Primary side saturation current recovery method and apparatus based on physical information injection

Abstract

The application discloses a primary side saturation current recovery method and device based on physical information injection, and relates to the technical field of current recovery, wherein the method comprises the following steps: constructing an initial primary side current recovery model comprising a full convolution network and a long short-term memory network; training the initial primary side current recovery model based on historical secondary side current data and historical primary side current data by using a loss function based on physical information weighting to obtain a target primary side current recovery model meeting preset conditions; detecting actual secondary side current data of a current transformer by using a preset saturation detection method to obtain binary saturation state data; and recovering saturation current based on the actual secondary side current data and the binary saturation state data by using the target primary side current recovery model to obtain a primary side saturation current recovery result. The method can improve the accuracy of primary side current recovery.

Description

Technical Field

[0001] This invention relates to the field of current recovery technology, and in particular to a method, apparatus, storage medium, and electronic device for primary-side saturated current recovery based on physical information injection. Background Technology

[0002] With the continuous expansion of modern power systems and the increasing complexity of distribution network structures, the operating conditions of power grids are becoming more and more variable, and the frequency and diversity of line transient processes have significantly increased. Current transformers, which undertake measurement tasks, directly affect the reliability of power grid operation through their measurement accuracy and dynamic response capabilities. However, when abnormal conditions such as short circuits, lightning strikes, or large current surges occur, the iron core of the current transformer is prone to saturation, causing severe distortion of the secondary current waveform. This results in an inability to accurately reflect changes in the primary current, leading to maloperation, failure to operate, or delayed operation of protection devices. Therefore, how to achieve accurate measurement and real-time reconstruction of the primary current under complex transient conditions has become a key issue for the stable operation of modern power systems.

[0003] Although the overall transmission model based on electromagnetic mechanisms can comprehensively characterize the physical properties of protective rheology and exhibits high modeling accuracy under various operating conditions such as steady state and fault, it still has certain limitations. This model relies on the accurate depiction of magnetization curve parameters and excitation branch characteristics, requiring the acquisition of a large number of structural and material parameters during the modeling stage. However, accurately obtaining this information is often difficult in practical engineering. Furthermore, the model's description of complex nonlinear factors such as core saturation, parasitic effects, and temperature changes still has some approximation; when the system's operating environment changes, the parameter identification and model correction processes are not only computationally intensive but also costly to maintain.

[0004] With the rapid development of artificial intelligence and machine learning technologies, data-driven modeling methods offer a novel research approach for characterizing the state of protective rheological systems and compensating for measurement errors. Unlike traditional analytical models, data-driven models do not rely on precise physical structural information. Instead, they automatically extract the nonlinear mapping relationship between inputs and outputs through learning from large amounts of data, thereby achieving high-precision reconstruction of electrical quantities and dynamic response prediction. Under conditions such as complex electromagnetic coupling, core nonlinear saturation, and transient disturbances, these methods exhibit stronger adaptability and robustness. However, purely data-driven models often lack a clear physical basis, and their internal decision-making processes lack interpretability. This "black box" characteristic limits the direct application of data-driven methods in fields such as relay protection where safety and reliability requirements are extremely high. If the model output lacks clear physical basis, even with high accuracy, it is difficult to gain sufficient trust in engineering practice.

[0005] Therefore, there is an urgent need to develop a new modeling approach that integrates physical mechanisms and artificial intelligence. This model can fully leverage the learning advantages of AI in extracting nonlinear laws from data, reducing reliance on precise physical parameters, while also maintaining the interpretability and credibility of the results through physical constraints. Through a well-designed model architecture, data learning and physical laws can be integrated to form a new method for protecting rheological saturation compensation that combines accuracy and reliability. Summary of the Invention

[0006] In view of this, the present invention provides a method, device, storage medium and electronic device for primary-side saturation current recovery based on physical information injection. The main purpose is to solve the problem that the secondary current waveform of the current transformer is severely distorted due to core saturation under fault transient conditions, and cannot accurately reflect the primary current.

[0007] To address the aforementioned problems, this application provides a method for primary-side saturation current recovery based on physical information injection, comprising: Construct an initial primary-side current recovery model that includes a fully convolutional network and a long short-term memory network; The initial primary current recovery model is trained using a loss function weighted by physical information based on historical secondary current data and historical primary current data to obtain a target primary current recovery model that meets preset conditions. A preset saturation detection method is used to detect the actual secondary current data of the current transformer to obtain binarized saturation state data; Based on the actual secondary current data and the binarized saturation state data, the target primary current recovery model is used to recover the saturation current, and the primary saturation current recovery result is obtained.

[0008] Optionally, the initial primary current recovery model is trained using a loss function weighted by physical information based on historical secondary current data and historical primary current data to obtain a target primary current recovery model that meets preset conditions, specifically including: Step 1: Use a preset saturation detection method to detect the historical secondary current data of the current transformer to obtain historical binary saturation state data; Step 2: Based on the historical secondary current data and the historical binary saturation state data, the initial primary current recovery model is used to perform current recovery to obtain the initial primary saturation current recovery value. Step 3: Update the weights of the loss function based on the historical binary saturation state data; Step 4: Based on the historical primary current data corresponding to the historical secondary current data and the initial primary saturation current recovery value, the loss function after weight update is used to calculate and process the initial loss value. Step 5: Based on the initial loss value, update the model parameters of the initial primary current recovery model using the backpropagation algorithm to obtain the current primary current recovery model; Step 6: Repeat steps 1 to 5 to update the model parameters of the current primary current recovery model until the preset iteration termination condition is met, and obtain the target primary current recovery model.

