Transformer excitation surge current identification method based on RBF neural network

Abstract

The invention belongs to the technical field of transformer protection, aims to solve the problem that an existing recognition method is poor in performance under the conditions of complex current waveform processing and non-ideal conditions, and discloses a transformer excitation surge current recognition method based on an RBF neural network. The method comprises the following steps: carrying out denoising and standardization preprocessing on excitation surge current and secondary side short-circuit current signals sampled by a sensor, mapping input data to a high-dimensional space by adopting a radial basis function, capturing nonlinear characteristics by calculating the distance between the input and the center of the basis function, generating a dichotomy result by weighting and fusing the nonlinear characteristics of a hidden layer, and carrying out classification on the dichotomy result. And the excitation surge current or the secondary side short-circuit current is discriminated, the RBF neural network is trained by adopting an orthogonal least square method, and weight calculation is optimized by screening an effective sample path. According to the method, efficient nonlinear feature extraction and accurate inrush current detection can be realized, the features of the excitation inrush current can be effectively identified from complex current signals, and the situations of false alarm and missing alarm are reduced.

Description

Technical Field

[0001] This application belongs to the technical field of transformer protection, and particularly relates to a method for identifying transformer inrush current based on RBF neural network. Background Technique

[0002] At present, most of the existing transformer inrush current identification technologies adopt the second-harmonic braking principle, the discontinuous angle principle, the waveform symmetry principle, etc. However, these methods perform poorly in dealing with complex current waveforms and non-ideal conditions.

[0003] In the prior art, the technical steps of using the second-harmonic braking principle include: collecting the current signal of the power system and performing spectrum analysis on it to extract the fundamental wave and second-harmonic components; analyzing the amplitude and phase characteristics of these frequency components, because inrush current will cause obvious changes in the second harmonic; using signal processing technology to monitor the changes of the second harmonic in real time and identify its characteristic patterns; comparing these characteristics with the preset inrush current model to determine whether inrush current has occurred in the system. Its existing technical defect is that this method mainly relies on identifying the characteristics of the second harmonic, but when the transformer load changes or there are other nonlinear factors in the system, the second harmonic may be interfered, resulting in inaccurate identification; in addition, the response time of the second harmonic to inrush current is relatively long, which may cause misoperation or delay.

[0004] In the prior art, the technical steps of using the discontinuous angle principle include: obtaining the current or voltage waveform of the power system and performing Fourier transform to extract the fundamental frequency and harmonic components; calculating the phase angle of each cycle in the waveform and finding the phase angle difference between adjacent cycles; analyzing these phase angle differences, especially observing the sudden change of the discontinuous angle, because inrush current will cause significant changes in the current or voltage waveform; judging whether inrush current has occurred in the system according to the change characteristics of the discontinuous angle. Its existing technical defect is that this method identifies inrush current by measuring the discontinuous angle of the current waveform, but in the case of complex load or internal transformer fault, the waveform discontinuity may not be obvious, thus affecting the accuracy of identification; the measurement of the discontinuous angle may also be affected by the current waveform noise, resulting in misjudgment.

[0005] In the prior art, the technical steps using the waveform symmetry principle include: collecting current or voltage waveform data and performing periodic analysis on it; comparing the symmetry of the waveforms, especially the symmetry between the positive and negative half-cycles of the waveforms; observing any significant symmetry disruptions, as inrush current will cause waveform asymmetry or distortion; analyzing the characteristics of these symmetry changes and comparing them with the normal waveform pattern to identify whether inrush current has occurred. The existing technical defects are as follows: using waveform symmetry to judge inrush current, but in practical applications, waveform symmetry may be disrupted due to load imbalance or system interference; this method has poor adaptability to asymmetric waveforms or non-standard fault conditions, which may lead to a decrease in recognition accuracy. Summary of the Invention

[0006] The main object of the present application is to provide a method for identifying transformer inrush current based on RBF neural network, to solve the problem that the existing identification methods perform poorly in dealing with complex current waveforms and non-ideal situations, and to accurately identify the inrush current generated by the transformer and the secondary short-circuit current.

