A real-time error compensation method for voltage transformers

Abstract

The application provides a voltage transformer error real-time compensation method, which is mainly applied to the field of power system safety. The method realizes high-precision dynamic compensation of the voltage transformer error by combining a CB module (CNN-BiLSTM bidirectional collaborative model) and an improved grey wolf optimization algorithm (DGWO). The CB module fuses space-time features by using a bidirectional information interaction channel to generate a candidate compensation solution. The DGWO algorithm introduces twin diversity operation and singleton derivation operation to optimize search efficiency and solution quality. Based on feature importance analysis (FIM) and a dynamic weighted preprocessing mechanism of knowledge base feedback, the adaptability and compensation accuracy of the model are further improved. The application solves the problems of poor real-time performance and insufficient precision of traditional methods, and significantly improves the measurement reliability of the power system and the operation safety of the equipment.

Description

Technical Field

[0001] This invention relates to the field of power system safety, and in particular to a real-time error compensation method for voltage transformers based on CB-DGWO. Background Technology

[0002] In power systems, voltage transformers are core equipment for energy metering, and their operational quality is closely linked to economic and social benefits. Due to inherent characteristics such as internal impedance, excitation current, and energy loss, as well as other factors, voltage transformers inevitably exhibit ratio and phase errors between their primary and secondary windings. Therefore, timely and accurate compensation for voltage transformer errors is crucial. However, traditional error compensation techniques have significant drawbacks. Not only is their accuracy insufficient for practical needs, but their response speed is also slow, often resulting in overcompensation. This not only leads to significant deviations in energy metering but can also, in severe cases, damage power equipment and cause unnecessary economic losses.

[0003] Chinese patent application CN202411119949.3, published on September 13, 2024, discloses a method for evaluating the error of a DC voltage transformer. This method extracts the fundamental and harmonic components of the secondary voltage through symplectic geometric mode decomposition, combines this with a CNN model to extract influencing factors (harmonics, temperature) and fundamental characteristics, performs deep feature fusion using an optimized SWGMN model, and finally achieves real-time evaluation of the error state through an RBF network. However, the above technical solution has at least the following problems: 1. The patented method focuses on error assessment of DC voltage transformers and does not cover AC voltage transformers or other types of transformers, thus its scope of application is limited; 2. Error calculation is based on the equivalent circuit of a resistive-capacitive DC voltage divider. If the actual transformer structure differs significantly, the accuracy of the model may decrease. 3. The core of the patent is error assessment rather than compensation, and it does not propose specific error correction measures, so its practical value is relatively limited. Summary of the Invention

[0004] This invention provides a real-time error compensation method for voltage transformers, involving CNN (Convolutional Neural Network) and BiLSTM (Bidirectional Long Short-Term Memory Network). The aim is to achieve accurate and rapid compensation of voltage transformer errors, ensuring the accuracy of power metering, the reliable operation of the power system, and the safe use of power equipment. It aims to solve the shortcomings of traditional voltage transformer error compensation methods in terms of real-time performance, accuracy, and adaptability.

[0005] To achieve the above objectives, the present invention provides a real-time error compensation method for voltage transformers, comprising the following technical solutions: The knowledge base receives real-time weight vectors, retrieves historical weight vectors from historical scenarios, and feeds the two types of vectors back to dynamic weighted preprocessing to obtain a fused weight vector. The CB module establishes a bidirectional information exchange channel between the CNN model and the BiLSTM model. The input time series data will be processed by the CNN model and the BiLSTM model, and a candidate solution set for the voltage transformer error compensation operation will be obtained by linear mapping. The improved Grey Wolf algorithm introduces enrichment operations, including twin diversity operations and singleton derivation operations, on the basis of the traditional Grey Wolf algorithm to process the candidate solution set and output the optimal compensation turn value.

[0006] Preferably, the time series data includes the primary voltage V1, secondary voltage V2, secondary current I, temperature T, error compensation turns ΔV, and effect record P (represented as a score of 0-10) of the voltage transformer within a certain past time period and at the current time. These data are arranged in chronological order and a suitable time step is determined, dividing them into continuous time windows. The data in each window serves as the input data for the CB module.

