Equipment state evaluation method and device for power distribution station house, equipment and storage medium

Abstract

The invention discloses an equipment state evaluation method and device of a power distribution station house, equipment and a storage medium, and relates to the technical field of power equipment, and the method comprises the steps: preprocessing the equipment data of secondary equipment of the power distribution station house to obtain a to-be-processed message, mapping the to-be-processed data into a continuous vector based on the discrete message type of the to-be-processed message to obtain a mapping result, performing position coding and feature extraction on a mapping result to obtain attention features, correcting the attention features by using a time decay mask, and compressing the attention features to obtain a state feature vector; using a preset encoder to perform trend feature extraction and reconstruction on the historical pre-test data to obtain a to-be-processed result, and performing feature reconstruction on the to-be-processed result to obtain a reconstructed feature vector; weight distribution is carried out on the state feature vector and the reconstruction feature vector, the state feature vector and the reconstruction feature vector are fused to obtain a fusion feature vector, state prediction is carried out to obtain an equipment state evaluation result, and the efficiency of carrying out state evaluation on the equipment is improved.

Description

Technical Field

[0001] This invention relates to the field of power equipment technology, and in particular to a method, apparatus, equipment and storage medium for assessing the equipment status of a power distribution station. Background Technology

[0002] Currently, the secondary equipment in power distribution substations is a crucial guarantee for the safe and stable operation of the power system, and the accuracy of its condition assessment directly affects the reliability of the power system. At present, condition assessments of secondary equipment in power distribution substations often rely on a single data source or simple data patching, which has the following shortcomings:

[0003] (1) Relying solely on communication sequence data ignores important information such as pre-test data and environmental data, making it difficult to fully reflect the true state of the equipment;

[0004] (2) The data processing method is simple and does not take into account the temporal characteristics of the data and the correlation between different data, resulting in low accuracy and reliability of the evaluation results;

[0005] (3) The lack of an effective feedback mechanism makes it impossible to optimize the evaluation model based on the actual maintenance results, which affects the model's adaptability and long-term effectiveness.

[0006] As can be seen from the above, how to improve the efficiency of equipment condition assessment in power distribution substations is an urgent problem to be solved. Summary of the Invention

[0007] In view of this, the purpose of this invention is to provide a method, apparatus, equipment, and storage medium for assessing the equipment status of a power distribution station, which can improve the efficiency of assessing the equipment status of the power distribution station, thereby enhancing the safety of the production process. The specific solution is as follows:

[0008] Firstly, this application provides a method for assessing the equipment status of a power distribution station, including:

[0009] Secondly, this application provides an equipment condition assessment device for a power distribution station, comprising:

[0010] Thirdly, this application provides an electronic device, comprising:

[0011] Memory, used to store computer programs;

[0012] A processor is used to execute the computer program to implement the aforementioned equipment status assessment method for a power distribution station.

[0013] Fourthly, this application provides a computer-readable storage medium for storing a computer program, wherein the computer program, when executed by a processor, implements the aforementioned equipment status assessment method for a power distribution station.

[0014] As can be seen from the above, before conducting equipment status assessment of the substation, this application needs to preprocess the equipment data corresponding to the secondary equipment of the substation, including historical pre-test data, real-time data of the secondary equipment, and environmental data, to obtain the messages to be processed. Then, using a preset event embedding layer and based on the discrete message type corresponding to the messages to be processed, the data to be processed is mapped into continuous vectors to obtain the mapping results. Next, a preset relative position encoding layer is used to perform position encoding on the mapping results to obtain the relative position encoding values ​​corresponding to each message to be processed. Finally, a preset multi-head self-attention layer is used to extract features from the relative position encoding values ​​to obtain attention features, and a time decay mask is used to correct the attention. The system first extracts force features, then compresses the attention features using the output layer to obtain a state feature vector. A preset encoder extracts and reconstructs trend features from historical test data to obtain the result to be processed. A preset decoder then reconstructs the result to obtain a reconstructed feature vector. The activation functions and hyperbolic tangent activation functions in the preset gating attention mechanism are used to assign weights to the state feature vector and the reconstructed feature vector. Based on the weight assignment results, the state feature vector and the reconstructed feature vector are fused to obtain a fused feature vector. A preset recurrent neural network is then used to predict the state based on the fused feature vector to obtain the equipment state evaluation result corresponding to the secondary equipment.

[0015] Therefore, this application first needs to preprocess the equipment data corresponding to the secondary equipment of the power distribution station, including historical pre-test data, real-time data of the secondary equipment, and environmental data, to obtain the messages to be processed. Then, using a preset event embedding layer and based on the discrete message type corresponding to the messages to be processed, the data to be processed is mapped into a continuous vector to obtain the mapping result. Next, a preset relative position encoding layer is used to position-encode the mapping result to obtain the relative position encoding value corresponding to each message to be processed. Secondly, a preset multi-head self-attention layer is used to extract features from the relative position encoding value to obtain attention features. A time decay mask is then used to correct the attention features, and finally, the input... The process involves: 1) compressing the attention features to obtain a state feature vector; 2) extracting and reconstructing trend features from historical test data using a pre-defined encoder to obtain the result to be processed; and 3) reconstructing the features of the result using a pre-defined decoder to obtain a reconstructed feature vector; and 4) weighting the state feature vector and the reconstructed feature vector using the activation functions and hyperbolic tangent activation function in a pre-defined gating attention mechanism, and then fusing the state feature vector and the reconstructed feature vector based on the weighting result to obtain a fused feature vector. A pre-defined recurrent neural network is then used to predict the state of the equipment based on this fused feature vector, resulting in an equipment state assessment result corresponding to the secondary equipment. This improves the efficiency of equipment state assessment in the power distribution station, thereby enhancing the safety of the production process. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0017] Figure 1 This application discloses a flowchart of a method for assessing the equipment status of a power distribution station.