[0007] Another object of the present application is to provide a method for identifying transformer inrush current based on RBF neural network, to solve the problem that differential protection malfunctions due to inrush current generated by three-phase transformers under certain specific operating conditions.

[0008] To achieve the above objects, the present application provides the following technical solutions:

[0009] A method for identifying transformer inrush current based on RBF neural network includes:

[0010] Step 1: Perform denoising and normalization preprocessing on the inrush current and secondary short-circuit current signals sampled by the sensor, extract the DC component and the 1st - 5th harmonic characteristic quantities through Fourier series expansion, eliminate the modulus difference by normalization, then remove the fixed fundamental wave component, and retain 5 dynamic characteristics as the input layer nodes;

[0011] Step 2: Use the radial basis function to map the input data to a high-dimensional space, capture the non-linear characteristics by calculating the distance between the input and the center of the basis function, and dynamically adjust the number of hidden layer nodes in combination with the characteristics of the training data after initially setting the number of hidden layer nodes based on an empirical formula;

[0012] Step 3: Set a single-neuron linear output layer, generate a binary classification result by weighted fusion of the non-linear characteristics of the hidden layer, and set the number of output nodes to 1 to distinguish inrush current or secondary short-circuit current;

[0013] Step 4: Train the RBF neural network using the orthogonal least squares method, optimize the weight calculation by screening the effective sample path, suppress the interference of abnormal data on parameter estimation, and improve the robustness and generalization ability of the model.

[0014] In some embodiments, in step 1, let the characteristic quantity be F = [F0, F1, F2, F3, F4, F5], where F1 is the fundamental wave component. After normalization, we get:

[0015]

[0016] In the formula, P represents the normalized sample, P1 represents the fundamental wave component, P0 represents the DC component, and P2 - P5 represent the higher harmonic components.

[0017] The value of P1 is fixed at 1. After removing the P1 term from the input, the input characteristic terms become: P = [P0, P2, P3, P4, P5].

[0018] In some embodiments, in step 1, the data after denoising and normalization pre - processing is converted into the format for network processing. At the input layer, the data is passed to the central nodes of the RBF neural network, and each central node calculates the response value of the input data according to its radial basis function.

[0019] In some embodiments, in step 1, the Gaussian function is used as the transfer function to implement the space mapping transformation. The expression of the Gaussian function applied to the non - linear hidden layer is as follows:

[0020]

[0021] In the formula, x represents the input vector of the input layer, c represents the center of the neurons in the non - linear hidden layer, and σ represents the center width of the non - linear hidden layer.

[0022] In some embodiments, in step 2, to determine the number of neurons in the non - linear hidden layer, the empirical formula used is as follows:

[0023]

[0024] In the formula, m and n respectively represent the number of input and output nodes, l represents the number of neurons in the hidden layer, and a represents a constant less than 10.

[0025] In some embodiments, in step 2, to determine the number of neurons in the non - linear hidden layer, the empirical formula used is as follows:

[0026]

[0027] In the formula, m and n respectively represent the number of input and output nodes, and l represents the number of neurons in the hidden layer.

[0028] In some embodiments, in step 3, the non - linear features provided by the hidden layer are further processed through the output layer to accurately judge the inrush current and fault current, and decide whether to trigger the protection measures.

[0029] In some embodiments, in step 3, the linear output layer has only one output neuron node.

[0030] In some embodiments, step 3 includes: adopting a single-neuron linear structure in the output layer of the RBF neural network, generating a binary classification result by weighted fusion of the non-linear features of the hidden layer, linearly combining the high-dimensional features mapped by the Gaussian function through the output layer, and accurately discriminating inrush current and secondary side short-circuit current by using the weight coefficients optimized by training. The number of output nodes is set to 1 to match the binary decision requirement.

[0031] In some embodiments, in step 4, the data fitting strategy is improved by the orthogonal least squares method, a dynamic sample screening mechanism and an orthogonal space mapping technique are introduced, and a forward stepwise regression strategy is adopted. In the iterative process, samples with strong correlation with the target variable are preferentially selected based on residual analysis to construct an optimized subset.

[0032] Compared with the prior art, the method for identifying inrush current of a transformer based on an RBF neural network provided by this application has the following beneficial effects:

[0033] This application uses the intelligent control method of the RBF neural network to accurately identify the inrush current and fault current generated by the transformer.