[0007] Preferably, the knowledge base provides preset parameters to the CB module when the system is first run, which helps the CB module know which features (such as temperature and voltage) are more important when there is no historical data, thereby generating candidate solutions more effectively.

[0008] Preferably, the feature importance analysis quantifies the contribution weight of each feature in real time, performs utility value gradient calculation, and generates a real-time weight vector.

[0009] Preferably, the bidirectional information interaction channel helps establish a bidirectional data flow mechanism between the CNN model and the BiLSTM model, enabling deep fusion of the CNN model and the BiLSTM model, as well as the collaborative optimization of spatiotemporal features, and enhancing the joint representation capability of the model.

[0010] Preferably, the bidirectional information interaction channel, from the CNN model to the BiLSTM model, introduces spatial features extracted by the CNN model into the input gate, forget gate, and cell state calculation of the BiLSTM model, and combines them with spatial features such as voltage change rate and current fluctuation to help the BiLSTM model extract temporal features and more accurately determine which temporal information needs to be retained or forgotten.

[0011] Preferably, the bidirectional information interaction channel, from the BiLSTM model to the CNN model, introduces the historical hidden state of the BiLSTM model into the convolution operation of the CNN model, so that the CNN model can perceive the temporal dependence when extracting local spatial features, thus enabling the spatial feature extraction to have temporal awareness capability, thereby extracting local spatial features.

[0012] Preferably, the linear mapping method maps the spatiotemporal features extracted by the fusion of the CNN model and the BiLSTM model into a set of candidate compensation turn values ​​through a linear transformation, which are then used by the improved Grey Wolf optimization algorithm to select the optimal solution.

[0013] Preferably, the improved Grey Wolf algorithm uses the twin diversity operation to generate new solutions with higher diversity from high-quality candidate solutions, expand the search range, discover potential better regions, and generate new solutions more frequently when the solution set diversity is high.

[0014] Preferably, the singleton derivation operation used in the improved Grey Wolf algorithm randomly changes certain characteristics of some candidate solutions, locally perturbs a single candidate solution, introduces randomness to escape local optima, and combines with the twin diversity operation to find the optimal number of turns compensation value.

[0015] Compared with the prior art, the beneficial effects achieved by the present invention are as follows: This invention provides a real-time error compensation method for voltage transformers. Combining a CB module with an improved Grey Wolf algorithm, it achieves real-time dynamic compensation for voltage transformer errors, improving the response speed and accuracy of compensation. The CB module uses a CNN model to extract spatial features (such as voltage change rate and current fluctuation characteristics) and a BiLSTM model to extract temporal features. A bidirectional interactive channel enables deep fusion of spatiotemporal features, thus more comprehensively capturing the complex correlations of power data. The improved Grey Wolf algorithm introduces twin diversity operations and singleton derivation operations on the basis of the traditional Grey Wolf algorithm. Its adaptive weighting factors (such as ρ, η, ξ, etc.) and probabilistic mechanisms (such as P...)... c P m The method can dynamically adjust based on the diversity of the solution set and the iteration progress, further optimizing the search efficiency and solution quality. Simultaneously, based on feature importance analysis and a knowledge base, it completes dynamic weighted preprocessing of the data, adjusts the feature extraction strategy of the CB module, and realizes the interaction mechanism between the CB module and the DGWO algorithm. In summary, the real-time error compensation method for voltage transformers provided by this invention improves the overall safety and operational reliability of the power system. Attached Figure Description

[0016] The accompanying drawings, which are included to provide a further understanding of this disclosure and form part of this application, illustrate exemplary embodiments of this disclosure and are used to explain this disclosure, but do not constitute an undue limitation of this disclosure.

[0017] Figure 1 This is a flowchart of a real-time error compensation method for voltage transformers provided in Embodiment 1 of the present invention.