[0018] Figure 2 This is a schematic diagram of the connection relationship of a specific Transformer structure for discrete data channel feature extraction disclosed in this application;

[0019] Figure 3 This is a schematic diagram of the connection relationship of a specific long short-term memory network structure for extracting time-series data channel features disclosed in this application;

[0020] Figure 4 This is a schematic diagram of the equipment status assessment device for a power distribution station disclosed in this application.

[0021] Figure 5 This is a structural diagram of an electronic device disclosed in this application. Detailed Implementation

[0022] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] Currently, secondary equipment in power distribution substations is a crucial guarantee for the safe and stable operation of power systems. However, condition assessments of this equipment often rely on single data sources or simple data splicing, which has the following shortcomings: it depends solely on communication sequence data, neglecting important information such as pre-test data and environmental data, making it difficult to comprehensively reflect the true state of the equipment; it fails to consider the temporal characteristics of the data and the correlation between different data points, resulting in low accuracy and reliability of the assessment results; and it lacks an effective feedback mechanism. Therefore, this application provides a method for assessing the condition of equipment in power distribution substations, which can improve the efficiency of condition assessment and thus enhance the safety of the production process.

[0024] See Figure 1 As shown in the figure, an embodiment of the present invention discloses a method for assessing the equipment status of a power distribution station, including:

[0025] Step S11: Preprocess the equipment data corresponding to the secondary equipment of the power distribution station, including historical pre-test data, real-time data of the secondary equipment and environmental data, to obtain the message to be processed. Then, use a preset event embedding layer and based on the discrete message type corresponding to the message to be processed, map the data to be processed into a continuous vector to obtain the mapping result. Then, use a preset relative position encoding layer to perform position encoding on the mapping result to obtain the relative position encoding value corresponding to each message to be processed.

[0026] In this embodiment, during the equipment status assessment of the substation, the present application embodiment first needs to collect historical pre-test data, real-time data of secondary equipment, and environmental data. Then, through data preprocessing, multi-source data is generated. In one specific implementation, the historical pre-test data is offline test data at the annual and monthly scales, including insulation resistance test, logic function correction record data, capacitance value, power ripple coefficient, logic function verification record, and other data. In another specific implementation, the real-time data of secondary equipment is a message sequence, including message type, timestamp, error code, and packet loss rate; the environmental data includes temperature and humidity data (obtained using temperature and humidity sensors), cabinet temperature rise curve data, and lightning strike alarm signal data (Boolean value: 0 indicates no warning, 1 indicates a warning).

[0027] It is worth mentioning that data preprocessing includes data cleaning, data transformation, data normalization, and environmental threshold filtering. Specifically, data cleaning removes outlier values ​​from sensors (such as temperature and humidity exceeding their physical range) and fills in missing values ​​using linear interpolation; data is converted to a unified timestamp format (millisecond-level Unix time) or a unified unit, such as standardizing the current unit if current is involved; data normalization uses Z-score standardization to eliminate dimensions; and environmental threshold filtering removes transient outliers in temperature and humidity (such as temperature fluctuations exceeding 5°C within 10 seconds).

[0028] Specifically, the equipment data corresponding to the secondary equipment in the substation, including historical pre-test data, real-time data of the secondary equipment, and environmental data, is preprocessed to obtain a message to be processed. This preprocessing can include: identifying the equipment data corresponding to the secondary equipment in the substation, including historical pre-test data, real-time data of the secondary equipment, and environmental data; where historical pre-test data is offline test data under a preset scale, including insulation resistance test data and logic function correction record data; real-time data of the secondary equipment is a message sequence including message type, timestamp, error code, and packet loss rate; environmental data includes temperature and humidity data, cabinet temperature rise curve data, and lightning strike alarm signal data; sensor anomaly removal is performed on the equipment data to obtain the removed data, and a preset interpolation algorithm is used to fill in missing values ​​in the removed data to obtain the filled data; the timestamp format of the filled data is standardized to obtain the unified data, and a preset standardization algorithm is used to perform dimension elimination on the unified data to obtain dimension-eliminated data; then, transient anomalies in temperature and humidity in the dimension-eliminated data are identified and removed to obtain the message to be processed.

[0029] In this embodiment, the discrete data channel feature extraction module adopts a Transformer structure, and its connection relationship is as follows: Figure 2As shown, it includes an event embedding layer, a relative position encoding layer, a multi-head self-attention layer, and an output layer; among which, the multi-head self-attention layer includes two multi-heads, namely the first multi-head and the second multi-head, so as to use this structure to extract message event features, capture hidden fault chains, so as to associate historical message events with the current state through the self-attention mechanism, generate a state feature vector, and thus accurately identify real-time abnormal patterns (such as message packet loss rate exceeding the threshold, error code mutation).

[0030] First, the multi-source data input event embedding layer is mapped to discrete message types to generate a continuous vector, i.e.:

[0031] ;

[0032] in, Represents a continuous vector. Indicates the event embedding layer, Indicates message Type number, This represents the embedding dimension, typically ranging from 128 to 256. express 3D real space.

[0033] In this embodiment, embedding multi-source data input events into the layer enables... , The message types of discrete messages are mapped to continuous vectors for feature extraction in subsequent steps.

