CVT error prediction method, system, device and medium based on step-by-step layer extraction

Abstract

The present invention provides a CVT error prediction method, system, device and medium based on step-by-step hierarchical extraction. The method decomposes the acquired CVT error history sequence by using variational mode decomposition (VMD) to obtain modal components; constructs a step-by-step hierarchical feature extraction structure, and outputs final features based on the modal components and the CVT error history sequence; constructs a CVT error prediction model based on an improved LSTM model. The improved LSTM model specifically introduces an attention mechanism in the calculation of the candidate cell state by the LSTM model. The CVT error prediction model uses the final feature as input and outputs a CVT error prediction result. The present invention achieves accurate prediction of CVT errors through the above steps, improves the measurement accuracy and reliability of the power system, and provides data-driven decision support for power system operators.

Description

Technical Field

[0001] The present invention belongs to the technical field of online monitoring of electric power metering, and mainly relates to a CVT error prediction method, system, equipment and medium based on step-by-step layered extraction. Background Art

[0002] Capacitor Voltage Transformer (CVT) is an indispensable precision measuring device in the power system. Its main function is to accurately capture and transmit voltage signals, and provide key data support for the protection, measurement, control and metering of the power system. In order to ensure the metering performance of CVT, relevant departments have formulated strict testing standards and regulations, requiring regular offline power outage verification of CVT. However, the traditional offline power outage verification method often faces the problem of "untimely detection" in practice, and is unable to detect and solve potential problems in time. Secondly, power outage detection takes a long time, which not only affects normal power supply, but also increases the economic losses of power companies. Finally, the costs of equipment, manpower and time required for offline detection are also relatively high, which brings a certain economic burden to power companies. Therefore, promoting the transformation of CVT from traditional "periodic power outage detection" to "online monitoring" has become an important task in the current power industry.

[0003] For example, CN117826057A "A Capacitive Voltage Transformer Online Monitoring and Intelligent Diagnosis System" discloses "A Capacitive Voltage Transformer Online Monitoring and Intelligent Diagnosis System, including a data acquisition module, a data transmission module, a data quality assessment module, a data storage module, a model training module, a model prediction module, an online monitoring module and an information push alarm module, which uses the Gray Wolf Algorithm to adjust and optimize the LSTM input hyperparameters. The model training loss function is used as the fitness function of the Gray Wolf Algorithm to optimize the hyperparameters such as the number of hidden layer neurons, the learning rate, and the number of batches, so as to obtain the prediction model with the best prediction effect." However, the feature extraction of the invention mainly depends on the hyperparameter optimization of the model, and it is impossible to extract specific representative features, and the model prediction robustness and accuracy are insufficient. In addition, the invention uses the LSTM model, and the LSTM model has limitations when processing long sequences, and it is impossible to focus on important parts of the input data, and the prediction accuracy and stability are insufficient. Summary of the invention

[0004] In order to solve the above-mentioned problems existing in the prior art, the present application provides a CVT error prediction method, system, device and medium based on step-by-step layered extraction.

[0005] The technical solution of this application is as follows:

[0006] Based on the step-by-step layered extraction CVT error prediction method, the method includes:

[0007] Obtain the CVT error history sequence of the capacitor voltage transformer, and use variational mode decomposition VMD to decompose the CVT error history sequence to obtain modal components;

[0008] Construct a feature extraction structure, the feature extraction structure comprising +1 hierarchical feature extraction layer, each feature extraction layer includes Unique expert networks and 1 shared expert network, among which: The unique expert network, shared expert network and the first feature extraction layer +1 Each of the unique expert networks in the feature extraction layer is provided with a corresponding gating function; the feature extraction structure is based on The modal components and CVT error history sequence output the final features;

[0009] A CVT error prediction model based on the improved LSTM model is constructed, wherein the CVT error prediction model takes the final features as input and outputs the CVT error prediction results, wherein: the improved LSTM model specifically introduces an attention mechanism in the calculation of the candidate cell state by the LSTM model;

[0010] Training the CVT error prediction model to obtain a trained CVT error prediction model; using the trained CVT error prediction model to predict the CVT to be predicted to obtain a prediction result;