[0009] Optionally, the step of using a preset saturation detection method to detect the actual secondary current data of the current transformer to obtain binary saturation state data specifically includes: Multiple parallel convolutional branches are used to extract local features from the actual secondary current data to obtain first local feature data corresponding to different parallel convolutional branches; The first local feature data and the actual secondary current data are spliced ​​together in the channel dimension to obtain the first spliced ​​feature data. The first spliced ​​feature data is subjected to feature fusion and dimensionality reduction processing to obtain the saturation state probability value corresponding to the actual secondary side current data; The saturation state probability value is converted using a preset threshold to obtain the binarized saturation state data.

[0010] Optionally, the step of using the target primary current recovery model to recover the saturation current based on the actual secondary current data and the binarized saturation state data to obtain the primary saturation current recovery result specifically includes: The target primary current recovery model is used to extract local features from the actual secondary current data and the binarized saturation state data to obtain the second local feature data corresponding to different convolutional branches of the full convolutional network. The actual secondary current data, the binarized saturation state data, and each of the second local feature data are spliced ​​together in the channel dimension to obtain the second spliced ​​feature data. The second spliced ​​feature data is subjected to feature fusion and dimensionality reduction processing to obtain a high-dimensional feature sequence; The high-dimensional feature sequence is extracted using the long short-term memory network of the target primary current recovery model to obtain the primary saturation current recovery value.

[0011] Optionally, the fully convolutional network employing the target primary current recovery model performs local feature extraction on the actual secondary current data and the binarized saturation state data to obtain second local feature data corresponding to different convolutional branches of the fully convolutional network, specifically including: The first branch of the fully convolutional network of the target primary current recovery model is used to extract features from the actual secondary current data and the binarized saturation state data to obtain medium-scale waveform features. The second branch of the fully convolutional network of the target primary current recovery model is used to extract features from the actual secondary current data and the binarized saturation state data to obtain short-time waveform transition features that characterize the changes in details. The third branch of the fully convolutional network of the target primary current recovery model is used to extract features from the actual secondary current data and the binarized saturation state data to obtain long-term distortion features that characterize the global morphology. The first branch includes 32 convolutional kernels with a kernel size of 5×1; The second branch includes a first convolutional layer and a second convolutional layer. The first convolutional layer includes 64 convolutional kernels with a kernel size of 3×1, and the second convolutional layer includes 64 convolutional kernels with a kernel size of 5×1. The third branch includes a third convolutional layer and a fourth convolutional layer. The third convolutional layer includes 64 convolutional kernels with a kernel size of 3×1, and the fourth convolutional layer includes 128 convolutional kernels with a kernel size of 3×1. The second local feature data includes the medium-scale waveform features, the short-time waveform transition features, and the long-time distortion features.

[0012] Optionally, the step of performing feature fusion and dimensionality reduction processing on the second concatenated feature data to obtain a high-dimensional feature sequence specifically includes: The second spliced ​​feature data is subjected to channel dimensionality reduction processing to obtain the first feature data; The first feature data is subjected to time-series correlation feature aggregation processing to obtain the second feature data; Feature extraction is performed on the second feature data to obtain the high-dimensional feature sequence.

[0013] Optionally, the step of using the long short-term memory network of the target primary-side current recovery model to extract key features from the high-dimensional feature sequence to obtain the primary-side saturation current recovery value specifically includes: The first LSTM network of the long short-term memory network of the target primary current recovery model is used to extract features from the high-dimensional feature sequence to obtain the hidden state features corresponding to the high-dimensional feature sequence. The second LSTM network of the long short-term memory network of the target primary current recovery model is used to extract context features from the hidden state features to obtain the primary saturation current recovery value.

[0014] To address the aforementioned problems, this application provides a primary-side saturation current recovery device based on physical information injection, comprising: The building block is used to construct the initial primary-side current recovery model, which includes a fully convolutional network and a long short-term memory network. The training module is used to train the initial primary current recovery model based on historical secondary current data and historical primary current data using a loss function weighted by physical information, so as to obtain a target primary current recovery model that meets preset conditions. The detection module is used to detect the actual secondary current data of the current transformer using a preset saturation detection method to obtain binary saturation state data. The recovery module is used to perform saturation current recovery based on the actual secondary current data and the binarized saturation state data using the target primary current recovery model, and obtain the primary saturation current recovery result.

[0015] To address the aforementioned problems, this application provides a storage medium storing a computer program that, when executed by a processor, implements the steps of the aforementioned method for restoring primary saturated current based on physical information injection.

[0016] To address the aforementioned problems, this application provides an electronic device, comprising at least a memory and a processor. The memory stores a computer program, and the processor, when executing the computer program in the memory, implements the steps of the aforementioned method for restoring primary saturated current based on physical information injection.

[0017] The beneficial effects of this application are as follows: This application addresses the problem of severe waveform distortion of the secondary current in current transformers under fault transient conditions due to core saturation, which fails to accurately reflect the primary current. It proposes an end-to-end recovery method integrating physical information and artificial intelligence. First, the secondary current sequence is processed using a saturation detection algorithm to generate a binary sequence identifying the saturation state. Then, a composite model based on fully convolutional and long short-term memory networks is constructed, using the secondary current sequence and the saturation state sequence as dual-channel inputs, and the primary current as the output target. The model is trained by introducing a loss function weighted by the saturation state, ultimately achieving high-precision recovery of the primary current. This invention maintains the strong fitting ability of the artificial intelligence model while improving interpretability and reliability through multi-level physical information injection. It is applicable to saturation compensation of current transformers under complex transient conditions and can provide accurate measurement data for relay protection and fault analysis.