[0034] Subsequently, in this embodiment of the application, the continuous vector is input into the relative position coding layer for position coding, and the relative position coding value between each message is obtained according to the parity of the embedding dimension.

[0035] In this embodiment, unlike the absolute position encoding of the traditional Transformer structure, the present application introduces relative position encoding, that is, sine and cosine calculations are performed according to the parity of the embedding dimension to obtain the relative position encoding value. The advantages of doing so are: it can adapt to the temporal characteristics of multi-source data and improve the encoding expression capability; it can conform to the message pattern of secondary devices and optimize the anomaly identification process.

[0036] In one specific implementation, this application embodiment needs to determine whether the embedding dimension is odd. If the embedding dimension is odd, the expression for calculating the relative position encoding value between each message is as follows:

[0037] ;

[0038] in, Indicates message With messages The relative positional encoding values ​​between them, where sin represents the sine function, Indicates message timestamp, Indicates message timestamp, Indicates a dimension index. This represents the frequency scaling factor.

[0039] If the embedding dimension is even, the expression for calculating the relative positional encoding values ​​between messages is as follows:

[0040] ;

[0041] in, This represents the cosine function.

[0042] In this embodiment, This represents the frequency scaling factor, which varies with the dimension index. As the exponent of the denominator increases, the frequency decreases (the period becomes longer); the odd / even dimension is encoded using sine and cosine respectively, allowing different dimensions to be indexed. Corresponding to different time scales, making small Capture millisecond-level message intervals, large Capture hourly intervals; the typical GOOSE message interval in a distribution substation is 2-4ms (small). (Scenario), while the equipment's historical pre-testing cycle is measured in months / years (large). In this scenario, parity coding can simultaneously cover both time scales, accurately depicting the temporal correlation between "sudden anomalies in real-time messages" and "deterioration of long-term pre-test data." Therefore, distinguishing between sine / cosine coding based on the parity of the embedding dimension is a design that aligns with the characteristics of multi-source time-series data from secondary equipment in power distribution stations. This not only accurately captures the multi-scale temporal patterns of message interactions but also provides an efficient and robust coding foundation for subsequent fusion modeling, ultimately improving the accuracy and practicality of state assessment. This approach is particularly suitable for scenarios like power systems that are sensitive to temporal correlations.

[0043] Specifically, a preset event embedding layer is used to map the data to be processed into a continuous vector based on the discrete message type corresponding to the message to be processed, obtaining a mapping result. Then, a preset relative position encoding layer is used to positionally encode the mapping result to obtain the relative position encoding value corresponding to each message to be processed. This can include: determining the discrete message type corresponding to the message to be processed, then using the preset event embedding layer to map the message to be processed into a continuous vector in a preset real number space based on the discrete message type and a preset embedding dimension, obtaining a mapping result; determining the parity of the preset embedding dimension, if the parity indicates that the preset embedding dimension is odd, then using a sine function and the preset relative position encoding layer, and based on... The frequency scaling factor is determined by the dimension index corresponding to each mapping result. Then, the relative position encoding value is obtained by encoding the mapping result with the relative position encoding value corresponding to each message to be processed based on the timestamp and frequency scaling factor corresponding to each mapping result. If the preset embedding dimension of the parity representation is even, the cosine function and the preset relative position encoding layer are used, and the frequency scaling factor is determined by the dimension index corresponding to each mapping result. Then, the relative position encoding value is obtained by encoding the mapping result with the relative position encoding value corresponding to each message to be processed based on the timestamp and frequency scaling factor corresponding to each mapping result. The value of the frequency scaling factor is positively correlated with the value of the dimension index.

[0044] Step S12: Use a preset multi-head self-attention layer to extract features from the relative position encoding value to obtain attention features, and use a time decay mask to correct the attention features. Then, use the output layer to compress the attention features to obtain a state feature vector.

[0045] In this embodiment, the relative position encoding values ​​between each message are input into the multi-head attention layer for feature extraction, so that each head focuses on communication quality, event type, temporal continuity, and error association. Through attention calculation, an attention score for the multi-head attention layer is generated, i.e.:

[0046] ;

[0047] in, This represents the attention score of the multi-head attention layer. Represents the attention function. Represents the query vector. Represents the key vector. Represents a value vector. Represents the normalization function. This represents the dimension of the multi-head attention layer. Indicates the scaling factor. The relative position encoding matrix of the message. Indicates transpose;

[0048] In this embodiment, in the multi-head self-attention layer, each head focuses on different patterns such as communication quality, event type, temporal continuity, and error association. Through key-value separation, the key matrix encodes only the event type, while the value matrix encodes message content such as length and error code, thereby obtaining the attention score of the multi-head attention layer, i.e.:

[0049] ;

[0050] S24. Based on the timestamps of each message, calculate the time decay coefficient between each message, i.e.:

[0051] ;

[0052] in, Indicates message With messages The time decay coefficient between them Represents an exponential function. This represents the attenuation rate hyperparameter, with a value range of 0.01 to 0.1.

[0053] In this embodiment, the purpose of calculating the time decay mask between each message is to strengthen the weight of recent events, thereby correcting the attention score of the multi-head attention layer and specifically solving the evaluation problem brought about by the characteristics of power system data.

[0054] It is worth mentioning that, in this embodiment of the application, a time decay mask needs to be generated based on the time decay coefficient between each message. The time decay mask is then multiplied element-wise with the attention score of the multi-head attention layer to obtain the corrected attention score, i.e.:

[0055] ;

[0056] in, This indicates the corrected attention score. Indicates the time decay mask. This indicates pixel-by-pixel multiplication.