[0018] Figure 2 This is a flowchart of the CB module provided in Embodiment 2 of the present invention.

[0019] Figure 3 This is a flowchart of the DGWO algorithm provided in Embodiment 3 of the present invention. Detailed Implementation

[0020] To make the objectives, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0021] Example 1

[0022] like Figure 1 As shown, this invention proposes a real-time error compensation method for voltage transformers, the method comprising:

[0023] During the data preprocessing stage, multi-dimensional real-time data generated by the voltage transformer during operation are collected, including primary side voltage (V1), secondary side voltage (V2), secondary side current (I), temperature (T), historical error compensation turns (ΔV), and compensation effect score (P). The compensation effect score quantifies the actual effect of each compensation with a score of 0-10 and is used for model optimization. In addition, the historical error compensation turns and the compensation effect score are initialized to 0.

[0024] Based on the sampling frequency of the power system, the collected multi-dimensional real-time data is arranged in chronological order and divided into fixed-length time windows (e.g., each window contains W time steps) to capture the dynamic trends of the data; the input data vector X for the t-th time window is... t It can be represented as: X t =[V 1,t-w+1 V 2,t-w+1 ,I t-w+1 ,T t-w+1 ,ΔV t-w+1 ,P t-w+1 ] Where t represents the current time and W is the window size, ensuring that the data has temporal continuity.

[0025] The input data vector X for each time window tIt needs to be standardized, using the following formula: Where, μ i σ i All of these are historical statistics.

[0026] The CB module is the core computing unit of this invention. It combines the spatial feature extraction capability of the Convolutional Neural Network (CNN model) with the temporal modeling capability of the Bidirectional Long Short-Term Memory Network (BiLSTM model). It establishes a bidirectional data flow mechanism between the CNN model and the BiLSTM model, realizing the deep fusion of the CNN model and the collaborative optimization of spatiotemporal features. When the CNN model extracts data such as voltage change rate, current fluctuation characteristics, voltage-current coupling characteristics, temperature-other parameter correlation characteristics, and compensation turn quantification characteristics, the temporal information of the BiLSTM model layer is introduced to assist convolution on the basis of conventional convolution operations. This completes the deep fusion of spatial and temporal features, extracts key information from the preprocessed time window data, and generates high-precision error compensation candidate solutions.

[0027] When the system is run for the first time, the knowledge base provides key parameters to the CB module through a preset parameter initialization mechanism, enabling the CB module to effectively generate candidate solutions and ensuring the rationality and security of the generated candidate solutions.

[0028] The initialization parameters stored in the knowledge base typically include the basic weight vector W. base The initial configuration of the convolutional kernel, the size of the hidden layer in the BiLSTM model, the range of candidate solutions generated, the activation function parameters, and the initial weighting formula are as follows: X weighted =X norm ⊙W base +∈,∈~N(0,0.01 2 )

[0029] When the system is not running for the first time, the CB module combines the real-time weights from the FIM with the historical weights from the knowledge base to obtain a fused weight vector, using the following fusion formula: W final =α*W real +(1-α)W history α=σ(10SNR) Where SNR is the current signal-to-noise ratio, and σ is the Sigmoid function.

[0030] The resulting fusion weight vector will be further constrained by the knowledge base's security policy. When weight allocation is abnormal, the knowledge base will forcibly truncate it to a safe range to ensure the safe operation of the device. W final[j]←chip{W final [j],W min [j],W max [j]} Among them, W min [j] represents the minimum safety weight: the lower bound of the weight of the j-th feature as specified by the knowledge base; W max [j] represents the maximum safety weight: the upper limit of the weight of the j-th feature as specified by the knowledge base; W final [j] represents the final fusion weight: the weight coefficient actually used by the j-th feature in the dynamic weighted preprocessing.

[0031] Based on the time series data and the final fusion weights, dynamic weight allocation is completed using the following formula: X weighted =X norm ⊙W final +∩,∈~N(0,0.01 2 ) Where ⊙ represents element-wise multiplication, ∈ represents anti-overfitting noise, and X weighted A matrix is ​​an optimized data matrix that strengthens key features and weakens secondary features.