[0011] Among them, The output of the gating function corresponding to the expert network in the feature extraction layer is expressed as:

[0012] ;

[0013] ;

[0014] ;

[0015] In the formula, Indicated in In the feature extraction layer, Input data of an expert network; Indicated in In the feature extraction layer, The output data of the gating function; represents the weighting function; Indicates In the feature extraction layer The data selected by the gating function;

[0016] in:

[0017] ;

[0018] ;

[0019] ;

[0020] In the formula, express The weight matrix of represents the normalized exponential function; represents the CVT error history sequence; Indicates the obtained modal components; represents the input of the unique expert network in the first feature extraction layer; represents the input of the shared expert network in the first feature extraction layer; Indicated in In the feature extraction layer, Input data of an expert network; Indicates In the feature extraction layer The output of the gating function; when hour, Indicates The feature extraction layer The combined output of a unique expert network and the corresponding shared expert network; when hour, Indicates The combined output of feature extraction layers; Indicates In the feature extraction layer The corresponding output of the expert network.

[0021] Preferably, the CVT error history sequence is obtained by decomposing the CVT error history sequence using variational mode decomposition (VMD) to obtain The modal components are expressed as:

[0022] ;

[0023] ;

[0024] In the formula, represents the CVT error history sequence; Indicates the obtained modal components; represents the number of modal components; represents the Lagrange multiplier; represents the adjustment parameter; Indicates The time derivative of the modal components; represents the integral variable; represents the time variable; Indicates at a point in time Upper modal component The value of Indicates the integration time point Upper modal component The value of .

[0025] Preferably, the The output of +1 feature extraction layer is expressed as:

[0026]

[0027] In the formula, Indicates The final feature; Indicates The feature extraction layer The combined output of a unique expert network and the corresponding shared expert network.

[0028] Preferably, an attention mechanism is introduced in the calculation of the candidate cell state by the LSTM model, which is expressed as:

[0029] ;

[0030] ;

[0031] ;

[0032] ;

[0033] ;

[0034] ;

[0035] In the formula, Represents the final feature At the moment The value of ; and Represents the hidden layer state of the previous moment and the current moment in the LSTM model; , and Indicates the cell state at the previous moment, the candidate cell state at the current moment, and the current cell state; , and Represents the output of the forget gate, input gate, and output gate; and Represents pre-set trainable weight parameters; Representation calculation The bias parameters when ; Represents the bias parameter of the output gate; Represents the sigmoid function; and Represent the first attention mechanism and the second attention mechanism respectively.

[0036] Based on the step-by-step hierarchical extraction of CVT error prediction system, the system includes a data processing module, a feature extraction module, an error prediction module and an output module, wherein:

[0037] The data acquisition module is used to obtain the CVT error history sequence of the capacitor voltage transformer, and the CVT error history sequence is decomposed by using variational mode decomposition VMD to obtain modal components, and the CVT error history sequence and The modal components are transmitted to the feature extraction module;

[0038] The feature extraction module has a built-in feature extraction structure, and the feature extraction structure includes +1 hierarchical feature extraction layer, each feature extraction layer includes Unique expert networks and 1 shared expert network, among which: The unique expert network, shared expert network and the first feature extraction layer +1 Each of the unique expert networks in the feature extraction layer is provided with a corresponding gating function; the feature extraction structure is based on The modal components and CVT error history sequence output the final features;

[0039] The error prediction module has a built-in trained CVT error prediction model, which is based on an improved LSTM model and takes the final features as input, and outputs the CVT error prediction result to the output module;

[0040] The output module is used to display the CVT error prediction result;

[0041] Among them, The output of the gating function corresponding to the expert network in the feature extraction layer is expressed as:

[0042] ;

[0043] ;

[0044] ;

[0045] In the formula, Indicated in In the feature extraction layer, Input data of an expert network; Indicated in In the feature extraction layer, The output data of the gating function; represents the weighting function; Indicates In the feature extraction layer The data selected by the gating function;

[0046] in:

[0047] ;

[0048] ;

[0049] ;