[0034] Based on the technical means of identifying inrush current by the RBF neural network in this application, efficient non-linear feature extraction and accurate inrush current detection are realized. Through the RBF neural network, the features of inrush current can be effectively identified from complex current signals, reducing the situations of false alarms and missed alarms. This method uses its powerful non-linear mapping ability to accurately distinguish inrush current from secondary side short-circuit current, improving the reliability and protection accuracy of the system. Through this application, the protection strategy of the transformer is optimized, ensuring the stable operation of the equipment and the overall safety of the power system.

[0035] Furthermore, for the inrush current generated by the three-phase transformer under certain specific operating conditions, which causes the differential protection to malfunction, this application proposes an identification method based on the RBF neural network to accurately identify the inrush current and internal fault current generated by the transformer, thereby playing a role in protecting the transformer. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] In order to more clearly illustrate the technical solutions of this application, the drawings required for the technical description will be briefly introduced below.

[0037] Figure 1 It is a flowchart of the method for identifying inrush current of a transformer based on an RBF neural network provided by this application;

[0038] Figure 2 It is a schematic diagram of the RBF neural network model provided by this application;

[0039] Figure 3 General flowchart of the RBF neural network provided for this application

[0040] Figure 4 Flowchart of the RBF neural network for identifying inrush current provided for this application Detailed implementation manners

[0041] The following is a further detailed description through specific implementation manners

[0042] As Figures 1 to 4 shown, this application provides a method for identifying inrush current of a transformer based on an RBF neural network, including:

[0043] Step 1: Input layer processing. Denoise and normalize the inrush current and secondary side short-circuit current signals sampled by the sensor, extract the DC component and 1st - 5th harmonic characteristic quantities through Fourier series expansion, eliminate the modulus difference through normalization, then remove the fixed fundamental wave component (P1), and retain 5 dynamic characteristics (P0, P2 - P5) as the input layer nodes

[0044] Denoise and normalize the above current signals, extract the DC component and 1st - 5th harmonic characteristic quantities through Fourier series expansion, remove the fixed fundamental wave component after eliminating the modulus difference through normalization, and retain 5 dynamic characteristics as the input of the RBF neural network; by optimizing the number of input layer nodes and the feature dimension, avoid the saturation problem of the network transfer function, combine normalization processing to improve the consistency of data distribution, effectively enhance the training stability of the neural network and the identification accuracy of inrush current, and provide highly reliable input features for fault diagnosis

[0045] Step 2: Nonlinear hidden layer processing. Use the radial basis function (Gaussian function) to map the input data to a high-dimensional space. After initially setting the number of hidden layer nodes based on an empirical formula, dynamically adjust the number of nodes in combination with the characteristics of the training data, and optimize the model complexity and generalization ability by gradually increasing or decreasing the nodes and verifying the performance

[0046] Use the radial basis function (Gaussian function) to map the input data to a high-dimensional space, capture the nonlinear characteristics by calculating the distance between the input and the center of the basis function, and improve the accuracy and robustness of inrush current pattern identification; the number of hidden layer nodes needs to balance the model complexity and generalization ability - too few will lead to underfitting, and too many will cause overfitting. After initially setting the number of nodes based on an empirical formula, determine the optimal number in combination with the characteristics of the training data through the dynamic adjustment method (such as gradually increasing or decreasing the nodes and verifying the performance), avoid the failure problem of the empirical formula under complex samples, and achieve the coordinated optimization of the network expression ability and calculation efficiency

[0047] Step 3: Linear output layer processing. Set up a single-neuron linear output layer to generate a binary classification result by weighted fusion of the non-linear features of the hidden layer. The number of output nodes is 1 to distinguish inrush current or secondary side short-circuit current.

[0048] In the output layer of the RBF neural network, a single-neuron linear structure is adopted to generate a binary classification result by weighted fusion of the non-linear features of the hidden layer; the output layer linearly combines the high-dimensional features mapped by the Gaussian function, and uses the weight coefficients optimized by training to achieve accurate discrimination of inrush current and secondary side short-circuit current. The number of output nodes is set to 1 to match the binary decision requirement (trigger / non-trigger protection), simplify the model structure while ensuring classification interpretability, reduce the computational complexity by combining linear operations, significantly improve the identification speed and reliability, and provide an efficient decision-making basis for the relay protection device.