[0018] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of the present invention more apparent and understandable, specific embodiments of the present invention are described below. Attached Figure Description

[0019] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings: Figure 1 A flowchart illustrating a primary-side saturation current recovery method based on physical information injection provided in an embodiment of this application is shown. Figure 2 A flowchart illustrating a primary-side saturation current recovery method based on physical information injection, according to another embodiment of this application, is shown. Figure 3 A schematic diagram of the model structure used in the preset saturation detection method of this application embodiment is shown; Figure 4 A schematic diagram of the model structure of the target primary current recovery model according to an embodiment of this application is shown; Figure 5 A schematic diagram showing the recovery result of the primary-side saturation current after recovery according to an embodiment of this application is illustrated. Figure 6 A structural block diagram of a primary-side saturated current recovery device based on physical information injection, according to another embodiment of this application, is shown. Detailed Implementation

[0020] Various embodiments and features of this application are described herein with reference to the accompanying drawings.

[0021] It should be understood that various modifications can be made to the embodiments described herein. Therefore, the above description should not be considered as limiting, but merely as an example of embodiments. Other modifications within the scope and spirit of this application will be apparent to those skilled in the art.

[0022] The accompanying drawings, which are included in and form part of this specification, illustrate embodiments of the present application and, together with the general description of the present application given above and the detailed description of the embodiments given below, serve to explain the principles of the present application.

[0023] These and other features of this application will become apparent from the following description of preferred forms of embodiments given as non-limiting examples, with reference to the accompanying drawings.

[0024] It should also be understood that although this application has been described with reference to some specific examples, those skilled in the art can certainly implement many other equivalent forms of this application.

[0025] The above and other aspects, features and advantages of this application will become more apparent when taken in conjunction with the accompanying drawings and in view of the following detailed description.

[0026] Specific embodiments of this application are described thereafter with reference to the accompanying drawings; however, it should be understood that the claimed embodiments are merely examples of this application, which can be implemented in various ways. Well-known and / or repeated functions and structures are not described in detail to avoid unnecessary or redundant details that could obscure the application. Therefore, the specific structural and functional details claimed herein are not intended to be limiting, but merely to teach those skilled in the art to use this application in a variety of substantially any suitable detailed structures.

[0027] This specification may use the phrases “in one embodiment,” “in another embodiment,” “in yet another embodiment,” or “in other embodiments,” all of which may refer to one or more of the same or different embodiments according to this application.

[0028] This application provides a method for restoring primary-side saturated current based on physical information injection, such as... Figure 1 As shown, it includes: Step S101: Construct an initial primary-side current recovery model including a fully convolutional network and a long short-term memory network; In the specific implementation process, an initial primary current recovery model that can recover the primary current is constructed. The initial primary current recovery model uses a composite architecture of a fully convolutional network FCN and a long short-term memory network LSTM. It takes the secondary current data sequence and the binary saturation state data sequence obtained by detecting the secondary current data sequence using a preset saturation detection method as input, and the recovered primary current data sequence of the same length as output.

[0029] Step S102: Based on historical secondary current data and historical primary current data, the initial primary current recovery model is trained using a loss function weighted by physical information to obtain a target primary current recovery model that meets preset conditions. In the specific implementation process, Step 1: Detect the historical secondary current data of the current transformer using a preset saturation detection method to obtain historical binary saturation state data; Step 2: Based on the historical secondary current data and the historical binary saturation state data, perform current recovery using the initial primary current recovery model to obtain the initial primary saturation current recovery value; Step 3: Update the weights of the loss function based on physical information weighting based on the historical binary saturation state data; Step 4: Calculate and process the historical primary current data corresponding to the historical secondary current data and the initial primary saturation current recovery value using the weighted updated loss function to obtain the initial loss value; Step 5: Update the model parameters of the initial primary current recovery model using the backpropagation algorithm based on the initial loss value to obtain the current primary current recovery model; Step 6: Repeat steps 1 to 5 to update the model parameters of the current primary current recovery model until the preset iteration termination condition is met to obtain the target primary current recovery model.

[0030] Step S103: Use a preset saturation detection method to detect the actual secondary current data of the current transformer and obtain binarized saturation state data; In the specific implementation process, multiple parallel convolutional branches are used to extract local features from the actual secondary current data to obtain first local feature data corresponding to different parallel convolutional branches; the local feature data and the actual secondary current data are concatenated in the channel dimension to obtain first concatenated feature data; the concatenated feature data is fused and dimensionality reduced to obtain saturation state probability values ​​corresponding to the actual secondary current data; the saturation state probability values ​​are converted using a preset threshold to obtain the binarized saturation state data.

[0031] Step S104: Based on the actual secondary current data and the binarized saturation state data, the target primary current recovery model is used to perform saturation current recovery to obtain the primary saturation current recovery result.

[0032] In the specific implementation process, the actual secondary current data and binary saturation state data of the current transformer are input into the trained target primary current recovery model to obtain the recovered primary current result. The fully convolutional network of the target primary current recovery model is used to extract local features from the actual secondary current data and the binary saturation state data, obtaining second local feature data corresponding to different convolutional branches of the fully convolutional network. The actual secondary current data, the binary saturation state data, and each of the second local feature data are concatenated along the channel dimension to obtain second concatenated feature data. Feature fusion and dimensionality reduction are performed on the second concatenated feature data to obtain a high-dimensional feature sequence. The long short-term memory network of the target primary current recovery model is used to extract key features from the high-dimensional feature sequence to obtain the primary saturation current recovery value.

[0033] This application addresses the problem of severe waveform distortion in the secondary current of current transformers under fault transient conditions due to core saturation, which fails to accurately reflect the primary current. It proposes an end-to-end recovery method integrating physical information and artificial intelligence. First, a saturation detection algorithm is used to process the secondary current sequence, generating a binary sequence indicating the saturation state. Then, a composite model based on fully convolutional and long short-term memory networks is constructed, using the secondary current sequence and the saturation state sequence as dual-channel inputs, and the primary current as the output target. The model is trained by introducing a loss function weighted by the saturation state, ultimately achieving high-precision recovery of the primary current. This invention maintains the strong fitting ability of the artificial intelligence model while improving interpretability and reliability through multi-level physical information injection. It is applicable to saturation compensation of current transformers under complex transient conditions and can provide accurate measurement data for relay protection and fault analysis.

