Current transformer error prediction method and system based on hybrid neural network

Abstract

The invention discloses a current transformer error prediction method and system based on a hybrid neural network, and relates to the technical field of current transformer error prediction, and the method comprises the steps: carrying out the feature extraction of a data set through employing SAE, and capturing the complex features of data. The reconstructed data set is fed into a CNN-LSTM model, the CNN extracts the features of the time sequence, and the LSTM finds the dependency relationship of the time; and dispersing different features to different models for extraction, predicting data through a regression device, and outputting an evaluation result. According to the current transformer error prediction method based on the hybrid neural network, a reconstruction data set is reconstructed after an abnormal value is judged, a sparsity penalty mechanism is utilized in an SAE model, learned representation is promoted to have a sparsity characteristic, the generalization ability and interpretability of the model are enhanced, an index average is added, and the error prediction accuracy of the current transformer is improved. The model is smoothed, parameter fluctuation in the model training process is reduced, and the training process is more stable.

Description

Technical Field

[0001] The present invention relates to the technical field of current transformer error prediction, and in particular to a current transformer error prediction method and system based on a hybrid neural network. Background Art

[0002] Current transformer is an indispensable device in the power system. It is used to monitor and measure the current on the transmission line, while ensuring the accuracy of relay protection and power metering. In the power transmission project of the power grid, the current transformer provides the power system with the necessary primary current information. Since the electromagnetic current transformer is affected by various factors such as temperature change, magnetic field interference and vibration during operation, its measurement accuracy and stability may be inaccurate. Therefore, it is of great significance to conduct long-term field measurement research on the equipment to understand the law of its performance change for the safe and stable operation of the power system. Traditional current transformer error evaluation methods mainly rely on hardware calibration and periodic inspections, which are often time-consuming and costly. It is also often difficult to achieve timely and accurate detection of faults and potential risks that occur during field testing, which will cause a series of abnormal problems in power metering. In addition, traditional methods cannot comprehensively and systematically evaluate the accuracy, health status and reliability of the device under operation. In order to accurately capture the operating status of the current transformer and monitor and evaluate the error of the current transformer online through corresponding technical means, it has become a research difficulty in current power grid technology.

[0003] In the study of current transformer metering performance evaluation, the main operating methods are regular power outage offline maintenance, live detection and predictive evaluation; in the online evaluation of current transformer errors, the traditional machine learning method is mainly used to learn the representation of current transformer monitoring data from manually extracted features. Compared with traditional methods, the performance evaluation method based on deep learning has the advantages of safety, efficiency and low cost. Many scholars have added it to the pre-evaluation and calibration method. However, due to the large amount of current transformer data, many interference factors, and difficulty in online evaluation, a single deep learning network model cannot effectively predict the error of the current transformer.

[0004] At present, the error prediction methods of current transformer monitoring data can be mainly divided into two categories, namely online monitoring technology based on traditional machine learning and online monitoring technology based on deep learning models. The idea of ​​online monitoring technology based on traditional machine learning is to improve the prediction performance by designing different feature extraction modes, selecting the optimal feature combination from them, and then constructing effective regressor input features. This type of method has good interpretability, but requires human participation in the feature design process, which is time-consuming, laborious and easily affected by human subjective factors. The idea of ​​online monitoring technology based on deep learning models is to directly input the pre-processed current transformer monitoring data into the neural network model, and automatically learn and optimize the constructed network model in an end-to-end manner, avoiding the tedious feature extraction process. However, the structure of the deep learning network model will have a greater impact on the final prediction results. An effective network architecture can capture more representative features. Therefore, it is necessary to design a reasonable and efficient deep learning network to accurately predict current transformer monitoring data. Summary of the invention

[0005] In view of the above-mentioned problems, the present invention is proposed.

[0006] Therefore, the technical problem solved by the present invention is: the existing online monitoring technology based on traditional machine learning and the online monitoring technology based on deep learning models require human participation, are time-consuming, labor-intensive and easily affected by human subjective factors, and how to avoid the cumbersome feature extraction process.