[0049] Step 4: Network training. Use the orthogonal least squares (OLS) method to train the RBF neural network, optimize the weight calculation by screening effective sample paths, suppress the interference of abnormal data on parameter estimation, and improve the robustness and generalization ability of the model.

[0050] Use the orthogonal least squares (OLS) method to train the RBF neural network, which is an improvement based on the traditional LS. By the optimal subset selection strategy, it reduces the interference of outliers, improves the fitting accuracy and model stability; OLS abandons the idea of full-data fitting, screens effective sample paths to optimize the weight calculation, suppresses the negative impact of abnormal data on parameter estimation, enhances the training robustness and generalization ability, and ensures the high reliability of the inrush current identification model.

[0051] In Step 1, first, it is necessary to preprocess the inrush current and secondary side short-circuit current signal data sampled by the sensor, including denoising and normalization, to ensure the consistency and effectiveness of the input data. Then, the processed data is converted into a format suitable for network processing. In the input layer, the data is passed to the center nodes of the RBF neural network, and each center node calculates the response value of the input data according to its radial basis function, thereby extracting the feature information related to the inrush current.

[0052] Such as Figure 2As shown, the RBF neural network model. The present invention has been improved on the original basis. In step 4, the least squares method (SL) is changed to the orthogonal least squares method (OSL). Through orthogonalization processing, the multicollinearity between independent variables is eliminated, and the stability and calculation accuracy of parameter estimation are improved. The first layer is the input layer. The input layer consists of signal source nodes. Its main function is to receive and transmit the sample data of the input inrush current and secondary short-circuit current. In the signal input layer, the role of the nodes is to transmit the data information from the input end to the next layer of the network, denoise and standardize the data, and then convert the processed data into a format suitable for network processing. The number of neurons in the input layer is equal to the number of sample elements in each group, set as x. Each neuron acts as a simple distribution unit in this layer, and they directly transmit the received input vector without modification to the hidden layer for further processing and analysis.

[0053] The second layer of the RBF neural network model is the non-linear hidden layer. The Gaussian function is selected as the transfer function in the non-linear hidden layer to achieve spatial mapping transformation. The Gaussian function is chosen because of its superior non-linear mapping ability. It flexibly performs non-linear transformation on the input space by adjusting the center point and width parameters. Such mapping ability enables the hidden layer to capture more complex data relationships, thus significantly enhancing the overall expression ability and modeling accuracy of the neural network, enabling the network to process more complex patterns and data features. The following formula is the expression of the Gaussian function applied to the non-linear hidden layer:

[0054]

[0055] In the formula, x represents the input vector of the input layer; c represents the center of the neurons in the non-linear hidden layer; σ represents the center width of the non-linear hidden layer.

[0056] The third layer of the RBF neural network model is the linear output layer. The linear output layer is responsible for processing the signals from the previous hidden layer. The linear output layer uses a linear function as the activation function to perform a linear weighting operation on the signals output by the neurons in the hidden layer. Specifically, the linear output layer generates the final output result by multiplying the output of the hidden layer by the weight coefficient and adding the bias term. This result reflects the final prediction or decision of the entire neural network and serves as the overall response of the model to the input data.

[0057] In step 1, determine the number of neuron nodes in the input layer. For the transformer inrush current waveform and the secondary short-circuit current waveform, after performing Fourier series expansion, it is found that the frequency components above the 5th harmonic are very small and can be ignored. Therefore, the main focus will be on the DC component and the 1st to 5th harmonic components remaining after Fourier series expansion. These components are called characteristic quantities and serve as the input quantities for the neural network. However, since the modulus values of these characteristic quantities vary greatly among different samples, these characteristic quantities cannot be directly used as the network input. Directly inputting them into the neural network may cause the transfer function of the neural network to enter the saturation region, resulting in too small network weights during the training process and affecting the model performance. Therefore, before inputting into the neural network, these characteristic quantities need to be normalized to ensure that the range of the characteristic quantities is within an appropriate interval, thereby improving the effect and stability of network training.