[0057] In this embodiment, the advantage of using a time decay mask to correct the attention score is that it strengthens the temporal correlation and accurately captures the fault chain. Since faults in the secondary equipment of the substation are often caused by a "recent abnormal event chain", such as GOOSE packet loss → protection device malfunction → alarm chain reaction, and in the time decay mask, recent events ( (small) Close to 1, high weight, old events ( big) Approaching 0, the weight is low, thus highlighting the diagnostic value of recent events; at the same time, the time decay mask can amplify the weight of "sudden instructions" (between sudden instructions). Small, Larger data points are used to capture abnormal patterns of "intensive interactions in a short period of time," avoiding the long-term proportion of heartbeat messages masking fault signals, thus adapting to the sudden patterns of power messages; suppressing noise interference and improving assessment robustness; due to the presence of environmental noise at the substation site (such as false alarms caused by lightning interference, and temporary message anomalies caused by temperature and humidity fluctuations), if these noise events are long after the current assessment time ( (large), will be due to The attenuation is suppressed, thereby reducing interference with the current state assessment and filtering invalid data. At the same time, when the communication of secondary equipment is affected by electromagnetic interference, the message timestamp will fluctuate slightly, such as the normal 2ms interval becoming 3ms. The exponential decay characteristic of the time attenuation mask is not sensitive to such slight fluctuations, thereby avoiding the model from misjudging the equipment state due to slight time fluctuations and improving robustness.

[0058] Subsequently, in this embodiment of the application, the modified attention score input to the output layer needs to be compressed to obtain the state feature vector, that is:

[0059] ;

[0060] in, Represents the state feature vector. Indicates the dimension of the input sequence. Indicates the first A sequence of elements.

[0061] In this embodiment, the corrected attention score needs to be calculated along the sequence dimension. The mean is taken, thus compressing it into a state feature vector, i.e., the state feature vector. It is a feature vector that includes communication dispersion (such as the uniformity of GOOSE message distribution) and key event weights (event type attention).

[0062] Specifically, a preset multi-head self-attention layer is used to extract features from the relative position encoded values ​​to obtain attention features. These attention features are then corrected using a time decay mask. Finally, the output layer compresses the attention features to obtain a state feature vector. This compression can include: using the preset multi-head self-attention layer to determine the query vector, key vector, value vector, relative position encoding matrix, and preset scaling factor corresponding to the relative position encoded values; and then using the preset multi-head self-attention layer and based on its dimension, the query vector, key vector, value vector, relative position encoding matrix, and preset scaling factor to determine the attention score. The timestamps corresponding to the relative position encoded values ​​are used. Then, an exponential function is used to determine the time decay coefficient based on a preset decay rate hyperparameter and each timestamp. The time decay mask is determined using the time decay coefficient. A preset multi-head self-attention layer is used to multiply the time decay mask and the attention score pixel by pixel to obtain the corrected features. The output layer is used to compress the corrected features based on a preset sequence dimension to obtain the state feature vector. The state feature vector is a feature vector that includes communication dispersion and event weights. The preset multi-head self-attention layer is used to focus on the communication quality, event type, temporal continuity and error association of the relative position encoded values.

[0063] Step S13: Use a preset encoder to extract and reconstruct the trend features of the historical test data to obtain the result to be processed, and use a preset decoder to reconstruct the features of the result to be processed to obtain the reconstructed feature vector.

[0064] In one specific implementation, the time-series data channel feature extraction module adopts a long short-term memory network structure, and its connection relationship is as follows: Figure 3 As shown, it includes a two-layer long short-term memory network encoder and decoder. The decoder includes a long short-term memory network layer (LSTM layer) and a fully connected layer. Therefore, this structure can capture the long-term trend of historical test data and ultimately obtain the health status of the device.

[0065] That is, firstly, historical pre-test data is input into a two-layer long short-term memory network encoder to extract long-term trend features and generate hidden state vectors, i.e.:

[0066] ;

[0067] in, Represents the hidden state vector. This represents a two-layer long short-term memory network encoder. This represents a sequence of historical pre-test data.

[0068] In this embodiment, a dual-layer long short-term memory network encoder is used, with 128 neurons in each layer. It inputs historical pre-test data sequences and outputs hidden state vectors to quantify long-term trends (long-term dependencies in historical pre-test data), such as capacitance decay rate and insulation degradation rate. The hidden state vector is the output vector of the dual-layer long short-term memory network encoder, representing the compressed characteristics of the device's historical health status, covering slow-changing processes such as capacitor aging and insulation degradation.

[0069] Subsequently, the hidden state vector is input into the decoder, and the mean squared error loss function is used for feature reconstruction to generate a reconstructed feature vector, namely:

[0070] ;

[0071] in, This represents the mean squared error loss function. This represents the total number of historical pre-test data. Indicates the first Reconstructed feature vectors of historical pre-test data, Indicates the first The true value of the historical pre-test data.

[0072] Furthermore, the hidden state vector is reconstructed using a decoder to constrain the consistency of the feature space. It uses mean squared error as the loss function and outputs reconstructed historical prediction data, such as predicted insulation resistance values. The decoder consists of a Long Short-Term Memory (LSTM) layer and a fully connected layer. The LSTM layer restores temporal dependencies, and the fully connected layer maps dimensions. Through mean squared error constraints, it reconstructs prediction data that conforms to temporal patterns (such as the annual changes in viscosity and insulation resistance) to ensure the consistency of the feature space in the temporal dimension.