[0007] In order to solve the above technical problems, the present invention provides the following technical solutions: a method for predicting current transformer errors based on a hybrid neural network, comprising using SAE to extract features from a data set and capture the complex features of the data. The reconstructed data set is fed into a CNN-LSTM model, where CNN extracts the features of the time series and LSTM searches for time dependencies. Different features are distributed to different models for extraction, and the data is predicted by a regressor to output an evaluation result.

[0008] As a preferred solution of the current transformer error prediction method based on hybrid neural network described in the present invention, wherein: the use of SAE to extract features from the data set, capturing the complex features of the data includes the encoder converting the input signal x=[x1, x2, ..., x c ] is mapped to the hidden layer to generate a low-dimensional feature vector h = [h1,h2,...,h d ] to compress the input data.

[0009] The decoder converts the feature vector h back into an output vector y = [y1, y2, ..., y d ], reconstruct the original input signal and restore the input.

[0010] As a preferred solution of the current transformer error prediction method based on hybrid neural network described in the present invention, wherein: the use of SAE to extract features from the data set and capture the complex features of the data also includes adding the KL divergence regularization term as a sparsity constraint to the loss function to achieve the sparsity effect of the AE autoencoder network. Let the sparsity parameters be ρ, α j (x) represents the signal activation degree of the jth neuron in the hidden layer of the autoencoder network. The loss function of SAE is the same as that of the neural network. The goal is to minimize the weighted sum of the reconstruction error and the sparsity penalty. The cost function of the autoencoder is expressed as:

[0011] J(W,b)

[0012] Among them, J(W,b) represents the cost function of the autoencoder, W and b are the weight and bias of the network respectively.

[0013] The cost function after adding the sparse constraint is expressed as:

[0014]

[0015] Among them, β is the weight of sparsity control, S2 represents the number of units in the hidden layer, and ρ represents the target activation rate. represents the average activation rate of the jth hidden unit.

[0016] The KL divergence term is expressed as:

[0017]

[0018] The first treatment When it is close to ρ, the second term deals with keep away The situation at that time.

[0019] As a preferred solution of the current transformer error prediction method based on hybrid neural network described in the present invention, wherein: the feeding of the reconstructed data set into the CNN-LSTM model includes constructing a CNN-LSTM current transformer evaluation model, predicting the error of the current transformer, and characterizing the electromagnetic current transformer error as a model including an output variable and multiple input variables:

[0020]

[0021] Among them, δ represents the ratio error, γ represents the angle difference, x represents the input variable, f1 represents the linear error, and f2 represents the nonlinear error.

[0022] As a preferred solution of the current transformer error prediction method based on hybrid neural network described in the present invention, wherein: the feeding of the reconstructed data set to the CNN-LSTM model also includes combining CNN with LSTM network, CNN consists of five parts: input layer, convolution layer, pooling layer, fully connected layer and output layer, and LSTM consists of input gate, forget gate, output gate and candidate gate.

[0023] As a preferred solution of the current transformer error prediction method based on hybrid neural network described in the present invention, wherein: the data is predicted by a regressor, and the output evaluation result includes adding an attention module to the LSTM output, and by assigning different weights to the data features in the prediction model, the prediction model performs information screening, captures the characteristic information of the data, and correctly captures the long-term interdependence in the sequence. After CNN-LSTM processes and extracts the features of the time series data, the multilayer perceptron makes a final prediction from the features.

[0024] As a preferred solution of the current transformer error prediction method based on hybrid neural network described in the present invention, the current transformer performance evaluation model including SAE-CNN-LSTM adopts mean square error as the loss function, directly measures the average square difference between the model prediction value and the actual value, and introduces L2 regularization technology in the process of calculating the loss function, which is expressed as:

[0025]

[0026] Among them, y t is the real input data, represents the output of the model, λ represents the weight attenuation when the penalty term is too large, ω i represents the model parameter weight, and T represents the detection cycle.

[0027] Another object of the present invention is to provide a current transformer error prediction system based on a hybrid neural network, which can avoid the shortcomings of weak characterization ability of a single deep learning model by constructing a SAE-CNN-LSTM hybrid neural network, enhance the accuracy and stability of current transformer error prediction, and solve the problems that the current online monitoring technology based on traditional machine learning and the online monitoring technology based on deep learning models cannot be integrated together, require human participation or cumbersome feature extraction.