[0032] For example, the operation of a 110kV voltage transformer at a high temperature (65 degrees Celsius) follows these steps: S1: Input data: X t =[226.5,108.2,14.8,63.1,0,0], (V1,V2,I,T,ΔV,P); S2: Weight Calculation FIM Real-Time Weights: W real = [0.3, 0.2, 0.1, 0.4, 0, 0]; Knowledge base historical weight: W history = [0.2, 0.1, 0.05, 0.65, 0, 0]; Safety constraints: W min =[0.2,0.1,0.0,0.1], W max = [0.9, 0.5, 0.3, 0.5]; S3: Fusion Result: S31: Let α = 0.7; S32: W final =0.7×[0.3,...]+0.3×[0.2,...]=[0.27,0.17,0.0715,0.475,0,0]; S33: Since 0.1 <= 0.475 <= 0.5, no truncation is required; S34: Preprocessing output, yielding dynamically weighted preprocessed time series data: W final = [226.5×0.27, 108.2×0.17, ..., 63.1×0.475]

[0033] Finally, the dynamically weighted preprocessed time series data serves as the input to the CB module. Its structure combines temporal and multidimensional features, strengthening key features and weakening secondary features. This effectively supports the collaborative processing of CNN and BiLSTM models in subsequent steps, laying the data foundation for real-time error compensation calculation.

[0034] The dynamically weighted time series data processed by the CB module will output multiple candidate solutions, which will be used by the subsequent DGWO algorithm for utility value analysis and output of the optimal solution.

[0035] The DGWO algorithm is the core optimization module of this invention for achieving real-time and accurate compensation of voltage transformer errors. Based on traditional gray wolf optimization, this algorithm innovatively introduces twin diversity operations and singleton derivation operations, significantly improving global search capability and convergence speed.

[0036] On the one hand, the DWGO algorithm processes the candidate solution set and outputs the optimal solution.

[0037] On the other hand, utility values ​​and candidate solution sets can be used as inputs for feature importance analysis to generate real-time weight vectors.

[0038] Specifically, Feature Importance Analysis (FIM) is the core module in the CB-DGWO system that quantifies the contribution weights of each feature in real time. Through real-time gradient analysis, it transforms the physical characteristics of power data into a computable weight vector, guiding system resource allocation and error compensation strategies. The specific implementation is as follows:

[0039] Prepare candidate solution set Each candidate solution L i =[L i,1 ,L i,2 ,L i,3 ,...,L i,d [] represents a set of compensation schemes, with the utility value vector being... Utility value f i The smaller the value, the more likely it is to be a candidate solution L. i The better.

[0040] The gradient of utility values ​​is calculated using numerical differentiation, and the formula is as follows: Where, Δ=(10 -5 The step size is a small perturbation, which avoids the complexity of analytical differentiation.

[0041] For each feature j, the absolute values ​​of the gradients of all candidate solutions are summarized and normalized. The formula for calculating the absolute value of the gradient is as follows: The formula for calculating the normalization is as follows: Subsequently, outliers were filtered using the 3σ principle, removing gradients that deviated excessively from the mean and retaining g. i，j If and only if |g i，j -μ j |≤3σ j ;where μ j , σ j The mean and standard deviation of the j-th feature gradient.

[0042] Finally, the generated real-time weight vector W is used as input data for the knowledge base and stored in the time series database.

[0043] Specifically, the knowledge base uses Local Sensitive Hash (LSH) to quickly retrieve historical data and obtain historical weight vectors, security constraint boundaries, and confidence scores.

[0044] The results obtained from the above steps, along with the real-time weight vector generated by the feature importance analysis, are fed back into the dynamic weighted preprocessing to perform a new round of fusion weight calculation.

[0045] Example 2

[0046] like Figure 2 As shown, this invention proposes a real-time error compensation method for voltage transformers. The method includes data processing by the CB module. The following are the steps of the CB module in processing the weighted time data sequence.