[0073] Specifically, a preset encoder is used to extract and reconstruct trend features from historical pre-test data to obtain the result to be processed. A preset decoder is then used to reconstruct features from the result to obtain a reconstructed feature vector. This process may include: using the long short-term network layer corresponding to the preset encoder in a preset long short-term memory network structure to restore the temporal dependencies of the historical pre-test data, obtaining a temporal dependency restoration result including long-term dependencies and hidden state vectors; and using the fully connected layer corresponding to the preset encoder to map the temporal dependency restoration result to obtain the result to be processed. Long-term dependencies include capacitance attenuation rate and insulation degradation rate. The hidden state vector is a compressed feature vector used to characterize the historical health status of secondary equipment. The total number of historical pre-test data, the historical reconstructed feature vector, and the true value of the historical pre-test data are determined. Then, the preset decoder and the mean squared error loss function are used to process the result to be processed based on the total number of historical pre-test data, the historical reconstructed feature vector, and the true value of the historical pre-test data to obtain the reconstructed feature vector.

[0074] Step S14: Use the activation function and hyperbolic tangent activation function in the preset gating attention mechanism to assign weights to the state feature vector and the reconstructed feature vector, and fuse the state feature vector and the reconstructed feature vector based on the weight assignment result to obtain a fused feature vector. Use a preset recurrent neural network to perform state prediction based on the fused feature vector to obtain the equipment state evaluation result corresponding to the secondary device.

[0075] In this embodiment, the state feature vector is dynamically allocated and the weights of the reconstructed feature vector are used by a gating attention mechanism. The fused feature vector is generated by weighted fusion and then input into a recurrent neural network for state prediction to generate the health status assessment result of the secondary device.

[0076] In this embodiment, the hidden state vector is fused with the state feature vector extracted from real-time data such as temperature and packet loss rate. During the fusion process, a gating attention mechanism is introduced to dynamically allocate weights. By weighted summation, a fused feature vector is generated, thereby solving the problem of unbalanced contribution of multi-source data and improving the accuracy and robustness of state assessment.

[0077] That is, firstly, a gating attention mechanism is used to assign weights to the state feature vector and the reconstructed feature vector of historical pre-test data, namely:

[0078] ;

[0079] in, Indicates the first Class feature vector weights, This represents the Sigmoid activation function, which maps scores to the interval [0, 1]. Let represent the attention score vector, and , Representing feature dimension, express The set of real numbers, This allows us to quantify the contribution of features to the current task. This represents the hyperbolic tangent activation function, which can capture complex relationships between features through nonlinear transformation. Represents a learnable matrix, and , express The set of real numbers, Indicates the first Class feature vectors Indicates the bias term. This indicates transpose.

[0080] In this embodiment, since the influence weights of historical pre-test data, real-time data of secondary equipment, and environmental data on equipment status change dynamically with the operating conditions, this embodiment of the application adopts a gating attention mechanism to automatically allocate weights according to the importance of each feature.

[0081] Subsequently, based on the weights of the state feature vector and the reconstructed feature vector from historical pretest data, a weighted sum is performed to obtain the fused feature vector, i.e.:

[0082] ;

[0083] in, Represents the fused feature vector. Indicates the number of feature types.

[0084] Specifically, the activation functions and hyperbolic tangent activation functions in the preset gating attention mechanism are used to assign weights to the state feature vector and the reconstructed feature vector. Based on the weight assignment results, the state feature vector and the reconstructed feature vector are fused to obtain the fused feature vector. This can include: processing a preset matrix, a preset set of real numbers, and a preset bias term using the hyperbolic tangent activation function to obtain an intermediate state result, and determining the feature dimension and task contribution corresponding to the state feature vector and the reconstructed feature vector; using the activation function and based on the intermediate state result, feature dimension, and task contribution, determining the weights corresponding to the state feature vector and the reconstructed feature vector respectively, and determining the number of feature types corresponding to each state feature vector and the reconstructed feature vector; and fusing the state feature vector and the reconstructed feature vector based on the number of feature types and each weight to obtain the fused feature vector.

[0085] In this embodiment, the fused feature vector is input into a recurrent neural network (GRU classifier) ​​for state prediction to obtain the health score (ranging from 0 to 100) and state (e.g., normal, abnormal, or severely abnormal) of the secondary equipment. Furthermore, to more accurately quantify the effectiveness of the state assessment of the secondary equipment in the substation, this embodiment uses the mean absolute error of the health score (MAE_HS) and the recall rate of severely abnormal states (Recall_SA) as evaluation metrics, as detailed below:

[0086] (1) Mean Absolute Error of Health Score (MAE_HS): The mean absolute deviation between the predicted health score and the actual health status, reflecting the accuracy of status quantification. The calculation formula is:

[0087] ;

[0088] in, Represents the total number of samples. Indicates the first The device health score predicted for each sample, and , Indicates the first The actual health status of the equipment corresponding to each sample is derived from expert annotations in the sample dataset, based on a comprehensive score of maintenance records, fault history, and test data.

[0089] (2) Serious Anomalous State Recall (Recall_SA): This represents the proportion of serious anomalous states correctly identified by the model, avoiding missed detections that could lead to accidents. The calculation formula is as follows:

[0090] ;

[0091] in, This indicates the number of truly serious anomalies that were correctly predicted. The closer the value is to 1, the less likely the method proposed in this invention is to miss serious anomalies, and the better it can avoid equipment failures or even power grid accidents caused by missed detections. This indicates the number of instances where an actual serious anomaly is predicted to occur under other conditions, in order to prevent power grid accidents caused by unannounced equipment failures.