[0028] As a preferred solution of the current transformer error prediction system based on hybrid neural network described in the present invention, it includes: a data processing module, a feature extraction module, and a regression prediction module. The data processing module is used to extract features from the data set using SAE to capture the complex features of the data. The feature extraction module is used to feed the reconstructed data set into the CNN-LSTM model, CNN extracts the features of the time series, and LSTM looks for time dependencies. The regression prediction module is used to distribute different features to different models for extraction, predict the data through the regressor, and output the evaluation results.

[0029] A computer device comprises a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement a step of a current transformer error prediction method based on a hybrid neural network.

[0030] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of a current transformer error prediction method based on a hybrid neural network.

[0031] Beneficial effects of the present invention: The hybrid neural network-based current transformer error prediction method provided by the present invention reconstructs a reconstructed data set after judging the abnormal value, and uses the sparsity penalty mechanism in the SAE model to promote the learned representation to have sparsity characteristics, thereby enhancing the generalization ability and interpretability of the model, adding the exponential mean to smooth the model, reducing the parameter fluctuation during the model training process, making the training process more stable, and significantly improving the learning efficiency. At the same time, it extracts some time features and long-term relationships in the time series, thereby achieving higher accuracy than a single neural network model. The present invention achieves better results in terms of efficiency, accuracy and generalization ability. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0033] Figure 1 An overall flow chart of a current transformer error prediction method based on a hybrid neural network provided in the first embodiment of the present invention.

[0034] Figure 2 A SAE model encoding and decoding block diagram of a current transformer error prediction method based on a hybrid neural network is provided for the first embodiment of the present invention.

[0035] Figure 3 A CNN model framework diagram of a current transformer error prediction method based on a hybrid neural network provided in the first embodiment of the present invention.

[0036] Figure 4 An LSTM model structure diagram of a current transformer error prediction method based on a hybrid neural network provided in the first embodiment of the present invention.

[0037] Figure 5 An attention mechanism module diagram of a current transformer error prediction method based on a hybrid neural network provided in the first embodiment of the present invention.

[0038] Figure 6 This is an example diagram of measured data of an electromagnetic current transformer according to a current transformer error prediction method based on a hybrid neural network provided in the second embodiment of the present invention.

[0039] Figure 7 A flowchart of data set division processing for a current transformer error prediction method based on a hybrid neural network provided in the second embodiment of the present invention.

[0040] Figure 8 A loss function variation curve of the SAE model of a current transformer error prediction method based on a hybrid neural network provided in the second embodiment of the present invention.

[0041] Fig. 9 A ratio difference MSE value diagram obtained by a CNN-LSTM model of a current transformer error prediction method based on a hybrid neural network provided in the second embodiment of the present invention.

[0042] Fig.10 A graph of the angular difference MSE values ​​obtained by a CNN-LSTM model of a current transformer error prediction method based on a hybrid neural network provided in the second embodiment of the present invention. DETAILED DESCRIPTION

[0043] In order to make the above-mentioned purposes, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are described in detail below in conjunction with the drawings of the specification. Obviously, the described embodiments are part of the embodiments of the present invention, but not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary persons in the art without creative work should fall within the scope of protection of the present invention.

[0044] Example 1, reference Figure 1-Figure 5 , is an embodiment of the present invention, and provides a current transformer error prediction method based on a hybrid neural network, comprising:

[0045] S1: Use SAE to extract features from the dataset and capture the complex features of the data.

[0046] Furthermore, SAE is used to extract features from the dataset, capturing the complex features of the data, including the encoder converting the input signal x = [x1, x2, ..., x c ] is mapped to the hidden layer to generate a low-dimensional feature vector h = [h1,h2,...,h d ] to compress the input data.

[0047] The decoder converts the feature vector h back into an output vector y = [y1, y2, ..., y d ], reconstruct the original input signal and restore the input.