[0058] Let the characteristic quantity be F = [F0, F1, F2, F3, F4, F5], where F1 is the fundamental wave component. After normalization, we get:

[0059]

[0060] The normalized sample P contains 6 features. In the formula, P1 represents the fundamental wave component, and its value is fixed at 1; P0 represents the DC component, and its numerical range varies from 0 to 2; P2 - P5 represent the higher harmonic components, and the numerical range varies from 0 to 1.

[0061] Since the value of P1 is always 1, which has no effect on the identification of the neural network, the P1 term can be removed from the input. Finally, the effective input feature terms are reduced to 5: P = [P0, P2, P3, P4, P5]. Therefore, the number of input nodes of the neural network is set to 5 to reflect these effective characteristic quantities.

[0062] In step 2, the nonlinear processing of the hidden layer is mainly achieved through the radial basis function (RBF). Each hidden layer node uses the Gaussian function as the activation function to map the input data to a high-dimensional feature space. In this process, the network calculates the response value based on the distance between the input data and the center of each basis function, thereby effectively capturing the nonlinear characteristics of the data. This mapping ability enables the network to identify and distinguish complex inrush current patterns, improving the accuracy and robustness of the identification.

[0063] In step 2, determine the number of neurons in the non-linear hidden layer. The number of nodes in the non-linear hidden layer has a much greater impact on the performance of the RBF neural network than the number of nodes in the input layer and the linear output layer. This is because the number of nodes in the hidden layer directly determines the network's expressive ability and learning ability. If the number of nodes in the hidden layer is set too small, the network cannot fully capture the complex features of the data, resulting in poor learning effects, low training accuracy, and even possible non-convergence. On the contrary, if the number of nodes in the hidden layer is set too large, although the network's expressive ability can be improved, it will also significantly increase the training time and cause overfitting problems, reducing the network's generalization ability and making its performance on new data worse. An empirical formula is used to guide this selection, and the empirical formula is as follows:

[0064]

[0065] Furthermore, the empirical formula adopted is as follows:

[0066]

[0067] In the formula, m and n respectively represent the number of input and output nodes; l represents the number of neurons in the hidden layer; a represents a constant less than 10.

[0068] The empirical formula adopted in this application is mainly used to provide an initial estimate of the number of nodes in the hidden layer. In practical applications, the trial method is combined to determine the optimal number of neurons in the hidden layer. This process starts from the initial value given by the empirical formula and then gradually adjusts the number of neurons in the hidden layer. By continuously increasing or decreasing the number of nodes in the hidden layer and observing the training results and performance of the network until a configuration that not only meets the training accuracy requirements but also keeps the network structure as simple as possible is found. This method can help optimize the network performance and effectively avoid overfitting or underfitting problems.

[0069] In step 3, a single-neuron linear output layer of the RBF neural network is adopted to weight and fuse the non-linear features of the non-linear hidden layer, and output a binary classification result (inrush current / secondary side short-circuit current). Precise discrimination is achieved by training and optimizing the weight coefficients, and a single node is used to match the protection trigger decision requirements, simplifying the structure and improving the calculation efficiency and reliability.

[0070] In the process of identifying inrush current based on the RBF neural network, the main function of the linear output layer is to convert the feature mapping after non-linear processing by the non-linear hidden layer into the final output result. The output layer usually consists of a group of linear neurons, and these neurons sum the outputs of the non-linear hidden layer with weights to obtain a linear combination. This linear combination is combined with the activation values of the hidden layer nodes through weight coefficients (usually obtained through training) to generate the final recognition result.

[0071] Specifically, the linear output layer further processes the non-linear features provided by the non-linear hidden layer to accurately judge the inrush current and the secondary side short-circuit current, and decide whether to trigger protection measures. This linear processing method ensures the interpretability of the network during classification decision-making, and also provides an intuitive adjustment means for the training and optimization of the model.

[0072] In step 3, determine the number of neuron nodes in the linear output layer. Since this application only identifies whether the abnormal current is an inrush current or a secondary side short-circuit current, there is only one output neuron node in the linear output layer.

