the probability that X = x is the joint probability of their simultaneous occurrence. X represents one of five states, and Y represents a characteristic parameter represents the strength of the relationship between the characteristic parameter and the state

[0049] During the operation a relationship strength close to 0 is low. In this application, the value of each state and each characteristic parameter is calculated, and the characteristic parameters with a relationship strength ≤0.3 with each state are excluded

[0050] For the five states of normal, short - circuit, open - circuit, aging, and shadow occlusion of a photovoltaic, the electrical and non - electrical characteristics with a large relationship strength with each state are different. The judgment of each state does not require taking all the above - mentioned characteristic parameters. Taking multiple characteristic parameters with a large relationship strength for each state can reduce the amount of data used

[0051] Step S3, sample data processing, specifically:

[0052] Step S31: Process the sample data according to steady - state characteristics, transient characteristics, and non - electrical characteristics. The sequence length of the sample data is 100 - 600, and the feature dimension is 13

[0053] Step S32: Perform an identification row vector on the sample data, where the first column is the mark of the data name; the second column uses numbers to represent the photovoltaic state types, where: 1 is normal, 2 is short - circuit, 3 is open - circuit, 4 is aging, 5 is shadow occlusion

[0054] Step S33: Perform normalization processing on the sample data, and normalize the same - type sample data between (-1, 1). Let the sample data be , and the maximum - minimum normalization calculation formula used here is as follows:

[0055] ,

[0056] where: is the normalized data and are the maximum and minimum values in the sample data and are the right - hand endpoint and left - hand endpoint of the normalization interval of the sample data

[0057] Step S34: Divide the normalized sample data into a training set and a test set according to a ratio of 7:3

[0058] Step S4, construct a photovoltaic fault prediction model based on the Bagging algorithm, specifically:

[0059] Step S41: Construct at least two convolutional neural network models;

[0060] In the convolutional neural network model, set the convolutional layer parameters as k = 3 and c = 64; where: k is the size of the convolutional kernel, and c is the number of output channels of this convolutional layer;

[0061] In the neural network model, set the activation function of the activation layer as the LeakyReLU function;

[0062] Step S44: Based on the Bagging algorithm, construct a photovoltaic fault diagnosis model through the constructed convolutional neural network models,

[0063] Step S5, training of the photovoltaic fault diagnosis model, specifically:

[0064] Step S51: Randomly extract sample data from the training set obtained in Step S3 and put it into a new training data subset Di, i where i = the number of convolutional neural network models;

[0065] Step S52: Given two data sets D1 and D2, with size m, m S is the number of the same samples in the two data sets, and the overlap rate of D1 and D2 is:

[0066]

[0067] Set a threshold R T to limit the overlap rate of each subset, R T ≤5%:

[0068]

[0069] Step S53: Train the photovoltaic fault diagnosis model through the training data subset Di i ;

[0070] Step S6: Use the data in the test set to test the photovoltaic fault diagnosis model, set the accuracy target threshold P of the photovoltaic fault diagnosis model ≥ 90%, if the performance target threshold is reached, the training is completed, otherwise return to Step S5;

[0071] Step S7, photovoltaic fault diagnosis, specifically:

[0072] Use the photovoltaic fault diagnosis model, input the data to be diagnosed, and judge the load category according to the position of the maximum value in the output vector. The final output is 1, 2, 3, 4, 5, where: 1 represents normal, 2 represents short circuit, 3 represents open circuit, 4 represents aging, and 5 represents shadow occlusion.

[0073] Specific construction method of the convolutional neural network model in step S41:

[0074] Construction method of convolutional neural network model 1, including the following steps:

[0075] Step 1: Construct a basic residual block, as Figure 2 shown, specifically:

[0076] Step 11: Set the convolutional layer with parameters k = 3 and c = 64; where: k is the size of the convolutional kernel, and c is the number of output channels of this convolutional layer;

[0077] Step 12: Set the Layer Normalization module, which is used to calculate the mean and variance of each sample in the feature dimension and perform a normalization operation, so that the output of each sample has a similar scale, helping to reduce the variation of different samples in the feature dimension and improving the robustness and generalization ability of the model;

[0078] Step 13: Set the activation function of the activation layer as the LeakyReLU function;

[0079] Step 2: Based on the basic residual block in Step 1, construct a neural network model, as Figure 3 shown, which includes two basic residual blocks, two bidirectional LSTM modules, a conversion module, and a fully connected layer module; among them, the two basic residual blocks are basic residual block 1 and basic residual block 2 (Baseblock1, Baseblock2);