[0050] In the formula, express The weight matrix of represents the normalized exponential function; represents the CVT error history sequence; Indicates the obtained modal components; represents the input of the unique expert network in the first feature extraction layer; represents the input of the shared expert network in the first feature extraction layer; Indicated in In the feature extraction layer, Input data of an expert network; Indicates In the feature extraction layer The output of the gating function; when hour, Indicates The feature extraction layer The combined output of a unique expert network and the corresponding shared expert network; when hour, Indicates The combined output of feature extraction layers; Indicates In the feature extraction layer The corresponding output of the expert network;

[0051] No. The output of +1 feature extraction layer is expressed as:

[0052] ;

[0053] In the formula, Indicates The final feature; Indicates The feature extraction layer The combined output of a unique expert network and the corresponding shared expert network.

[0054] Preferably, the CVT error prediction model is constructed based on an improved LSTM model. Specifically, the improved LSTM model introduces an attention mechanism in the calculation of the candidate cell state by the LSTM model, which is expressed as follows:

[0055] ;

[0056] ;

[0057] ;

[0058] ;

[0059] ;

[0060] ;

[0061] In the formula, Represents the final feature At the moment The value of ; and Represents the hidden layer state of the previous moment and the current moment in the LSTM model; , and Indicates the cell state at the previous moment, the candidate cell state at the current moment, and the current cell state; , and Represents the output of the forget gate, input gate, and output gate; and Represents pre-set trainable weight parameters; Representation calculation The bias parameters when ; Represents the bias parameter of the output gate; Represents the sigmoid function; and Represent the first attention mechanism and the second attention mechanism respectively.

[0062] An electronic device comprises a memory, a processor and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, a CVT error prediction method based on step-by-step hierarchical extraction as described in any one of the embodiments is implemented.

[0063] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements a CVT error prediction method based on step-by-step hierarchical extraction as described in any one of the embodiments.

[0064] Compared with the prior art, the present invention has the following beneficial effects:

[0065] 1) The present invention provides a CVT error prediction method, system, device and medium based on step-by-step layered extraction, decomposes the CVT error history sequence through variational mode decomposition (VMD), effectively extracts the features of different modes, and improves the accuracy of the prediction model;

[0066] 2) The present invention provides a CVT error prediction method, system, device and medium based on hierarchical extraction, constructs a hierarchical feature extraction structure to extract features, gradually extracts features from low-level to high-level, and the output of each layer of the network provides more refined feature information for the next layer. The hierarchical processing better captures complex error patterns; the feature extraction structure uses a gating function to dynamically select and adjust the weight of the feature to enhance the ability to extract key information; the feature extraction structure is in the first + The shared expert network in the 1st feature extraction layer does not set a gating function, and processes the global features of the entire CVT error history sequence, so that the feature extraction structure can learn the common features in the sequence, while avoiding overfitting and improving the model generalization ability; the feature extraction structure increases the accuracy of the feature extraction process according to the above design, and effectively integrates local features and global information;

[0067] 3) The present invention provides a CVT error prediction method, system, device and medium based on step-by-step hierarchical extraction, which utilizes two attention mechanisms to update the LSTM cell state, enhances the model's ability to extract important information from CVT error data, better adapts to complex data patterns and fluctuations, and enhances the model's generalization ability on different data sets, thereby improving the accuracy and stability of prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0068] Figure 1 is a flow chart of an embodiment of the present invention;

[0069] Figure 2 is a structural diagram of a feature extraction module according to an embodiment of the present invention;

[0070] Figure 3 This is the internal structure of the improved LSTM model of the embodiment of the present invention. DETAILED DESCRIPTION

[0071] The specific implementation modes of the present invention are described below so that those skilled in the art can understand the present invention. However, it should be clear that the present invention is not limited to the scope of the specific implementation modes. For those of ordinary skill in the art, as long as various changes are within the spirit and scope of the present invention as defined and determined by the attached claims, these changes are obvious, and all inventions and creations utilizing the concept of the present invention are protected.

[0072] The present invention provides the following technical solutions: a CVT error prediction method, system, equipment and medium based on step-by-step layered extraction.