[0048] It should be noted that the number of neurons in the input layer and the output layer of the SAE model is exactly the same, so in each round of training, the network continuously updates the weights and biases to make the input data and the reconstructed data as consistent as possible, minimizing the information loss in the feature extraction process. In order to enhance the ability of the autoencoder (AE) network to mine deep features, the present invention introduces a sparsity constraint in the network, activating only a small number of neurons to improve the sparsity of the extracted features.

[0049] It should also be noted that using SAE to extract features from the data set and capture the complex features of the data also includes adding the KL divergence regularization term as a sparsity constraint to the loss function to achieve the sparsity effect of the AE autoencoder network. Let the sparsity parameters be ρ, α j (x) represents the signal activation degree of the jth neuron in the hidden layer of the autoencoder network. The loss function of SAE is the same as that of the neural network. The goal is to minimize the weighted sum of the reconstruction error and the sparsity penalty. The cost function of the autoencoder is expressed as:

[0050] J(W,b)

[0051] Among them, J(W,b) represents the cost function of the autoencoder, W and b are the weight and bias of the network respectively.

[0052] The cost function after adding the sparse constraint is expressed as:

[0053]

[0054] Among them, β is the weight of sparsity control, S2 represents the number of units in the hidden layer, ρ represents the target activation rate, and ρ j represents the average activation rate of the jth hidden unit.

[0055] The KL divergence term is expressed as:

[0056]

[0057] The first treatment When it is close to ρ, the second term deals with The situation when far away from ρ.

[0058] S2: Feed the reconstructed dataset into the CNN-LSTM model. CNN extracts the features of the time series and LSTM finds the time dependencies.

[0059] Furthermore, feeding the reconstructed dataset into the CNN-LSTM model includes building a CNN-LSTM current transformer evaluation model to predict the error of the current transformer and characterize the electromagnetic current transformer error as a model containing one output variable and multiple input variables:

[0060]

[0061] Among them, δ represents the ratio error, γ represents the angle difference, x represents the input variable, f1 represents the linear error, and f2 represents the nonlinear error.

[0062] It should be noted that the CNN and LSTM networks are combined to construct the CNN-LSTM deep learning network model by combining the advantages of the two models. Among them, CNN is mainly composed of five parts: input layer, convolution layer, pooling layer, fully connected layer and output layer. Figure 3 The CNN network framework constructed in the present invention is shown in FIG. LSTM can effectively process the long-term dependency of time series and is composed of three gate control units. Figure 4 As shown in the figure, the LSTM structure includes 4 boxes, which represent the input gate, forget gate, output gate and a candidate gate (state unit).

[0063] The forget gate is used to control the influence of the state of the previous time step on the current time step, which can be expressed as:

[0064] n t =σ(W n ·[h t-1 ,x t ]+b n )

[0065] The function of the input gate is to determine the influence of the current transformer input data of the current time step on the state unit and output, which can be expressed as the following formula:

[0066] l t =σ(W l ·[h t-1 ,x t ]+b l )

[0067] rt =tanh(W r ·[h t-1 ,x t ]+b r )

[0068] c t =n t *c t-1 +l t *r t

[0069] The output gate controls the effect of the state unit on the output. The state unit is used to store past information and can add or delete information when needed.

[0070] m t =σ(W m ·[h t-1 ,x t ]+b m )

[0071] h t =m t *tanh(c t )

[0072] Among them, x t is the current transformer monitoring data at time t, σ(·) represents the Sigmoid activation function, which is used to map the output between 0 and 1, where 0 represents the loss of all information and 1 represents the retention of all information. W n , W l , W r , W m represents the weight matrix. b n , b l , b r , b m is a parameter. tanh is the hyperbolic tangent activation function, n t is the output of the forget gate, l t is the output of the input gate, r t is the candidate status information, c t is the value of the memory unit at time t, m t is the output of the output gate, h t Output for unit memory.

[0073] Furthermore, due to the huge amount of transformer data, when processing data, we face the phenomenon of forgetting as the length of the time series increases. As time goes by, LSTM may forget the previous key information, which will affect the complete capture of data features and thus affect the accuracy and efficiency of prediction. In order to solve this problem, the attention mechanism is introduced, such as Figure 5 As shown, the model can focus more on the important parts of the sequence.