[0073] In step 4, the RBF neural network is trained using the Orthogonal Least Squares (OLS) method, and the effective sample paths are screened to optimize the weight calculation and suppress the interference of abnormal data.

[0074] The OLS method is developed on the basis of the Least Squares (LS) method. Its core idea is to determine the subset by selecting the optimal path, thereby improving the accuracy of fitting. In the least squares method, all data points are taken into account, but this method is significantly affected by outliers, resulting in inaccurate fitting results because the deviation degree of the outliers will have an adverse impact on the final fitting result. To solve this problem, the OLS method adopts a fitting strategy based on partial data. By selecting data points for fitting during the training process, the OLS method can reduce the negative impact of outliers on model training, thereby obtaining more reliable and accurate fitting results. The advantage of this method is that it can effectively handle the outliers in the data, making the training process more stable and improving the generalization ability of the model.

[0075] The OLS method improves the data fitting strategy. That is, on the basis of the traditional least squares (LS) method, when the data characteristics or model assumptions do not meet the effectiveness and optimality, by constructing a quadratic loss function about the model parameters (that is, minimizing the sum of the squares of the residuals between the predicted value and the actual observed value), and using the extreme value solution method in calculus (such as taking the derivative and setting the derivative to zero), find the parameter estimation value that makes this loss function reach the minimum value, introduce a dynamic sample screening mechanism and an orthogonal space mapping technology, abandon the traditional full-data fitting mode, adopt a forward stepwise regression strategy, and preferentially select samples strongly correlated with the target variable to construct an optimization subset based on residual analysis during the iteration process, so as to effectively isolate the interference of outliers on the weight calculation. In the task of identifying transformer inrush current, this strategy can adaptively ignore the asymmetric waveform distortion data segments caused by mutation situations, and combined with dynamic residual threshold control, ensure that the model suppresses the influence of transient noise on the classification boundary while retaining key discriminant information such as second harmonic characteristics.

[0076] The key point of using the RBF neural network to identify inrush current in this application lies in effectively extracting and processing the features of the input data. First, high-frequency data needs to be collected from the real-time monitoring system of the transformer to accurately capture the instantaneous changes of the inrush current. Secondly, the design of the RBF neural network should focus on selecting appropriate centers and widths of the basis functions, which affects the model's ability to identify the features of the inrush current. In addition, the training data set of the model needs to cover various possible inrush current scenarios and normal operating states to ensure the generalization ability and accuracy of the network.

[0077] In one embodiment, the implementation process of the transformer inrush current identification method based on the RBF neural network provided in this application is as follows:

[0078] (1) Data collection

[0079] Collect current signal data, including data of inrush current and secondary side short-circuit current.

[0080] (2) Data preprocessing

[0081] Denoising: Filter out the noise in the current signal.

[0082] Normalization: Normalize the data to ensure the consistency of the input data.

[0083] (3) Feature extraction

[0084] Extract key features (such as amplitude, phase angle, etc.) from the preprocessed data.

[0085] (4) Input layer

[0086] Input the extracted feature data into the input layer of the RBF neural network.

[0087] (5) Nonlinear hidden layer

[0088] Radial basis function: Each hidden layer node calculates the distance between the input data and the center, and uses the Gaussian function to generate the activation value.

[0089] Nonlinear mapping: Perform nonlinear feature mapping through the Gaussian function.

[0090] (6) Linear output layer

[0091] Linear combination: Weighted sum the activation values of the hidden layer and the weights to obtain the final output.

[0092] Classification decision: Judge the inrush current and secondary side short-circuit current, and output the identification result.

[0093] (7) Model training

[0094] Weight adjustment: Adjust the weights from the hidden layer to the output layer according to the actual results.

[0095] Error minimization: Optimize network parameters using training data to reduce prediction errors.

[0096] (8) Result output

[0097] Identification: Output the identification results of inrush current and secondary short-circuit current, and take corresponding protection measures.

[0098] The above are only specific embodiments of the present application, but the protection scope of the present application is not limited thereto. Any changes or substitutions that can be easily thought of by those skilled in the art within the technical scope disclosed by the present application should be covered within the protection scope of the present application.