[0073] Example 1

[0074] This embodiment provides a CVT error prediction method based on step-by-step layer extraction, and the specific steps include:

[0075] S1. Obtain the CVT error history sequence of the capacitor voltage transformer, and use variational mode decomposition (VMD) to decompose the CVT error history sequence to obtain modal components;

[0076] S11. Obtain CVT error history sequence , the CVT error history sequence is decomposed using variational mode decomposition VMD to obtain modal components, the component data is ;

[0077] S111, Variational Mode Decomposition

[0078] Variational Mode Decomposition (VMD) is an advanced signal processing technique used to decompose complex nonlinear and non-stationary signals into a series of Intrinsic Mode Functions (IMFs), each of which represents a component on a different time scale. It is based on the variational principle and local feature extraction, and decomposes the signal into multiple IMFs with different frequency and amplitude modulation through an iterative optimization process, making the final signal decomposition more refined and adaptable, and suitable for various real-time signal analysis and mode extraction applications.

[0079] The variational mode decomposition (VMD) is used to decompose the CVT error history sequence to obtain modal components, the objective function of variational mode decomposition VMD is expressed as:

[0080] ; (1)

[0081] ; (2)

[0082] In the formula, represents the CVT error history sequence; Indicates the obtained modal components; represents the number of modal components; represents the Lagrange multiplier; represents the adjustment parameter; Indicates The time derivative of the modal components; represents the integral variable; represents the time variable; Indicates at a point in time Upper modal component The value of Indicates the integration time point Upper modal component The value of .

[0083] S2. Construct a feature extraction structure, the feature extraction structure includes +1 hierarchical feature extraction layer, each feature extraction layer includes Unique expert networks and 1 shared expert network, among which: The unique expert network, shared expert network and the first feature extraction layer +1 Each of the unique expert networks in the feature extraction layer is provided with a corresponding gating function; the feature extraction structure is based on The modal components and CVT error history sequence output the final features;

[0084] S21, the above The output of the gating function corresponding to the expert network in the feature extraction layer is expressed as:

[0085] ; (3)

[0086] ; (4)

[0087] ; (5)

[0088] In the formula, Indicated in In the feature extraction layer, Input data of an expert network; Indicated in In the feature extraction layer, The output data of the gating function; represents the weighting function; Indicates In the feature extraction layer The data selected by the gating function;

[0089] in:

[0090] ; (6)

[0091] ; (7)

[0092] ; (8)

[0093] In the formula, express The weight matrix of represents the normalized exponential function; represents the CVT error history sequence; Indicates the obtained modal components; represents the input of the unique expert network in the first feature extraction layer; represents the input of the shared expert network in the first feature extraction layer; Indicated in In the feature extraction layer, Input data of an expert network; Indicates In the feature extraction layer The output of the gating function; when hour, Indicates The feature extraction layer The combined output of a unique expert network and the corresponding shared expert network; when hour, Indicates The combined output of feature extraction layers; Indicates In the feature extraction layer The corresponding output of the expert network;

[0094] S22, please refer to Figure 2 , the shared network input data is input into the shared expert network, and the unique expert input data is input into the unique expert network; in the previous In each feature extraction module, the gate calculation methods of the unique expert network and the shared expert network are different; for each component, there is a unique expert network, which can retain its own information uniqueness and learn other component information; the role of the shared expert network is to learn the information of all components, jointly extract the low-level representation at the high-level shared expert, and obtain high-level shared knowledge;

[0095] S23, the The output of +1 feature extraction layer is expressed as:

[0096] ; (9)

[0097] In the formula, Indicates The final feature; Indicates The feature extraction layer The combined output of a unique expert network and the corresponding shared expert network.

[0098] S3. Construct a CVT error prediction model based on the improved LSTM model, wherein the CVT error prediction model takes the final feature as input and outputs a CVT error prediction result, wherein: the improved LSTM model specifically introduces an attention mechanism in the calculation of the candidate cell state by the LSTM model;

[0099] S31. Improved LSTM model

[0100] The input gate and output gate in the classic long short-term memory network LSTM model use sigmoid functions to calculate and The degree of memory is used when updating the current cell state. and Using the above idea, two attention mechanisms are used to calculate and After weighting, when updating the current cell state, the outputs of the two attention mechanisms can be directly added together. , and As Query, key and Value; , and As Query, key and Value. and Internally, the methods for calculating Query and key correlations should be different.