[0074] S3: Distribute different features to different models for extraction, predict the data through the regressor, and output the evaluation results.

[0075] Furthermore, the data is predicted through the regressor, and the output evaluation results include adding an attention module to the LSTM output. By assigning different weights to the data features in the prediction model, the prediction model performs information screening, captures the characteristic information of the data, and correctly captures the long-term interdependencies in the sequence. After CNN-LSTM processes and extracts the features of the time series data, the multi-layer perceptron makes the final prediction from the features.

[0076] It should be noted that the current transformer performance evaluation model including SAE-CNN-LSTM uses mean square error as the loss function, which directly measures the average square difference between the model prediction value and the actual value. The L2 regularization technology is introduced in the process of calculating the loss function, which is expressed as:

[0077]

[0078] Among them, y t is the real input data, represents the output of the model, λ represents the weight attenuation when the penalty term is too large, ω i Represents the model parameter weight, and T represents the detection cycle. By adding a penalty term proportional to the sum of the squares of the model weight parameters to the loss function, L2 regularization can effectively limit the amplitude of the model parameters, help prevent the model from overfitting on the training data, and enhance the generalization ability of the model.

[0079] Example 2, reference Figure 6-Figure 10 , which is an embodiment of the present invention, provides a current transformer error prediction method based on a hybrid neural network. In order to verify the beneficial effects of the present invention, scientific demonstration is carried out through economic benefit calculation and simulation experiments.

[0080] First, we use a long-term measured data set collected from the Wuhan High Voltage Research Institute for electromagnetic current transformers. Researchers at the Wuhan High Voltage Research Institute record a set of data every half an hour, which provides a rich data foundation for model construction and training in this invention. Figure 6 The figure shows the data changes of one column of the electromagnetic current transformer measured data.

[0081] In order to more effectively train, evaluate and tune the hybrid neural network model, the present invention adopts a data set segmentation strategy, dividing the overall data set into three parts: training set, validation set and test set in a ratio of 8:1:1. Among them, the training set is mainly used in the model training process to fit the rules in the data as much as possible. The validation set is used to monitor the performance of the model and help adjust the hyperparameters of the model. The test set is used after the model training is completed to evaluate the final performance of the model. Figure 7 Shown is a flow chart of the data set partitioning process.

[0082] There are many parameters that affect the hybrid neural network model, including learning rate, batch size, number of training times, and learner selection. Through continuous experiments, the mean of MSE is compared to find the most suitable parameter value. Table 1 shows the optimal MSE parameter combination settings obtained by the model through experiments.

[0083] Table 1 Model parameter settings

[0084] Parameter name SAE-CNN-LSTM Model Training batch 950 Verify batch 100 Learning Rate 0.001 Iterations 1000 Learner Adam LSTM Layer 2

[0085] In order to evaluate the effectiveness of the SAE-CNN-LSTM hybrid neural network, the present invention sets up an ablation experiment, as shown in Table 2, to compare the current transformer prediction performance of the SAE-CNN model, the CNN model and the SAE-LSTM model. In order to ensure the fairness of the comparative model evaluation, the present invention uses the average value of 10 MSEs to evaluate the performance of the model with the same parameters and training cycles in the ablation experiment.

[0086] Table 2 Hybrid neural network ablation experiment

[0087]

[0088]

[0089] It can be seen from Table 2 that the average MSE of the angle difference and the average MSE of the ratio difference of the SAE-CNN-LSTM model are 0.1605 and 0.1336, respectively, which are the smallest compared with other combinations. Compared with the single CNN model, the MSE value of SAE-CNN-LSTM is 1.8258 smaller and the MAE value is 0.6595 smaller, showing the good prediction ability of the hybrid neural network. Compared with the independent CNN or LSTM model, the hybrid model can more accurately evaluate the performance of the current transformer. In the comparison between SAE-CNN and CNN models, the advantage of using SAE for data feature extraction is obvious. SAE can learn the high-order abstract features of the data, making the extracted features more representative and discriminative, thereby providing more valuable information for the model. Therefore, compared with the single CNN model, the SAE-CNN model can achieve a more significant overall accuracy improvement in the performance evaluation of the current transformer.