[0034] Another embodiment of this application provides a method for restoring primary-side saturation current based on physical information injection, such as... Figure 2 As shown, it includes: Step S201: Construct an initial primary-side current recovery model including a fully convolutional network and a long short-term memory network; In the specific implementation process of this step, an initial primary current recovery model that can recover the primary current is constructed. The initial primary current recovery model uses a composite architecture of a fully convolutional network FCN and a long short-term memory network LSTM. It takes the secondary current data sequence and the binary saturation state data sequence obtained by detecting the secondary current data sequence using a preset saturation detection method as input, and the recovered primary current data sequence of the same length as output.

[0035] Step S202: Use a preset saturation detection method to detect the historical secondary current data of the current transformer to obtain historical binary saturation state data; In this step, multiple parallel convolutional branches are used to extract local features from the historical secondary current data, obtaining historical first local feature data corresponding to different parallel convolutional branches. The historical first local feature data and the historical secondary current data are then concatenated along the channel dimension to obtain historical first concatenated feature data. Feature fusion and dimensionality reduction are performed on the historical first concatenated feature data to obtain historical saturation state probability values ​​corresponding to the historical secondary current data. Finally, a preset threshold is used to transform the historical saturation state probability values ​​to obtain the historical binarized saturation state data. The preset threshold can be 0.5, and can be set according to actual needs.

[0036] Step S203: Based on the historical secondary current data and the historical binary saturation state data, the initial primary current recovery model is used to perform current recovery to obtain the initial primary saturation current recovery value; In this step, the historical secondary current data and the historical binary saturation state data are input into the initial primary current recovery model to recover the current and obtain the initial primary saturation current recovery value. The initial primary current recovery model employs a dual-channel input strategy. This strategy allows the model to establish a physical understanding of waveform reliability from the initial information reception stage: it knows which data can be used directly and can identify which segments have significant distortion. Compared to traditional single-channel input neural networks, this structure significantly improves the model's feature perception ability and learning efficiency in the saturation region, providing clear physical guidance and stronger convergence stability for subsequent error compensation and waveform restoration.

[0037] Simulation methods can also be used to obtain training samples when training the model. The PSCAD / EMTDC electromagnetic transient simulation system is used to establish a dataset covering various typical operating conditions. The simulation model adopts a 500kV transmission network structure, and the protection rheological model is constructed based on the JA core magnetization theory. By systematically traversing key parameters such as the initial phase angle of the fault, fault impedance, fault distance, initial remanence, and transformer ratio, representative training and test sample sets are generated. The sampling frequency is set to 5kHz, and the length of a single recovery window is two power frequency cycles (200 sampling points). The complete dataset is randomly divided into training and test sets in a 9:1 ratio, meaning 90% of the samples are used for training and 10% for testing.

[0038] Step S204: Update the weights of the loss function based on physical information weighting based on the historical binarized saturation state data; In the specific implementation process of this step, the mathematical expression of the loss function based on physical information weighting is as follows:

[0039] in, This is the actual primary side current. The weighting coefficient is used to recover the primary side current. Saturation state of sampling points Related, and their relationship is expressed as:

[0040] When the sampling points are in a saturated state In the case of saturation, the loss function assigns higher weights to this region to amplify its impact on model parameter updates; while in the unsaturated state... The weights are relatively low, thus guiding the model to concentrate more learning resources on accurate compensation modeling in the saturated region. Hyperparameters The weighting level can be controlled by setting it to 0.8. This value shows a good balance effect in actual training, enabling the model to significantly improve the recovery accuracy of saturated sections while ensuring the overall waveform continuity.

[0041] Step S205: Based on the historical primary current data corresponding to the historical secondary current data and the initial primary saturation current recovery value, the loss function after weight update is used to calculate and process the initial loss value. In this step, the historical primary current data corresponding to the historical secondary current data and the initial primary saturation current recovery value are substituted into the weighted updated loss function for calculation to obtain the initial loss value.

[0042] Step S206: Based on the initial loss value, the model parameters of the initial primary current recovery model are updated using the backpropagation algorithm to obtain the current primary current recovery model; In this step, the gradient values ​​of the initial loss value and the model parameters of the initial primary current recovery model are calculated. Since saturation points have high weight, the errors caused by these points contribute more to the final gradient. A preset optimizer is used to update the model parameters based on the calculated gradient values, and the update direction is significantly biased towards reducing the prediction error in the saturation region.

[0043] Step S207: Repeat steps S202 to S206 to update the model parameters of the current primary current recovery model until the preset iteration termination condition is met, and obtain the target primary current recovery model. In this step, the initial learning rate is set to a low 0.0001 to ensure stable training; the batch size is 64; the training epochs are set to 200, and early stopping is not used to allow the model to converge fully. When the preset number of training epochs is reached, the target primary current recovery model is obtained. Model performance is evaluated on the test set using three metrics: root mean square error (RMSE), mean absolute error (MAE), and standard deviation (Std), calculated as follows:

[0044]

[0045]

[0046] in, To recover the current value sequence With the actual primary side current value sequence The difference sequence is called the error sequence, and it is defined as follows:

[0047] Overall test set results show that the recovery accuracy of this strategy reaches RMSE = 3.7%, MAE = 2.2%, and Std = 2.9%. The results indicate that this method can achieve stable recovery of the primary current under various saturation conditions, demonstrating strong robustness and generalization ability.