[0092] Specifically, using a pre-defined recurrent neural network to predict the state based on the fused feature vector to obtain the device state assessment result corresponding to the secondary device can include: inputting the fused feature vector into the pre-defined recurrent neural network to obtain the device state assessment result corresponding to the secondary device; the device state assessment result includes the health score and device state corresponding to the secondary device; determining the actual device state corresponding to the secondary device, then using an absolute error function and determining the mean absolute deviation based on the device state assessment result and the actual device state, and processing the device state assessment result using a pre-defined recall rate determination formula to obtain the recall rate determination result; adjusting the parameters corresponding to the pre-defined recurrent neural network based on the mean absolute deviation and the recall rate determination result to obtain a new recurrent neural network, and using the new recurrent neural network to assess the device state of the secondary device to obtain a new device state assessment result.

[0093] As can be seen from the above, the embodiments of this application first need to preprocess the equipment data corresponding to the secondary equipment of the power distribution station, including historical pre-test data, real-time data of the secondary equipment, and environmental data, to obtain the message to be processed. Then, using a preset event embedding layer and based on the discrete message type corresponding to the message to be processed, the data to be processed is mapped into a continuous vector to obtain the mapping result. Next, a preset relative position encoding layer is used to position-encode the mapping result to obtain the relative position encoding value corresponding to each message to be processed. Then, a preset multi-head self-attention layer is used to extract features from the relative position encoding value to obtain attention features. The attention features are then corrected using a time decay mask, and then... The output layer compresses the attention features to obtain a state feature vector. Then, a pre-defined encoder extracts and reconstructs trend features from historical test data to obtain the result to be processed. A pre-defined decoder then reconstructs the feature vector from the result to obtain a reconstructed feature vector. Finally, the activation functions and hyperbolic tangent activation function in the pre-defined gating attention mechanism are used to assign weights to the state feature vector and the reconstructed feature vector. Based on the weight assignment result, the state feature vector and the reconstructed feature vector are fused to obtain a fused feature vector. A pre-defined recurrent neural network then uses this fused feature vector to predict the state of the equipment, obtaining the equipment state assessment result corresponding to the secondary equipment. This improves the efficiency of equipment state assessment in the power distribution station, thereby enhancing the safety of the production process.

[0094] Accordingly, see Figure 4 As shown, this application also provides an equipment condition assessment device for a power distribution station, comprising:

[0095] The mapping result determination module 11 is used to preprocess the equipment data corresponding to the secondary equipment of the power distribution station, including historical pre-test data, real-time data of the secondary equipment and environmental data, to obtain the message to be processed. Then, it uses a preset event embedding layer and based on the discrete message type corresponding to the message to be processed, it maps the data to be processed into a continuous vector to obtain the mapping result. Then, it uses a preset relative position encoding layer to perform position encoding on the mapping result to obtain the relative position encoding value corresponding to each message to be processed.

[0096] The feature compression module 12 is used to extract features from the relative position encoding value using a preset multi-head self-attention layer to obtain attention features, correct the attention features using a time decay mask, and then compress the attention features using an output layer to obtain a state feature vector.

[0097] The reconstructed feature vector determination module 13 is used to extract and reconstruct trend features from the historical test data using a preset encoder to obtain the result to be processed, and to reconstruct the features of the result to be processed using a preset decoder to obtain the reconstructed feature vector.

[0098] The equipment status assessment result determination module 14 is used to assign weights to the state feature vector and the reconstructed feature vector using the activation function and hyperbolic tangent activation function in the preset gating attention mechanism, and to fuse the state feature vector and the reconstructed feature vector based on the weight assignment result to obtain a fused feature vector, so as to use a preset recurrent neural network to perform state prediction based on the fused feature vector to obtain the equipment status assessment result corresponding to the secondary equipment.

[0099] In some specific embodiments, the mapping result determination module 11 may specifically include:

[0100] The equipment data determination unit is used to determine the equipment data corresponding to the secondary equipment of the power distribution station, including historical pre-test data, real-time data of the secondary equipment, and environmental data; wherein, the historical pre-test data is offline test data under a preset scale, including insulation resistance test data and logic function correction record data; the real-time data of the secondary equipment is a message sequence including message type, timestamp, error code, and packet loss rate; the environmental data includes temperature and humidity data, cabinet temperature rise curve data, and lightning strike alarm signal data;

[0101] The data outlier removal unit is used to perform sensor outlier removal operation on the device data to obtain the removed data, and to perform missing value filling operation on the removed data using a preset interpolation algorithm to obtain the filled data.

[0102] The data timestamp format unification unit is used to unify the timestamp format of the filled data to obtain unified data, and to perform a dimensionless operation on the unified data using a preset standardization algorithm to obtain dimensionless data. Then, it identifies and removes the transient abnormal values ​​of temperature and humidity in the dimensionless data to obtain the message to be processed.

[0103] In some specific embodiments, the mapping result determination module 11 may specifically include:

[0104] The mapping result determination unit is used to determine the discrete message type corresponding to the message to be processed, and then use a preset event embedding layer to map the message to be processed into a continuous vector in a preset real number space based on the discrete message type and the preset embedding dimension to obtain the mapping result.