[0080] The two basic residual blocks are connected in sequence and are respectively connected to the two bidirectional LSTM modules through a reshaping module. The reshaping module performs the reshape operation, which is used to rearrange and expand the multi-dimensional array of the basic residual block;

[0081] The two bidirectional LSTM modules are connected in sequence. The parameter of the previous bidirectional LSTM module is set as N = 128 (BidirectionalLSTM, N = 128), and the parameter of the latter bidirectional LSTM module is set as N = 256 (BidirectionalLSTM, N = 256), where N is the number of hidden units of each bidirectional LSTM module;

[0082] The conversion module performs the Faltten operation, which is used to expand the multi-dimensional array into a one-dimensional array without changing the order of elements;

[0083] The fully connected layer module is used to perform a linear transformation on the input data and generate an output through an activation function, with the parameter set as n = 1, where n is the number of output nodes.

[0084] Method for constructing convolutional neural network model two, comprising the following steps:

[0085] Step S1: Construct a basic residual block, as Figure 2 shown, specifically:

[0086] Step S11: Set a convolutional layer with parameters k = 3 and c = 64; where: k is the size of the convolutional kernel and c is the number of output channels of this convolutional layer;

[0087] Step S12: Set a layer normalization processing module (LayerNormalization) to calculate the mean and variance of each sample in the feature dimension and perform a normalization operation, so that the output of each sample has a similar scale, which helps to reduce the variation of different samples in the feature dimension and improve the robustness and generalization ability of the model;

[0088] Step S13: Set the activation function of the activation layer to the LeakyReLU function;

[0089] Step S2: Based on the basic residual block in Step S1, construct a neural network model, as Figure 4 shown, which includes five sequentially connected basic residual blocks, a channel attention module, a conversion module, and a fully connected layer module; among them, the five basic residual blocks are: basic residual block 1, basic residual block 2, basic residual block 3 (Baseblock1, Baseblock2, Baseblock3);

[0090] The channel attention module (Channelattention) is used to weight different channel features on the channel scale;

[0091] The conversion module performs a Faltten operation to expand a multi-dimensional array into a one-dimensional array without changing the order of elements;

[0092] The fully connected layer module is used to perform a linear transformation on the input data and generate an output through an activation function, with the parameter set as n = N, where: n is the number of output nodes and N = 3;

[0093] Implementation method of the Bagging algorithm in Step S44, as Figure 5 shown:

[0094] (1) Establish training subsets: Generate multiple sub-sample sets from the original dataset by sampling with replacement (i.e., bootstrap sampling);

[0095] (2) Model training: Independently train a model on each sub-sample set, and these models can be any one of the two models (the models can be repeated);

[0096] (3)Result summary: The final prediction result is determined through a voting mechanism (majority voting); specifically:

[0097] Generally speaking, for a Bagging ensemble containing M models, for a new input x, its output y can be expressed as the average or voting result of all model outputs:

[0098] ,

[0099] where represents the prediction result of the m-th model for the input x.

[0100] The multi-feature fusion photovoltaic fault diagnosis method based on the Bagging algorithm of the present invention constructs the steady-state features and transient features of photovoltaic faults, combines the electrical data and non-electrical data of photovoltaic faults, constructs a fused feature library, and further constructs a photovoltaic fault prediction model based on the Bagging algorithm. After the training and testing of the model are completed, fault diagnosis is carried out through the fused features of the electrical data and non-electrical data of the photovoltaic system, and the diagnostic accuracy is significantly improved.

[0101] The present invention respectively performs feature selection through methods based on mutual information and recursive feature elimination, and obtains the optimal feature combination that can maximize the model performance by adding or removing specific feature variables. Different fused features are selected for the five states of normal, short circuit, open circuit, aging, and shadow occlusion of the photovoltaic, thereby reducing the amount of data required for model training and diagnosis, and further improving the performance of the model.