[0048] Step S208: Use a preset saturation detection method to detect the actual secondary current data of the current transformer and obtain binarized saturation state data; In the specific implementation process of this step, the model structure diagram of the preset saturation detection method is as follows: Figure 3As shown, multiple parallel convolutional branches are used to extract local features from the actual secondary current data to obtain the first local feature data corresponding to different parallel convolutional branches; the data size of the actual secondary current data is 1×L, where 1 represents the dimension and L represents the data length. The first parallel convolutional branch includes a first convolutional array of 64 kernels with a size of 3×1 and a second convolutional array of 64 kernels with a size of 5×1. The actual secondary-side current data is sequentially processed by the first and second convolutional arrays to obtain the first local feature data corresponding to the first parallel convolutional branch. The second parallel convolutional branch includes a third convolutional array of 64 kernels with a size of 3×1 and a fourth convolutional array of 128 kernels with a size of 3×1. The actual secondary-side current data is sequentially processed by the third and fourth convolutional arrays to obtain the first local feature data corresponding to the second parallel convolutional branch. The third parallel convolutional branch includes a fifth convolutional array of 32 kernels with a size of 3×1. The actual secondary-side current data is processed by convolutional arrays using the fifth convolutional array to obtain the first local feature data corresponding to the third parallel convolutional branch. This process yields local features at different scales. Multiple parallel convolutional branches include the first parallel convolutional branch, the second parallel convolutional branch, and the third parallel convolutional branch. The first local feature data and the actual secondary current data are concatenated along the channel dimension to obtain the first concatenated feature data; the data size of the first concatenated feature data is 225×L, where 225 represents the data dimension. Feature fusion and dimensionality reduction are performed on the first concatenated feature data to obtain the saturation state probability value corresponding to the actual secondary current data; features are extracted from the first concatenated feature data using 64 convolutions and a 1×1 convolution kernel to obtain the first fused feature data; features are extracted from the first fused feature data using 32 convolution kernels of size 3×1 to obtain the second fused feature data; convolution calculation is performed on the second fused feature data using a 1×1 convolution kernel to obtain the saturation state probability value; the saturation state probability value is converted using a preset threshold to obtain the binarized saturation state data. The preset threshold can be 0.5, and can be set according to actual needs. Finally, the output layer maps each sampling point to a probability value in the [0,1] interval, representing the probability that the point is in a saturated state. By setting a threshold, the saturation state probability value sequence is converted into a binary state sequence, where 1 represents saturation at the sampling point and 0 represents non-saturation, thereby enabling the identification of the saturation state of the current sampling point.

[0049] Step S209: Use the fully convolutional network of the target primary current recovery model to extract local features from the actual secondary current data and the binarized saturation state data to obtain the second local feature data corresponding to different convolutional branches of the fully convolutional network. In the specific implementation process of this step, the schematic diagram of the target primary current recovery model is as follows: Figure 4 As shown, the first branch of the fully convolutional network of the target primary current recovery model is used to extract features from the actual secondary current data and the binarized saturation state data to obtain medium-scale waveform features; the second branch of the fully convolutional network of the target primary current recovery model is used to extract features from the actual secondary current data and the binarized saturation state data to obtain short-time waveform transition features representing detailed changes; the third branch of the fully convolutional network of the target primary current recovery model is used to extract features from the actual secondary current data and the binarized saturation state data to obtain long-time distortion features representing the global morphology. The first branch includes 32 convolutional kernels of size 5×1; the second branch includes a first convolutional layer and a second convolutional layer, the first convolutional layer including 64 convolutional kernels of size 3×1 and the second convolutional layer including 64 convolutional kernels of size 5×1; the third branch includes a third convolutional layer and a fourth convolutional layer, the third convolutional layer including 64 convolutional kernels of size 3×1 and the fourth convolutional layer including 128 convolutional kernels of size 3×1; the second local feature data includes the medium-scale waveform features, the short-time waveform transition features, and the long-time distortion features.

[0050] Step S210: The actual secondary current data, the binarized saturation state data, and each of the second local feature data are spliced ​​together in the channel dimension to obtain the second spliced ​​feature data; In this step, the actual secondary current data, the binarized saturation state data, and each of the second local feature data are spliced ​​together along the channel dimension to obtain the second spliced ​​feature data; the data size of the second spliced ​​feature data is 226×L; where 226 represents the data dimension.

[0051] Step S211: Perform feature fusion and dimensionality reduction processing on the second spliced ​​feature data to obtain a high-dimensional feature sequence; In this step, channel dimensionality reduction is performed on the second spliced ​​feature data to obtain the first feature data; features are extracted from the second spliced ​​feature data using 64 convolutions and 1×1 kernels to obtain the first feature data; temporal correlation feature aggregation is performed on the first feature data to obtain the second feature data; features are extracted from the first feature data using 32 convolutions with 3×1 kernels to obtain the second feature data; features are then extracted from the second feature data to obtain the high-dimensional feature sequence. Convolution calculation is performed on the second feature data using 16 convolutions with 1×1 kernels to obtain the high-dimensional feature sequence; the data size of the high-dimensional feature sequence is 16×L. The high-dimensional feature sequence comprehensively reflects the local morphology, global trend, and saturation effect of the input signal at different time scales.

[0052] Step S212: Use the long short-term memory network of the target primary current recovery model to extract key features from the high-dimensional feature sequence to obtain the primary saturation current recovery value.

[0053] In this step, the first LSTM network of the Long Short-Term Memory (LSTM) network of the target primary-side current recovery model is used to extract features from the high-dimensional feature sequence, obtaining the hidden state features corresponding to the high-dimensional feature sequence. The second LSTM network of the LSTM network of the target primary-side current recovery model is then used to extract context features from the hidden state features, obtaining the primary-side saturation current recovery value. This high-dimensional feature sequence is then input into the second part of the LSTM network, which employs a two-layer stacked structure, with each layer containing 64 hidden units to enhance the ability to express long-term dependencies in the time series. The network terminates with a fully connected layer to generate the output sequence; as shown... Figure 5 The diagram shown is a schematic of the recovery result of the primary side saturation current after recovery in this application.