[0105] The first frequency scaling factor determination unit is used to determine the parity corresponding to the preset embedding dimension. If the parity indicates that the preset embedding dimension is odd, then the frequency scaling factor is determined based on the dimension index corresponding to each of the mapping results using a sine function and a preset relative position encoding layer. Then, the relative position encoding value is obtained for the mapping result based on the timestamp corresponding to each of the mapping results and the frequency scaling factor to obtain the relative position encoding value corresponding to each of the messages to be processed.

[0106] The second frequency scaling factor determination unit is configured to, if the parity characterization of the preset embedding dimension is even, utilize the cosine function and the preset relative position encoding layer, and determine the frequency scaling factor based on the dimension index corresponding to each mapping result, and then perform relative position encoding on the mapping result based on the timestamp corresponding to each mapping result and the frequency scaling factor to obtain the relative position encoding value corresponding to each message to be processed; wherein, the value of the frequency scaling factor is positively correlated with the value of the dimension index.

[0107] In some specific embodiments, the feature compression module 12 may specifically include:

[0108] The attention score determination unit is used to determine the query vector, key vector, value vector, relative position encoding matrix and preset scaling factor corresponding to the relative position encoding value using a preset multi-head self-attention layer, and then determine the attention score using the preset multi-head self-attention layer and based on the dimension of the multi-head attention layer and based on the query vector, the key vector, the value vector, the relative position encoding matrix and the preset scaling factor.

[0109] The time decay mask determination unit is used to determine the timestamp corresponding to each of the relative position encoding values, and then use an exponential function and a preset decay rate hyperparameter and each of the timestamps to determine the time decay coefficient, so as to determine the time decay mask using the time decay coefficient.

[0110] The state feature vector determination unit is used to multiply the temporal decay mask and the attention score pixel by pixel using the preset multi-head self-attention layer to obtain the corrected features, and to compress the corrected features using the output layer based on a preset sequence dimension to obtain a state feature vector; the state feature vector is a feature vector including communication dispersion and event weight; the preset multi-head self-attention layer is used to focus on the communication quality, event type, temporal continuity and error association of the relative position encoding value.

[0111] In some specific embodiments, the reconstructed feature vector determination module 13 may specifically include:

[0112] The time-dependent reconstruction result generation unit is used to perform time-dependent reconstruction on the historical test data using the long short-term network layer corresponding to the preset encoder in the preset long short-term memory network structure, to obtain a time-dependent reconstruction result including long-term dependencies and hidden state vectors, and to map the time-dependent reconstruction result using the fully connected layer corresponding to the preset encoder to obtain the result to be processed; the long-term dependencies include capacitance decay rate and insulation degradation rate; the hidden state vector is a compressed feature vector used to characterize the historical health status of the secondary equipment;

[0113] The reconstructed feature vector determination subunit is used to determine the total number of historical pre-test data, the historical reconstructed feature vector, and the true value of the historical pre-test data corresponding to the historical pre-test data. Then, using a preset decoder and mean squared error loss function, the result to be processed is processed based on the total number of historical pre-test data, the historical reconstructed feature vector, and the true value of the historical pre-test data to obtain the reconstructed feature vector.

[0114] In some specific embodiments, the equipment status assessment result determination module 14 may specifically include:

[0115] The intermediate state result determination unit is used to process the preset matrix, the preset real number set and the preset bias term using the hyperbolic tangent activation function to obtain the intermediate state result, and to determine the feature dimension and task contribution corresponding to the state feature vector and the reconstructed feature vector.

[0116] The feature type number determination unit is used to determine the weights corresponding to the state feature vector and the reconstructed feature vector respectively by using the activation function and based on the intermediate state result, the feature dimension and the task contribution, and to determine the number of feature types corresponding to each state feature vector and the reconstructed feature vector;

[0117] The feature vector fusion determination unit is used to fuse the state feature vector and the reconstructed feature vector based on the number of feature types and each of the weights to obtain a fused feature vector.

[0118] In some specific embodiments, the equipment status assessment result determination module 14 may specifically include:

[0119] A fusion feature vector processing unit is used to input the fusion feature vector into a preset recurrent neural network to obtain a device status evaluation result corresponding to the secondary device; the device status evaluation result includes a health score and device status corresponding to the secondary device;

[0120] The recall rate determination result determination unit is used to determine the actual device state corresponding to the secondary device, then use the absolute error function and the device state evaluation result and the actual device state to determine the mean absolute deviation, and use the preset recall rate determination formula to process the device state evaluation result to obtain the recall rate determination result.

[0121] The neural network parameter adjustment unit is used to adjust the parameters corresponding to the preset recurrent neural network based on the mean absolute deviation and the recall rate determination result to obtain a new recurrent neural network, so as to use the new recurrent neural network to evaluate the equipment status of the secondary equipment and obtain a new equipment status evaluation result.

[0122] Furthermore, embodiments of this application also disclose an electronic device, Figure 5 This is a structural diagram of an electronic device 20 according to an exemplary embodiment. The content of the diagram should not be construed as limiting the scope of this application. The electronic device 20 may specifically include: at least one processor 21, at least one memory 22, a power supply 23, a communication interface 24, an input / output interface 25, and a communication bus 26. The memory 22 stores a computer program, which is loaded and executed by the processor 21 to implement the relevant steps in the equipment status assessment method for a power distribution station disclosed in any of the foregoing embodiments. Furthermore, the electronic device 20 in this embodiment may specifically be a computer.

[0123] In this embodiment, the power supply 23 is used to provide operating voltage for each hardware device on the electronic device 20; the communication interface 24 can create a data transmission channel between the electronic device 20 and external devices, and the communication protocol it follows can be any communication protocol applicable to the technical solution of this application, and is not specifically limited here; the input / output interface 25 is used to acquire external input data or output data to the outside world, and its specific interface type can be selected according to specific application needs, and is not specifically limited here.