Claims

1. A transformer excitation inrush current identification method based on RBF neural network, characterized in that: include: Step 1: De-noise and standardize the excitation inrush current and secondary short-circuit current signals sampled by the sensor, extract the DC component and 1-5 harmonic feature quantities through Fourier series expansion, remove the fixed fundamental component after normalization to eliminate the module value difference, and retain 5 dynamic features as input layer nodes; Step 2: Use radial basis functions to map the input data to a high-dimensional space, and capture nonlinear features by calculating the distance between the input and the center of the basis function. After initially setting the number of hidden layer nodes based on empirical formulas, dynamically adjust the number of nodes based on the characteristics of the training data. Step 3: Set a single neuron linear output layer to generate a binary classification result by weighted fusion of the nonlinear features of the hidden layer, and the number of output nodes is 1 to distinguish the excitation inrush current or fault current; Step 4: Use the orthogonal least squares method to train the RBF neural network, optimize the weight calculation by screening the effective sample path, suppress the interference of abnormal data on parameter estimation, and improve the model robustness and generalization ability.

2. The transformer excitation inrush current identification method based on RBF neural network according to claim 1 is characterized in that: In step 1, let the feature quantity be F=[F0, F1, F2, F3, F4, F5], F1 is the fundamental component, and after normalization, we get: In the formula, P represents the normalized sample, P1 represents the fundamental component, P0 represents the DC component, and P2~P5 represent the higher harmonic components; The value of P1 is fixed to 1. If the P1 item is removed from the input, the input feature item is: P=[P0,P2,P3,P4,P5].

3. The transformer excitation inrush current identification method based on RBF neural network according to claim 1 is characterized in that: In step 1, the denoised and standardized preprocessed data is converted into a format for network processing. In the input layer, the data is passed to the central node of the RBF neural network, and each central node calculates the response value of the input data according to its radial basis function.

4. The transformer excitation inrush current identification method based on RBF neural network according to claim 1 is characterized in that: In step 1, the Gaussian function is used as the transfer function to realize the spatial mapping transformation. The expression of the Gaussian function applied to the nonlinear hidden layer is as follows: Where x represents the input vector of the input layer, c represents the neuron center of the nonlinear hidden layer, and σ represents the center width of the nonlinear hidden layer.

5. The transformer excitation inrush current identification method based on RBF neural network according to claim 1 is characterized in that: In step 2, the number of neuron nodes in the nonlinear hidden layer is determined using the following empirical formula: Where m and n represent the number of input and output nodes respectively, l represents the number of neurons in the hidden layer, and a represents a constant less than 10.

6. The transformer excitation inrush current identification method based on RBF neural network according to claim 1 is characterized in that: In step 2, the number of neuron nodes in the nonlinear hidden layer is determined using the following empirical formula: In the formula, m and n represent the number of input and output nodes respectively, and l represents the number of neurons in the hidden layer.

7. The transformer excitation inrush current identification method based on RBF neural network according to claim 1 is characterized in that: In step 3, the nonlinear features provided by the hidden layer are further processed through the output layer to accurately determine the excitation inrush current and fault current and decide whether to trigger the protection measures.

8. The transformer excitation inrush current identification method based on RBF neural network according to claim 1 is characterized in that: In step 3, the linear output layer has only one output neuron node.

9. The transformer excitation inrush current identification method based on RBF neural network according to claim 1 is characterized in that: Step 3 includes: using a single neuron linear structure in the output layer of the RBF neural network, generating a binary classification result by weighted fusion of the nonlinear features of the hidden layer, linearly combining the high-dimensional features mapped by the Gaussian function through the output layer, and using the weight coefficients optimized by training to achieve accurate discrimination of the excitation inrush current and the secondary short-circuit current. The number of output nodes is set to 1 to match the binary decision-making requirements.

10. The transformer excitation inrush current identification method based on RBF neural network according to claim 1, characterized in that: In step 4, the data fitting strategy is improved by using the orthogonal least squares method, a dynamic sample screening mechanism and orthogonal space mapping technology are introduced, and a forward stepwise regression strategy is adopted. In the iterative process, samples with strong correlation with the target variable are preferentially selected based on residual analysis to construct an optimized subset.

Images