[0054] This application addresses the problem of severe waveform distortion in the secondary current of current transformers under fault transient conditions due to core saturation, which fails to accurately reflect the primary current. It proposes an end-to-end recovery method integrating physical information and artificial intelligence. First, a saturation detection algorithm is used to process the secondary current sequence, generating a binary sequence indicating the saturation state. Then, a composite model based on fully convolutional and long short-term memory networks is constructed, using the secondary current sequence and the saturation state sequence as dual-channel inputs, and the primary current as the output target. The model is trained by introducing a loss function weighted by the saturation state, ultimately achieving high-precision recovery of the primary current. This invention maintains the strong fitting ability of the artificial intelligence model while improving interpretability and reliability through multi-level physical information injection. It is applicable to saturation compensation of current transformers under complex transient conditions and can provide accurate measurement data for relay protection and fault analysis.

[0055] Another embodiment of this application provides a primary-side saturation current recovery device based on physical information injection, such as... Figure 6 As shown, it includes: Module 1 is used to build an initial primary-side current recovery model that includes a fully convolutional network and a long short-term memory network. Training module 2 is used to train the initial primary current recovery model based on historical secondary current data and historical primary current data using a loss function weighted by physical information, so as to obtain a target primary current recovery model that meets preset conditions. Detection module 3 is used to detect the actual secondary current data of the current transformer using a preset saturation detection method to obtain binarized saturation state data. Recovery module 4 is used to perform saturation current recovery based on the actual secondary current data and the binarized saturation state data using the target primary current recovery model, and obtain the primary saturation current recovery result.

[0056] In the specific implementation process, the training module 2 is specifically used for: Step 1, detecting the historical secondary current data of the current transformer using a preset saturation detection method to obtain historical binary saturation state data; Step 2, performing current recovery based on the historical secondary current data and the historical binary saturation state data using the initial primary current recovery model to obtain the initial primary saturation current recovery value; Step 3, updating the weights of the loss function based on physical information weighting based on the historical binary saturation state data; Step 4, calculating and processing the historical primary current data corresponding to the historical secondary current data and the initial primary saturation current recovery value using the weighted loss function to obtain the initial loss value; Step 5, updating the model parameters of the initial primary current recovery model using the backpropagation algorithm based on the initial loss value to obtain the current primary current recovery model; Step 6, repeating steps 1 to 5 to update the model parameters of the current primary current recovery model until the preset iteration termination condition is met to obtain the target primary current recovery model.

[0057] In the specific implementation process, the detection module 3 is specifically used to: extract local features from the actual secondary side current data using multiple parallel convolutional branches to obtain first local feature data corresponding to different parallel convolutional branches; concatenate each of the first local feature data and the actual secondary side current data in the channel dimension to obtain first concatenated feature data; perform feature fusion and dimensionality reduction on the first concatenated feature data to obtain a saturation state probability value corresponding to the actual secondary side current data; and convert the saturation state probability value using a preset threshold to obtain the binarized saturation state data.

[0058] In the specific implementation process, the recovery module 4 is specifically used to: extract local features from the actual secondary current data and the binarized saturation state data using the fully convolutional network of the target primary current recovery model, to obtain second local feature data corresponding to different convolutional branches of the fully convolutional network; concatenate the actual secondary current data, the binarized saturation state data, and each of the second local feature data in the channel dimension to obtain second concatenated feature data; perform feature fusion and dimensionality reduction on the second concatenated feature data to obtain a high-dimensional feature sequence; and extract key features from the high-dimensional feature sequence using the long short-term memory network of the target primary current recovery model to obtain the primary saturation current recovery value.

[0059] In specific implementation, the recovery module 4 is further used to: extract features from the actual secondary current data and the binarized saturation state data using the first branch of the fully convolutional network of the target primary current recovery model to obtain medium-scale waveform features; extract features from the actual secondary current data and the binarized saturation state data using the second branch of the fully convolutional network of the target primary current recovery model to obtain short-time waveform transition features representing detailed changes; and extract features from the actual secondary current data and the binarized saturation state data using the third branch of the fully convolutional network of the target primary current recovery model to obtain features representing the global morphology. The long-term distortion features are as follows: the first branch includes 32 convolutional kernels of size 5×1; the second branch includes a first convolutional layer and a second convolutional layer, the first convolutional layer includes 64 convolutional kernels of size 3×1, and the second convolutional layer includes 64 convolutional kernels of size 5×1; the third branch includes a third convolutional layer and a fourth convolutional layer, the third convolutional layer includes 64 convolutional kernels of size 3×1, and the fourth convolutional layer includes 128 convolutional kernels of size 3×1; the second local feature data includes the medium-scale waveform features, the short-term waveform transition features, and the long-term distortion features.

[0060] In the specific implementation process, the recovery module 4 is also used to: perform channel dimensionality reduction processing on the second spliced ​​feature data to obtain the first feature data; perform time-series correlation feature aggregation processing on the first feature data to obtain the second feature data; and perform feature extraction on the second feature data to obtain the high-dimensional feature sequence.

[0061] In the specific implementation process, the recovery module 4 is also used to: use the first LSTM network of the long short-term memory network of the target primary current recovery model to extract features from the high-dimensional feature sequence to obtain hidden state features corresponding to the high-dimensional feature sequence; use the second LSTM network of the long short-term memory network of the target primary current recovery model to extract context features from the hidden state features to obtain the primary saturation current recovery value.