[0090] The loss function change curve of the SAE model can be used as an important indicator to evaluate the quality of SAE model training. Figure 8 The figure shows the trend of the loss function curve of the SAE model. The loss curve of the SAE model is divided into three parts: the total loss curve of the model, the sparse loss curve and the reconstruction loss curve. Among them, the total loss of the model is the result of the superposition of the sparse loss curve and the reconstruction loss curve. The total loss curve is to reflect the weighted sum of the minimum reconstruction error and sparsity penalty in the SAE model, and comprehensively characterize the quality of model training. Reconstruction can assist the SAE model to reconstruct the input data while maintaining the sparsity of the learning features. Through observation, it is found that the loss function curve of the SAE model does not show underfitting and overfitting. Although there is some noise interference from the data set, the effect of SAE model training is still good, which builds better data input for the subsequent evaluation of the CNN-LSTM model.

[0091] like Fig. 9 and Fig.10 The figure shows the MSE values ​​of the ratio difference and angle difference obtained by the CNN-LSTM model. The MSE value of the ratio difference is 0.1237, and the MSE value of the angle difference is 0.1593. For the model, the smaller the MSE index value, the smaller the error and the more accurate the result. The MSE values ​​of the ratio difference and angle difference reflect that the SAE-CNN-LSTM model has good performance for current transformers.

[0092] Embodiment 3, an embodiment of the present invention, provides a current transformer error prediction system based on a hybrid neural network, including a data processing module, a feature extraction module, and a regression prediction module.

[0093] The data processing module is used to extract features from the data set using SAE to capture the complex features of the data. The feature extraction module is used to feed the reconstructed data set into the CNN-LSTM model. CNN extracts the features of the time series, and LSTM looks for time dependencies. The regression prediction module is used to distribute different features to different models for extraction, predict the data through the regressor, and output the evaluation results.

[0094] If the function is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions for a computer device (which can be a personal computer, server, or network device, etc.) to perform all or part of the steps of the methods of each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk, etc. Various media that can store program codes.

[0095] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as an ordered list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by an instruction execution system, device or apparatus (such as a computer-based system, a system including a processor, or other system that can fetch instructions from an instruction execution system, device or apparatus and execute instructions), or in conjunction with such instruction execution systems, devices or apparatuses. For the purposes of this specification, "computer-readable medium" can be any device that can contain, store, communicate, propagate or transmit a program for use by an instruction execution system, device or apparatus, or in conjunction with such instruction execution systems, devices or apparatuses.

[0096] More specific examples of computer-readable media (a non-exhaustive list) include the following: an electrical connection with one or more wires (electronic device), a portable computer disk case (magnetic device), a random access memory (RAM), a read-only memory (ROM), an erasable and programmable read-only memory (EPROM or flash memory), an optical fiber device, and a portable compact disk read-only memory (CDROM). In addition, the computer-readable medium may even be a paper or other suitable medium on which the program is printed, since the program may be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, deciphering, or processing in another suitable manner as necessary, and then stored in a computer memory.

[0097] It should be understood that the various parts of the present invention can be implemented by hardware, software, firmware or a combination thereof. In the above-mentioned embodiments, multiple steps or methods can be implemented by software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented by hardware, as in another embodiment, it can be implemented by any one of the following technologies known in the art or their combination: a discrete logic circuit having a logic gate circuit for implementing a logic function for a data signal, a dedicated integrated circuit having a suitable combination of logic gate circuits, a programmable gate array (PGA), a field programmable gate array (FPGA), etc. It should be noted that the above embodiments are only used to illustrate the technical solution of the present invention and are not limited. Although the present invention is described in detail with reference to the preferred embodiments, it should be understood by those skilled in the art that the technical solution of the present invention can be modified or replaced by equivalents without departing from the spirit and scope of the technical solution of the present invention, which should be included in the scope of the claims of the present invention.

[0098] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A current transformer error prediction method based on hybrid neural network, It is characterized in that include: Use SAE to extract features from the data set and capture the complex features of the data; The reconstructed dataset is fed into the CNN-LSTM model. CNN extracts the features of the time series and LSTM finds the time dependency. Different features are distributed to different models for extraction, the data is predicted through the regressor, and the evaluation results are output.