[0047] In the data processing stage of the CB module, the first step is to extract spatial features from the CNN model.

[0048] Let the input power data after dynamic weighting preprocessing be X. t The dimensions are [number of samples, time step, number of features].

[0049] Due to the dynamic weighted preprocessing, the number of convolutional kernels corresponding to high-weight features will increase: Convolution kernel number j =Base×(1+W) final [j]) For example, if the base number of convolution kernels is 32 and the V1 weight is 0.7, then this channel is allocated 54 convolution kernels.

[0050] Specifically, the weighted data is directly input into the CNN model, so that key features dominate the convolutional output.

[0051] Unlike traditional CNN models, this module introduces the historical hidden states of a BiLSTM model into the convolutional layers to achieve dynamic interaction of spatiotemporal features. The temporal features output by the BiLSTM model are used to correct the convolutional kernel weights of the CNN model or modulate the convolutional input in real time through the feedback channel, forming a closed-loop control. The output C of the Lth convolutional layer... L The calculation formula is: Among them, K m The m-th convolutional kernel is used to capture local features (such as the rate of voltage change); E t-1:t-N Let be the hidden state of the BiLSTM model layer for the first N time steps; avg represents the averaging operation, providing temporal context; M is the number of convolutional kernels; γ is the fusion coefficient, used to dynamically adjust the weights of spatial and temporal features. The hidden state mean of BiLSTM is avg(E). t-1:t-N The coefficient γ participates in the convolution output.

[0052] The power data feature map S of the pooling layer output by the improved Switch activation function L Dimensionality reduction is performed to preserve the important features of the power data, and max pooling is used with a pooling window size of [P,1] and a step size of S. The output P of the pooling layer is... L The formula is: Where i represents the sample index and j represents the time step index.

[0053] The fully connected layer is responsible for transforming the pooled feature matrix into a large one-dimensional feature vector. Let the weight matrix of the fully connected layer be Q, and the bias be m. Then the output R of the fully connected layer... L Represented as: R L =Q·P l +m

[0054] The spatial features R output by the CNN model L Based on the concatenation with the original time series data, the time series modeling of the BiLSTM model is completed.

[0055] The time-series features output by the BiLSTM model are used to correct the convolutional kernel weights or modulate the convolutional input of the CNN model in real time through the feedback channel, forming a closed-loop control. This breaks through the limitations of the traditional serial architecture and enables the model to optimize the spatial feature extraction strategy in real time according to the dynamic operating conditions of the power system (such as load changes and temperature drift), thereby significantly improving the accuracy of error compensation.

[0056] Specifically, based on the traditional LSTM, the iteration of LSTM neurons is optimized by introducing the trait pooling results of the CNN model as an additional input for gating, where: for t =σ(M f ·[e t-1 C t-1 ,ax-pool(C L )]+β f ), in t =σ(M i ·[e t-1 ,x t ,max-pool(C L )]+β i ), C t =for t ⊙C t+1 +i t ⊙tanh(M c [,x t ,max-pool(C L )]+β c ), out t =σ(M o ·[e t-1 C t ,max-pool(C L )+β o ]), In the formula, t is the Hadarmad product. for t 、in t C t out t These represent the values ​​of the forget gate, input gate, cell state, and output gate, respectively. t Represents the input at time t, e t-1 M represents the hidden state value at time t-1. f M i M c M o These represent the forget gate, input gate, cell state, and output gate, respectively, in the process of extracting time-series features from power data. t-1 The weighting coefficient, β f β i β c β o These represent the bias values ​​of the forget gate, input gate, cell state, and output gate, respectively, during the extraction of time-series features from power data.

[0057] The hidden layer state e(t) at time t is obtained from the input gate out(t) and the current cell state c(t), then: e t =out(t)⊙tanh(C t )

[0058] The output of the BiLSTM model is the concatenation of the results obtained from two inverse LSTMs. In this way, the BiLSTM model can fully capture the temporal relationship of the power data collected during the voltage transformer error compensation operation and output a vector H containing the temporal characteristics of the power data.