[0062] This application addresses the problem of severe waveform distortion in the secondary current of current transformers under fault transient conditions due to core saturation, which fails to accurately reflect the primary current. It proposes an end-to-end recovery method integrating physical information and artificial intelligence. First, a saturation detection algorithm is used to process the secondary current sequence, generating a binary sequence indicating the saturation state. Then, a composite model based on fully convolutional and long short-term memory networks is constructed, using the secondary current sequence and the saturation state sequence as dual-channel inputs, and the primary current as the output target. The model is trained by introducing a loss function weighted by the saturation state, ultimately achieving high-precision recovery of the primary current. This invention maintains the strong fitting ability of the artificial intelligence model while improving interpretability and reliability through multi-level physical information injection. It is applicable to saturation compensation of current transformers under complex transient conditions and can provide accurate measurement data for relay protection and fault analysis.

[0063] Another embodiment of this application provides a storage medium storing a computer program, which, when executed by a processor, implements the following method steps: Step 1: Construct an initial primary-side current recovery model including a fully convolutional network and a long short-term memory network; Step 2: Based on historical secondary current data and historical primary current data, the initial primary current recovery model is trained using a loss function weighted by physical information to obtain a target primary current recovery model that meets the preset conditions. Step 3: Use a preset saturation detection method to detect the actual secondary current data of the current transformer and obtain the binarized saturation state data; Step 4: Based on the actual secondary current data and the binarized saturation state data, the target primary current recovery model is used to recover the saturation current and obtain the primary saturation current recovery result.

[0064] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0065] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is used as an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above.

[0066] The specific implementation process of the above method steps can be found in any of the above embodiments of the primary side saturation current recovery method based on physical information injection, and will not be repeated here.

[0067] This application addresses the problem of severe waveform distortion in the secondary current of current transformers under fault transient conditions due to core saturation, which fails to accurately reflect the primary current. It proposes an end-to-end recovery method integrating physical information and artificial intelligence. First, a saturation detection algorithm is used to process the secondary current sequence, generating a binary sequence indicating the saturation state. Then, a composite model based on fully convolutional and long short-term memory networks is constructed, using the secondary current sequence and the saturation state sequence as dual-channel inputs, and the primary current as the output target. The model is trained by introducing a loss function weighted by the saturation state, ultimately achieving high-precision recovery of the primary current. This invention maintains the strong fitting ability of the artificial intelligence model while improving interpretability and reliability through multi-level physical information injection. It is applicable to saturation compensation of current transformers under complex transient conditions and can provide accurate measurement data for relay protection and fault analysis.

[0068] Another embodiment of this application provides an electronic device, which can be a server. The electronic device includes a processor, a memory, a network interface, and a database connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile and / or volatile storage media and internal memory. The non-volatile storage media stores an operating system, computer programs, and a database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The network interface is used to communicate with external clients via a network connection. When the processor executes the program, it implements the server-side functions or steps of a primary-side saturation current recovery method based on physical information injection.

[0069] In one embodiment, an electronic device is provided, which can be a client. The electronic device includes a processor, memory, a network interface, a display screen, and an input device connected via a system bus. The processor provides computing and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The network interface is used to communicate with an external server via a network connection. When the program is executed by the processor, it implements client-side functions or steps of a primary-side saturation current recovery method based on physical information injection.

[0070] Another embodiment of this application provides an electronic device, including at least a memory and a processor. The memory stores a computer program, and the processor, when executing the computer program in the memory, performs the following method steps: Step 1: Construct an initial primary-side current recovery model including a fully convolutional network and a long short-term memory network; Step 2: Based on historical secondary current data and historical primary current data, the initial primary current recovery model is trained using a loss function weighted by physical information to obtain a target primary current recovery model that meets the preset conditions. Step 3: Use a preset saturation detection method to detect the actual secondary current data of the current transformer and obtain the binarized saturation state data; Step 4: Based on the actual secondary current data and the binarized saturation state data, the target primary current recovery model is used to recover the saturation current and obtain the primary saturation current recovery result.

[0071] The specific implementation process of the above method steps can be found in any of the above embodiments of the primary side saturation current recovery method based on physical information injection, and will not be repeated here.

[0072] This application addresses the problem of severe waveform distortion in the secondary current of current transformers under fault transient conditions due to core saturation, which fails to accurately reflect the primary current. It proposes an end-to-end recovery method integrating physical information and artificial intelligence. First, a saturation detection algorithm is used to process the secondary current sequence, generating a binary sequence indicating the saturation state. Then, a composite model based on fully convolutional and long short-term memory networks is constructed, using the secondary current sequence and the saturation state sequence as dual-channel inputs, and the primary current as the output target. The model is trained by introducing a loss function weighted by the saturation state, ultimately achieving high-precision recovery of the primary current. This invention maintains the strong fitting ability of the artificial intelligence model while improving interpretability and reliability through multi-level physical information injection. It is applicable to saturation compensation of current transformers under complex transient conditions and can provide accurate measurement data for relay protection and fault analysis.

[0073] The above embodiments are merely exemplary embodiments of this application and are not intended to limit this application. Those skilled in the art can make various modifications or equivalent substitutions to this application within the scope and nature of this application, and such modifications or equivalent substitutions should also be considered to fall within the scope of protection of this application.