[0101] See also Figure 3 , the improved LSTM model uses the attention mechanism to calculate the cell state within the LSTM, which is expressed as:

[0102] ; (10)

[0103] ; (11)

[0104] ; (12)

[0105] ; (13)

[0106] ; (14)

[0107] ; (15)

[0108] In the formula, Represents the final feature At the moment The value of ; and Represents the hidden layer state of the previous moment and the current moment in the LSTM model; , and Indicates the cell state at the previous moment, the candidate cell state at the current moment, and the current cell state; , and Represents the output of the forget gate, input gate, and output gate; Express The weight parameters used when performing linear transformation; Represents the weight parameters used by the output gate. In the early stage of model training, these two parameters are randomly initialized and will be gradually determined as the model is trained; Representation calculation The bias parameters when ; Represents the bias parameter of the output gate; Represents the sigmoid function; and Represent the first attention mechanism and the second attention mechanism respectively; and Represents the Query and key of the attention mechanism, and calculates the correlation between the Query and key. The output result is the correlation between Query and key. The result of multiplication, The output result is the correlation between Query and key. The result of multiplication; when updating the current cell state, the outputs of the two attention mechanisms are directly added.

[0109] Among them, formula (10) represents the calculation of the input gate, formula (11) represents the calculation of the candidate cell state, formula (12) represents the calculation of the update gate, formula (13) represents the calculation of the cell state at the current moment, formula (14) represents the calculation of the output gate, and formula (15) represents the calculation of the hidden layer state at the current moment.

[0110] S4. Train the CVT error prediction model to obtain a trained CVT error prediction model; use the trained CVT error prediction model to predict the CVT to be predicted to obtain a prediction result.

[0111] In this embodiment, in order to better demonstrate the prediction results of the CVT error prediction model, the following specific examples are given:

[0112] Based on the simulation platform, the voltage transformer error data under different temperatures, humidity, frequencies, and secondary loads are collected. The temperature changes from -40℃ to 40℃, the humidity changes from 10% to 80%, the frequency changes from 30kHz to 500kHz, and the power value of the secondary side load equipment is set between 0.2-50VA. Some error measurement results are shown in Table 1;

[0113] Table 1 shows the prediction results. The difference between the actual error value and the predicted error value of the CVT error prediction model is small, indicating that the prediction effect is good;

[0114] Table 1 Prediction results display

[0115]

[0116] Table 2 shows the evaluation index values ​​of the comparison models. The CVT error prediction model is compared with the classic LSTM model. The mean square error (MSE) of the prediction result of the CVT error prediction model is 0.1524, which is 0.1178 lower than that of the classic LSTM model; the mean absolute error (MAE) is 0.1120, which is 0.0938 lower than that of the classic LSTM model. The prediction result is significantly better than that of the classic LSTM model.

[0117] Table 2 Comparison model evaluation index values

[0118]

[0119] Example 2

[0120] This embodiment provides a CVT error prediction system based on step-by-step hierarchical extraction, the system comprising a data processing module, a feature extraction module, an error prediction module and an output module, wherein:

[0121] The data acquisition module is used to obtain the CVT error history sequence of the capacitor voltage transformer, and the CVT error history sequence is decomposed by using variational mode decomposition VMD to obtain modal components, and the CVT error history sequence and The modal components are transmitted to the feature extraction module;

[0122] The feature extraction module has a built-in feature extraction structure, and the feature extraction structure includes +1 hierarchical feature extraction layer, each feature extraction layer includes Unique expert networks and 1 shared expert network, among which: The unique expert network, shared expert network and the first feature extraction layer +1 Each of the unique expert networks in the feature extraction layer is provided with a corresponding gating function; the feature extraction structure is based on The modal components and CVT error history sequence output the final features;

[0123] The error prediction module has a built-in trained CVT error prediction model, which is based on an improved LSTM model and takes the final features as input, and outputs the CVT error prediction result to the output module;

[0124] The output module is used to display the CVT error prediction result;

[0125] Among them, The output of the gating function corresponding to the expert network in the feature extraction layer is expressed as:

[0126] ;

[0127] ;

[0128] ;

[0129] In the formula, Indicated in In the feature extraction layer, Input data of an expert network; Indicated in In the feature extraction layer, The output data of the gating function; represents the weighting function; Indicates In the feature extraction layer The data selected by the gating function;

[0130] in:

[0131] ;

[0132] ;

[0133] ;

[0134] In the formula, express The weight matrix of represents the normalized exponential function; represents the CVT error history sequence; Indicates the obtained modal components; represents the input of the unique expert network in the first feature extraction layer; represents the input of the shared expert network in the first feature extraction layer; Indicated in In the feature extraction layer, Input data of an expert network; Indicates In the feature extraction layer The output of the gating function; when hour, Indicates The feature extraction layer The combined output of a unique expert network and the corresponding shared expert network; when hour, Indicates The combined output of feature extraction layers; Indicates In the feature extraction layer The corresponding output of the expert network.

[0135] Example 3

[0136] This embodiment provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, the CVT error prediction method based on step-by-step hierarchical extraction as described in any one of Embodiment 1 is implemented.

[0137] Example 4

[0138] This embodiment provides a computer-readable storage medium on which a computer program is stored. When the program is executed by a processor, the CVT error prediction method based on step-by-step hierarchical extraction as described in any one of Embodiments 1 is implemented.

[0139] It is worth noting that the system, electronic device and computer-readable storage medium described in the present invention are based on the same principles as the method described in Example 1, and will not be described in detail here.

[0140] The above descriptions are merely embodiments of the present invention and are not intended to limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made using the contents of the present invention specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. A CVT error prediction method based on step-by-step hierarchical extraction, characterized in that: The method comprises: Obtain the CVT error history sequence of the capacitor voltage transformer, and use variational mode decomposition VMD to decompose the CVT error history sequence to obtain modal components; Construct a feature extraction structure, the feature extraction structure comprising +1 hierarchical feature extraction layer, each feature extraction layer includes Unique expert networks and 1 shared expert network, among which: The unique expert network, shared expert network and the first feature extraction layer +1 Each of the unique expert networks in the feature extraction layer is provided with a corresponding gating function; the feature extraction structure is based on The modal components and CVT error history sequence output the final features; A CVT error prediction model based on the improved LSTM model is constructed, wherein the CVT error prediction model takes the final features as input and outputs the CVT error prediction results, wherein: the improved LSTM model specifically introduces an attention mechanism in the calculation of the candidate cell state by the LSTM model; Training the CVT error prediction model to obtain a trained CVT error prediction model; using the trained CVT error prediction model to predict the CVT to be predicted to obtain a prediction result; Among them, The output of the gating function corresponding to the expert network in the feature extraction layer is expressed as: ； ； ； In the formula, Indicated in In the feature extraction layer, Input data of an expert network; Indicated in In the feature extraction layer, The output data of the gating function; represents the weighting function; Indicates In the feature extraction layer The data selected by the gating function; in: ； ； ； In the formula, express The weight matrix of represents the normalized exponential function; Represents the CVT error history sequence; Indicates the obtained modal components; represents the input of the unique expert network in the first feature extraction layer; represents the input of the shared expert network in the first feature extraction layer; Indicated in In the feature extraction layer, Input data of an expert network; Indicates In the feature extraction layer The output of the gating function; when hour, Indicates The feature extraction layer The combined output of a unique expert network and the corresponding shared expert network; when hour, Indicates The combined output of feature extraction layers; Indicates In the feature extraction layer The corresponding output of the expert network.

2. The CVT error prediction method based on step-by-step hierarchical extraction according to claim 1 is characterized in that: The CVT error history sequence is decomposed using variational mode decomposition (VMD) to obtain The modal components are expressed as: ； ； In the formula, Represents the CVT error history sequence; Indicates the obtained modal components; represents the number of modal components; represents the Lagrange multiplier; represents the adjustment parameter; Indicates The time derivative of the modal components; represents the integral variable; represents the time variable; Indicates at a point in time Upper modal component The value of Indicates the integration time point Upper modal component The value of .