[0059] Finally, the CB module extracts the spatial features R of the power data from the CNN model layers. L The power data time-series features H extracted by the BiLSTM model are fused together. Through the mapping relationship g, the fused feature vector is transformed into candidate compensation turns (candidate solutions), and multiple candidate solutions L are output. i : L i =g(H,R) L ).

[0060] Example 3

[0061] like Figure 3 As shown, this invention proposes a real-time error compensation method for voltage transformers. The method includes using the DWGO algorithm to process the candidate solution set and output the optimal solution.

[0062] During the optimal solution-finding phase of the DGWO algorithm, the first step is to receive the candidate compensation turns set generated by the CB module. Each L i There is a corresponding compensation scheme.

[0063] Candidate solutions are mapped to individual gray wolves and sorted into four categories according to their utility values: 1. Alpha Wolf: The current optimal solution (minimum utility value) 2. β Wolf: Suboptimal Solution 3. δ Wolf: Third optimal solution 4. ωwolf: Other candidate solutions

[0064] The formula for calculating the utility value of a candidate solution is as follows: Where, a smaller utility value indicates a better compensation plan, N represents the total number of data samples, and V measured,n V represents the measured value of the voltage transformer in the nth sample. n Indicates candidate solution L i The predicted voltage value of the nth sample obtained by the corresponding compensation scheme (the compensated voltage value is calculated by using fractional-turn compensation according to the compensation scheme).

[0065] Subsequently, the DGWO algorithm gradually narrows the encirclement and dynamically adjusts the wolf pack's position within the solution space formed by the candidate solutions by continuously adjusting and optimizing the positions and distribution of the candidate solutions, thus gradually approaching the global optimal solution.

[0066] The distance formula between the candidate solution and the optimal solution (the solution with the minimum utility value) is as follows: D i =ρ·|sigmoid(r1)·L p -L i |+(1-ρ)·std(L 1:i-1,i+1:n ) Where ρ is the adaptive weighting factor. num is the current iteration number, NUM is the total number of iterations, r1 is a random number in [0,1], and L p Let L be the optimal solution position vector. i std(L) represents the position of the current candidate solution. 1:i-1,i+1:n The standard deviation () represents the position of other candidate solutions besides the current candidate solution, and is used to measure the dispersion of candidate solutions.

[0067] The formula for updating the position of candidate solutions is as follows: L i (t+1)=L p -η·D i Where η is the learning factor, η0 is the initial learning factor, and θ is the adjustment parameter.

[0068] Perform twin diversity operations.

[0069] Specifically, the twin diversity operation is based on the diversity of candidate solutions, and the probability calculation formula for the twin diversity operation is as follows: Among them, P C1 and P C0 Here, var(L) represents the lower and upper limits of the predefined twin probability, var(L) represents the variance of the current candidate solution set, and max(var(L)) represents the maximum value of the variance of the candidate solution set throughout the entire optimization process.

[0070] Select multiple candidate solutions with high utility values ​​as parents, and randomly select two candidate solutions from them, according to P. C The formula for the location of the newly created candidate solution after completing the twin operation is as follows: Among them, L a L bλ represents the positions of the two selected candidate solutions, and λ is a random number within the range [0,1]. Perturbation factor rangdn(0,1) represents a random number that follows a normal distribution with a mean of 0 and a standard deviation of 1.

[0071] Perform singleton derivation operations.

[0072] By expanding the candidate solution set using the singleton derivation operation to prevent premature convergence, and simulating abrupt changes in nature, certain characteristics of some candidate solutions are randomly altered, leading the algorithm into new search directions and enhancing global search performance. The probabilistic process of the singleton derivation operation is as follows: Among them, P m1 and P m0 f represents the preset lower and upper bounds of the derivative probability. i It is the utility value corresponding to the candidate solution. It is the average utility value of the candidate solution set.