Claims

1. A primary side saturation current recovery method based on physical information injection, characterized in that, include: Construct an initial primary-side current recovery model that includes a fully convolutional network and a long short-term memory network; The initial primary current recovery model is trained using a loss function weighted by physical information based on historical secondary current data and historical primary current data to obtain a target primary current recovery model that meets preset conditions. A preset saturation detection method is used to detect the actual secondary current data of the current transformer to obtain binarized saturation state data; Based on the actual secondary current data and the binarized saturation state data, the target primary current recovery model is used to recover the saturation current, and the primary saturation current recovery result is obtained. The process of restoring the saturated current using the target primary current recovery model based on the actual secondary current data and the binarized saturation state data, to obtain the primary saturated current recovery result, specifically includes: The target primary current recovery model is used to extract local features from the actual secondary current data and the binarized saturation state data to obtain the second local feature data corresponding to different convolutional branches of the full convolutional network. The actual secondary current data, the binarized saturation state data, and each of the second local feature data are spliced ​​together in the channel dimension to obtain the second spliced ​​feature data. The second spliced ​​feature data is subjected to feature fusion and dimensionality reduction processing to obtain a high-dimensional feature sequence; The long short-term memory network of the target primary current recovery model is used to extract key features from the high-dimensional feature sequence to obtain the primary saturation current recovery value. The fully convolutional network employing the target primary current recovery model extracts local features from the actual secondary current data and the binarized saturation state data, obtaining second local feature data corresponding to different convolutional branches of the fully convolutional network, specifically including: The first branch of the fully convolutional network of the target primary current recovery model is used to extract features from the actual secondary current data and the binarized saturation state data to obtain medium-scale waveform features. The second branch of the fully convolutional network of the target primary current recovery model is used to extract features from the actual secondary current data and the binarized saturation state data to obtain short-time waveform transition features that characterize the changes in details. The third branch of the fully convolutional network of the target primary current recovery model is used to extract features from the actual secondary current data and the binarized saturation state data to obtain long-term distortion features that characterize the global morphology. The first branch includes 32 convolutional kernels with a kernel size of 5×1; The second branch includes a first convolutional layer and a second convolutional layer. The first convolutional layer includes 64 convolutional kernels with a kernel size of 3×1, and the second convolutional layer includes 64 convolutional kernels with a kernel size of 5×1. The third branch includes a third convolutional layer and a fourth convolutional layer. The third convolutional layer includes 64 convolutional kernels with a kernel size of 3×1, and the fourth convolutional layer includes 128 convolutional kernels with a kernel size of 3×1. The second local feature data includes the medium-scale waveform features, the short-time waveform transition features, and the long-time distortion features.

2. The method of claim 1, wherein, The initial primary current recovery model is trained using a loss function weighted by physical information based on historical secondary current data and historical primary current data to obtain a target primary current recovery model that meets preset conditions, specifically including: Step 1: Use a preset saturation detection method to detect the historical secondary current data of the current transformer to obtain historical binary saturation state data; Step 2: Based on the historical secondary current data and the historical binary saturation state data, the initial primary current recovery model is used to perform current recovery to obtain the initial primary saturation current recovery value. Step 3: Update the weights of the loss function based on the historical binary saturation state data; Step 4: Based on the historical primary current data corresponding to the historical secondary current data and the initial primary saturation current recovery value, the loss function after weight update is used to calculate and process the initial loss value. Step 5: Based on the initial loss value, update the model parameters of the initial primary current recovery model using the backpropagation algorithm to obtain the current primary current recovery model; Step 6: Repeat steps 1 to 5 to update the model parameters of the current primary current recovery model until the preset iteration termination condition is met, and obtain the target primary current recovery model.

3. The method of claim 1, wherein, The method of using a preset saturation detection method to detect the actual secondary current data of the current transformer and obtain binary saturation state data specifically includes: Multiple parallel convolutional branches are used to extract local features from the actual secondary current data to obtain first local feature data corresponding to different parallel convolutional branches; The first local feature data and the actual secondary current data are spliced ​​together in the channel dimension to obtain the first spliced ​​feature data. The first spliced ​​feature data is subjected to feature fusion and dimensionality reduction processing to obtain the saturation state probability value corresponding to the actual secondary side current data; The saturation state probability value is converted using a preset threshold to obtain the binarized saturation state data.

4. The method of claim 1, wherein, The step of performing feature fusion and dimensionality reduction processing on the second concatenated feature data to obtain a high-dimensional feature sequence specifically includes: The second spliced ​​feature data is subjected to channel dimensionality reduction processing to obtain the first feature data; The first feature data is subjected to time-series correlation feature aggregation processing to obtain the second feature data; Feature extraction is performed on the second feature data to obtain the high-dimensional feature sequence.

5. The method of claim 1, wherein, The Long Short-Term Memory network employing the target primary-side current recovery model extracts key features from the high-dimensional feature sequence to obtain the primary-side saturation current recovery value, specifically including: The first LSTM network of the long short-term memory network of the target primary current recovery model is used to extract features from the high-dimensional feature sequence to obtain the hidden state features corresponding to the high-dimensional feature sequence. The second LSTM network of the long short-term memory network of the target primary current recovery model is used to extract context features from the hidden state features to obtain the primary saturation current recovery value.

6. A primary-side saturation current recovery device based on physical information injection, used to implement the primary-side saturation current recovery method based on physical information injection as described in any one of claims 1 to 5, characterized in that, include: The building block is used to construct the initial primary-side current recovery model, which includes a fully convolutional network and a long short-term memory network. The training module is used to train the initial primary current recovery model based on historical secondary current data and historical primary current data using a loss function weighted by physical information, so as to obtain a target primary current recovery model that meets preset conditions. The detection module is used to detect the actual secondary current data of the current transformer using a preset saturation detection method to obtain binary saturation state data. The recovery module is used to perform saturation current recovery based on the actual secondary current data and the binarized saturation state data using the target primary current recovery model, and obtain the primary saturation current recovery result.

7. A storage medium, characterized by The storage medium stores a computer program, which, when executed by a processor, implements the steps of the primary-side saturated current recovery method based on physical information injection as described in any one of claims 1-5.

8. An electronic device, comprising: It includes at least a memory and a processor, wherein the memory stores a computer program, and the processor, when executing the computer program in the memory, implements the steps of the primary-side saturated current recovery method based on physical information injection as described in any one of claims 1-5.

Images