3. The CVT error prediction method based on step-by-step hierarchical extraction according to claim 2 is characterized in that: No. The output of +1 feature extraction layer is expressed as: ； In the formula, Indicates The final feature; Indicates The feature extraction layer The combined output of a unique expert network and the corresponding shared expert network.

4. The CVT error prediction method based on step-by-step layered extraction according to claim 3 is characterized in that: The attention mechanism is introduced into the calculation of the candidate cell state by the LSTM model, which is expressed as follows: ； ； ； ； ； ； In the formula, Represents the final feature At the moment The value of ; and Represents the hidden layer state of the previous moment and the current moment in the LSTM model; , and Indicates the cell state at the previous moment, the candidate cell state at the current moment, and the current cell state; , and Represents the output of the forget gate, input gate, and output gate; and Represents pre-set trainable weight parameters; Representation calculation The bias parameters when ; Represents the bias parameter of the output gate; Represents the sigmoid function; and Represent the first attention mechanism and the second attention mechanism respectively.

5. The CVT error prediction system based on step-by-step hierarchical extraction is characterized by: The system includes a data processing module, a feature extraction module, an error prediction module and an output module, wherein: The data acquisition module is used to obtain the CVT error history sequence of the capacitor voltage transformer, and the CVT error history sequence is decomposed by using variational mode decomposition VMD to obtain modal components, and the CVT error history sequence and The modal components are transmitted to the feature extraction module; The feature extraction module has a built-in feature extraction structure, and the feature extraction structure includes +1 hierarchical feature extraction layer, each feature extraction layer includes Unique expert networks and 1 shared expert network, among which: The unique expert network, shared expert network and the first feature extraction layer +1 Each of the unique expert networks in the feature extraction layer is provided with a corresponding gating function; the feature extraction structure is based on The modal components and CVT error history sequence output the final features; The error prediction module has a built-in trained CVT error prediction model, which is based on an improved LSTM model and takes the final features as input, and outputs the CVT error prediction result to the output module; The output module is used to display the CVT error prediction result; Among them, The output of the gating function corresponding to the expert network in the feature extraction layer is expressed as: ； ； ； In the formula, Indicated in In the feature extraction layer, Input data of an expert network; Indicated in In the feature extraction layer, The output data of the gating function; represents the weighting function; Indicates In the feature extraction layer The data selected by the gating function; in: ； ； ； In the formula, express The weight matrix of represents the normalized exponential function; Represents the CVT error history sequence; Indicates the obtained modal components; represents the input of the unique expert network in the first feature extraction layer; represents the input of the shared expert network in the first feature extraction layer; Indicated in In the feature extraction layer, Input data of an expert network; Indicates In the feature extraction layer The output of the gating function; when hour, Indicates The feature extraction layer The combined output of a unique expert network and the corresponding shared expert network; when hour, Indicates The combined output of feature extraction layers; Indicates In the feature extraction layer The corresponding output of the expert network; No. The output of +1 feature extraction layer is expressed as: ； In the formula, Indicates The final feature; Indicates The feature extraction layer The combined output of a unique expert network and the corresponding shared expert network.

6. The CVT error prediction system based on step-by-step hierarchical extraction according to claim 5, characterized in that: The CVT error prediction model is built based on the improved LSTM model. The improved LSTM model specifically introduces an attention mechanism in the calculation of the candidate cell state by the LSTM model, which is expressed as follows: ； ； ； ； ； ； In the formula, Represents the final feature At the moment The value of ; and Represents the hidden layer state of the previous moment and the current moment in the LSTM model; , and Indicates the cell state at the previous moment, the candidate cell state at the current moment, and the current cell state; , and Represents the output of the forget gate, input gate, and output gate; and Represents pre-set trainable weight parameters; Representation calculation The bias parameters when ; Represents the bias parameter of the output gate; Represents the sigmoid function; and Represent the first attention mechanism and the second attention mechanism respectively.

7. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, the CVT error prediction method based on step-by-step hierarchical extraction as described in any one of claims 1 to 4 is implemented.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the CVT error prediction method based on step-by-step hierarchical extraction as described in any one of claims 1 to 4 is implemented.

Images