Claims

1. A multi-feature fusion photovoltaic fault diagnosis method based on the Bagging algorithm, It is characterized in that it includes the following steps: Step 1, sample data collection: Collect electrical data and non-electrical data of the distributed photovoltaic system in normal and faulty states to obtain a sample data set; Step 2, construct a fused feature sample data set: Extract the transient features and steady-state features of the distributed photovoltaic system from the sample data set in Step 1, and associate them with the corresponding non-electrical data respectively to obtain fused features; In Step 2, the steady-state features include output power P, output current I, voltage V, average output power P j , average output current I j , average voltage V j , absolute value of power difference ΔP before and after fault occurrence, absolute value of output current difference ΔI before and after fault occurrence, and absolute value of voltage difference ΔV before and after fault occurrence; In step two, the transient characteristics include transient time ΔT, transient power change coefficient k p , transient current change coefficient k I and transient voltage change coefficient k V ; The transient time ΔT and the transient power change coefficient k p , the transient current change coefficient k I and the transient voltage change coefficient k V are expressed by the following formula: ΔT = T et - T st where ΔT is the transient time, k p is the transient power change coefficient, k I is the transient current change coefficient, k V is the transient voltage change coefficient; T et is the transient start time, T st is the transient end time; ΔP peak is the maximum value of the output power change during the transient process; ΔI peak is the maximum value of the current change during the transient process; ΔV peak is the maximum value of the voltage change during the transient process; ΔP is the absolute value of the power difference before and after the fault occurs, ΔI is the absolute value of the output current difference before and after the fault occurs, and ΔV is the absolute value of the voltage difference before and after the fault occurs; Step 3, construct a photovoltaic fault diagnosis model based on the Bagging algorithm, divide the fused feature sample data set into a training set and a test set, and use the training set to train the photovoltaic fault diagnosis model; Step 4, use the test set to test the photovoltaic fault diagnosis model. When the accuracy rate reaches the target threshold P, the training is completed; otherwise, return to Step 3; Step 5, photovoltaic fault diagnosis: Collect electrical data and non-electrical data of the distributed photovoltaic system to be diagnosed, extract fused features and perform photovoltaic fault diagnosis through the training of the photovoltaic fault diagnosis model.

2. The multi-feature fusion photovoltaic fault diagnosis method based on the Bagging algorithm according to claim 1, wherein In Step 1, the normal and fault states include normal, short circuit, open circuit, aging, and shadow occlusion. The electrical data includes the current I, voltage V, output power P, and photovoltaic cell temperature T of each photovoltaic string. The non-electrical data includes the operating duration T s , the ambient temperature T h , the ambient humidity H, and the irradiance E.

3. The multi-feature fusion photovoltaic fault diagnosis method based on the Bagging algorithm according to claim 1, wherein In Step 3, the fused feature sample data set is divided into a training set and a test set according to a ratio of 7:

3.

4. The multi-feature fusion photovoltaic fault diagnosis method based on the Bagging algorithm according to claim 1, wherein The fault diagnosis model in Step 3 includes multiple convolutional neural networks, and the multiple convolutional neural networks are aggregated by the Bagging algorithm.

5. The multi-feature fusion photovoltaic fault diagnosis method based on the Bagging algorithm according to claim 4, wherein, The convolutional neural network includes a first basic residual block, a second basic residual block, a first bidirectional LSTM module, a second bidirectional LSTM module, a first conversion module, and a first fully connected layer module, Both the first basic residual block and the second basic residual block include a convolutional layer, a layer normalization processing module, and an activation layer connected in sequence, The first basic residual block and the second basic residual block are respectively connected to the first bidirectional LSTM module and the second bidirectional LSTM module through a reshaping module; The first basic residual block, the second basic residual block, and the first conversion module are connected in sequence, and the first bidirectional LSTM module, the second bidirectional LSTM module, and the first conversion module are connected in sequence; The first fully connected layer is connected to the first conversion module.

6. The multi-feature fusion photovoltaic fault diagnosis method based on the Bagging algorithm according to claim 4, wherein It includes a third basic residual block, a fourth basic residual block, a fifth basic residual block, a channel attention module, a second conversion module, and a second fully connected layer module connected in sequence. The third basic residual block, the fourth basic residual block, and the fifth basic residual block all include a convolutional layer, a layer normalization processing module, and an activation layer connected in sequence.

7. The multi-feature fusion photovoltaic fault diagnosis method based on the Bagging algorithm according to claim 5 or 6, characterized in that The parameters of the convolutional layer are k = 3, c = 64, where k is the size of the convolutional kernel and c is the number of output channels of this convolutional layer; the activation function of the activation layer is the LeakyReLU function.

8. The multi-feature fusion photovoltaic fault diagnosis method based on the Bagging algorithm according to claim 1, wherein, The target threshold P in Step 4 is 90%.

Images