[0073] The formula for singleton derivation is as follows: in, It is a new candidate solution generated after the singleton derivation operation, where ξ is the derivation strength coefficient, which decreases with the number of iterations. ξ0 is the initial derivative intensity coefficient.

[0074] Hunting behavior simulation and optimal solution output.

[0075] Three Wolves Guiding Mechanism: Update ω wolf based on the combined positional information of α, β, and δ wolves: Among them, L best This represents the current position of the global optimal solution, where ω is the inertia weight. ω0 is the initial inertia weight, L i This indicates the position of the current optimal solution.

[0076] When the maximum number of iterations NUM is reached or the rate of change of the fitness of the optimal solution is less than the threshold, the α wolf position is output as the final compensation turn value.

[0077] The final output compensation value at this stage achieves high-precision dynamic compensation of errors by adjusting the number of turns on the secondary side of the voltage transformer in real time, which significantly improves the reliability of power system metering.

[0078] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. It should be noted that for those skilled in the art, any changes or substitutions that can be easily conceived without departing from the technical principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for real-time error compensation of a voltage transformer, characterized in that, The method includes: The knowledge base receives real-time weight vectors, retrieves historical weight vectors from historical scenarios, and feeds the two types of vectors back to dynamic weighted preprocessing to obtain a fused weight vector. The CB module establishes a bidirectional information exchange channel between the CNN model and the BiLSTM model. The input time series data will be processed by the CNN model and the BiLSTM model, and a candidate solution set for the voltage transformer error compensation operation will be obtained by linear mapping. The improved Grey Wolf algorithm processes the candidate solution set and outputs the optimal compensation turn value.

2. The real-time error compensation method for voltage transformers according to claim 1, characterized in that, The time series data is arranged in chronological order and with a determined appropriate time step, divided into continuous time windows, with the data in each window serving as the input data for the CB module.

3. The real-time error compensation method for voltage transformers according to claim 1 or 2, characterized in that, When the system runs for the first time, the knowledge base provides preset parameters for the CB module, enabling the CB module to effectively generate candidate solutions.

4. The real-time error compensation method for voltage transformers according to claim 3, characterized in that, The feature importance analysis quantifies the contribution weight of each feature in real time, performs utility value gradient calculation, and generates a real-time weight vector.

5. The real-time error compensation method for voltage transformers according to claim 4, characterized in that, The aforementioned bidirectional information interaction channel establishes a bidirectional data flow mechanism between the CNN model and the BiLSTM model, enabling deep fusion of the CNN model and the BiLSTM model, as well as collaborative optimization of spatiotemporal features.

6. The real-time error compensation method for voltage transformers according to claim 5, characterized in that, The aforementioned bidirectional information interaction channel, from the CNN model to the BiLSTM model, introduces spatial features extracted by the CNN model into the input gate, forget gate, and cell state calculation of the BiLSTM model, thereby enabling the BiLSTM model to extract temporal features.

7. The real-time error compensation method for voltage transformers according to claim 1 or 6, characterized in that, The aforementioned bidirectional information interaction channel, from the BiLSTM model to the CNN model, introduces the historical hidden state of the BiLSTM model into the convolution operation of the CNN model, enabling the CNN model to perceive temporal dependencies and thus extract local spatial features.

8. The real-time error compensation method for voltage transformers according to claim 7, characterized in that, The linear mapping method described above maps the spatiotemporal features extracted by fusing the CNN model and the BiLSTM model into a set of candidate compensation turn values ​​through a linear transformation, which are then used by the improved Grey Wolf optimization algorithm to select the optimal solution.

9. The real-time error compensation method for voltage transformers according to claim 1, characterized in that, The improved Grey Wolf algorithm uses the twin diversity operation to generate new solutions with higher diversity from high-quality candidate solutions, expanding the search range and discovering potential better regions.

10. The real-time error compensation method for voltage transformers according to claim 9, characterized in that, The improved Grey Wolf algorithm uses a singleton derivation operation to locally perturb a single candidate solution, introducing randomness to escape local optima.

Images