2. The current transformer error prediction method based on hybrid neural network as claimed in claim 1, Features: The use of SAE to extract features from a data set captures complex features of the data, including the encoder converting the input signal x=[x 1 ,x 2 ,…,x c ] is mapped to the hidden layer to generate a low-dimensional feature vector h = [h 1 ,h 2 ,…,h d ], to achieve compression of input data; The decoder converts the feature vector h back into an output vector y = [y 1 ,y 2 ,...,y d ], reconstruct the original input signal and restore the input.

3. The current transformer error prediction method based on hybrid neural network as claimed in claim 2, Features: The use of SAE to extract features from the data set and capture the complex features of the data also includes adding the KL divergence regularization term as a sparsity constraint to the loss function to achieve the sparsity effect of the AE autoencoder network. Let the sparsity parameters be ρ, α j (x) represents the signal activation degree of the jth neuron in the hidden layer of the autoencoder network. The loss function of SAE is the same as that of the neural network. The goal is to minimize the weighted sum of the reconstruction error and the sparsity penalty. The cost function of the autoencoder is expressed as: J(W,b) Where J(W,b) represents the cost function of the autoencoder, W and b are the weight and bias of the network respectively; The cost function after adding the sparse constraint is expressed as: Among them, β is the weight of sparsity control, S 2 represents the number of units in the hidden layer, ρ represents the target activation rate, represents the average activation rate of the jth hidden unit; The KL divergence term is expressed as: The first treatment When it is close to ρ, the second term deals with The situation when far away from ρ.

4. The current transformer error prediction method based on hybrid neural network as claimed in claim 3, Features: Feeding the reconstructed data set into the CNN-LSTM model includes constructing a CNN-LSTM current transformer evaluation model, predicting the error of the current transformer, and characterizing the electromagnetic current transformer error as a model including an output variable and multiple input variables: Among them, δ represents the ratio difference, γ represents the angle difference, x represents the input variable, and f 1 represents the linear error, f 2 Represents nonlinear error.

5. The current transformer error prediction method based on hybrid neural network as claimed in claim 4, Features: Feeding the reconstructed data set to the CNN-LSTM model also includes combining the CNN with the LSTM network. The CNN consists of five parts: an input layer, a convolutional layer, a pooling layer, a fully connected layer, and an output layer. The LSTM consists of an input gate, a forget gate, an output gate, and a candidate gate.

6. The current transformer error prediction method based on hybrid neural network as claimed in claim 5, Features: The data is predicted by the regressor, and the output evaluation result includes adding an attention module to the LSTM output. By assigning different weights to the data features in the prediction model, the prediction model performs information screening, captures the feature information of the data, and correctly captures the long-term interdependencies in the sequence. After CNN-LSTM processes and extracts the features of the time series data, the multilayer perceptron makes the final prediction from the features.

7. The current transformer error prediction method based on hybrid neural network as claimed in claim 6, Features: The current transformer performance evaluation model including SAE-CNN-LSTM adopts mean square error as the loss function, directly measuring the average square difference between the model prediction value and the actual value, and introduces L in the process of calculating the loss function. 2 Regularization technique, expressed as: Among them, y t is the real input data, represents the output of the model, λ represents the weight attenuation when the penalty term is too large, ω i represents the model parameter weight, and T represents the detection cycle.

8. A system using the current transformer error prediction method based on a hybrid neural network as claimed in any one of claims 1 to 7, Features: Including data processing module, feature extraction module, regression prediction module; The data processing module is used to extract features from the data set using SAE to capture complex features of the data; The feature extraction module is used to feed the reconstructed data set into the CNN-LSTM model, where CNN extracts the features of the time series and LSTM finds the time dependency; The regression prediction module is used to distribute different features to different models for extraction, predict the data through the regressor, and output the evaluation results.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program. It is characterized in that When the processor executes the computer program, the steps of the current transformer error prediction method based on a hybrid neural network as described in any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, It is characterized in that When the computer program is executed by a processor, the steps of the current transformer error prediction method based on a hybrid neural network according to any one of claims 1 to 7 are implemented.

Images