[0124] In addition, the memory 22, as a carrier for resource storage, can be a read-only memory, random access memory, disk or optical disk, etc. The resources stored thereon can include operating system 221, computer program 222, etc., and the storage method can be temporary storage or permanent storage.

[0125] The operating system 221 is used to manage and control the various hardware devices on the electronic device 20 and the computer program 222, which may be Windows Server, Netware, Unix, Linux, etc. In addition to including a computer program capable of performing the equipment status assessment method for the substation disclosed in any of the foregoing embodiments, the computer program 222 may further include computer programs capable of performing other specific tasks.

[0126] Furthermore, this application also discloses a computer-readable storage medium for storing a computer program; wherein, when the computer program is executed by a processor, it implements the aforementioned equipment status assessment method for a power distribution station. Specific steps of this method can be found in the corresponding content disclosed in the foregoing embodiments, and will not be repeated here.

[0127] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to in the method section.

[0128] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0129] The steps of the methods or algorithms described in conjunction with the embodiments disclosed herein can be implemented directly by hardware, a software module executed by a processor, or a combination of both. The software module can be located in random access memory (RAM), main memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, hard disk, removable disk, CD-ROM, or any other form of storage medium known in the art.

[0130] Finally, it should be noted that in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0131] The technical solutions provided in this application have been described in detail above. Specific examples have been used to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this application. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A method for assessing the equipment condition of a substation, characterized in that, include: The equipment data corresponding to the secondary equipment of the power distribution station, including historical pre-test data, real-time data of the secondary equipment and environmental data, are preprocessed to obtain the message to be processed. The message to be processed is then mapped into a continuous vector using a preset event embedding layer based on the discrete message type corresponding to the message to be processed, and the mapping result is obtained. Then, the mapping result is position encoded using a preset relative position encoding layer to obtain the relative position encoding value corresponding to each message to be processed. The relative position encoding value is extracted using a preset multi-head self-attention layer to obtain attention features. The attention features are then corrected using a time decay mask. Finally, the attention features are compressed using an output layer to obtain a state feature vector. The historical test data is subjected to trend feature extraction and reconstruction using a preset encoder to obtain the result to be processed, and the result to be processed is subjected to feature reconstruction using a preset decoder to obtain the reconstructed feature vector. The state feature vector and the reconstructed feature vector are weighted using the activation function and hyperbolic tangent activation function in the preset gating attention mechanism. The state feature vector and the reconstructed feature vector are then fused based on the weighting result to obtain a fused feature vector. A preset recurrent neural network is then used to perform state prediction based on the fused feature vector to obtain the equipment state evaluation result corresponding to the secondary device.

2. The equipment condition assessment method for a power distribution station according to claim 1, characterized in that, The process involves preprocessing the equipment data corresponding to the secondary equipment in the substation, including historical pre-test data, real-time data of the secondary equipment, and environmental data, to obtain a message to be processed, including: The equipment data corresponding to the secondary equipment in the substation includes historical pre-test data, real-time data of the secondary equipment, and environmental data. The historical pre-test data is offline test data under a preset scale, including insulation resistance test data and logic function correction record data. The real-time data of the secondary equipment is a message sequence including message type, timestamp, error code, and packet loss rate. The environmental data includes temperature and humidity data, cabinet temperature rise curve data, and lightning strike alarm signal data. The device data is subjected to sensor outlier removal operation to obtain the removed data, and the missing value operation is performed on the removed data using a preset interpolation algorithm to obtain the filled data; The timestamp format of the filled data is unified to obtain unified data. A preset standardization algorithm is then used to perform a dimensionless operation on the unified data to obtain dimensionless data. Then, the transient abnormal values ​​of temperature and humidity in the dimensionless data are identified and removed to obtain the message to be processed.

3. The equipment status assessment method for a power distribution station according to claim 1, characterized in that, The process of mapping the data to be processed into a continuous vector using a preset event embedding layer and based on the discrete message type corresponding to the message to be processed, obtaining a mapping result, and then using a preset relative position encoding layer to perform position encoding on the mapping result to obtain a relative position encoding value corresponding to each message to be processed, includes: Determine the discrete message type corresponding to the message to be processed, and then use a preset event embedding layer to map the message to be processed into a continuous vector in a preset real number space based on the discrete message type and the preset embedding dimension to obtain the mapping result. Determine the parity corresponding to the preset embedding dimension. If the parity indicates that the preset embedding dimension is odd, then use a sine function and a preset relative position encoding layer, and determine the frequency scaling factor based on the dimension index corresponding to each of the mapping results. Then, based on the timestamp corresponding to each of the mapping results and the frequency scaling factor, perform relative position encoding on the mapping results to obtain the relative position encoding value corresponding to each of the messages to be processed. If the parity indicates that the preset embedding dimension is even, then the cosine function and the preset relative position encoding layer are used, and the frequency scaling factor is determined based on the dimension index corresponding to each of the mapping results. Then, the relative position encoding value of the mapping result is obtained based on the timestamp corresponding to each of the mapping results and the frequency scaling factor, so as to obtain the relative position encoding value corresponding to each of the messages to be processed. The numerical value of the frequency scaling factor is positively correlated with the numerical value of the dimension index.

4. The equipment status assessment method for a power distribution station according to claim 1, characterized in that, The process involves extracting features from the relative position encoded values ​​using a preset multi-head self-attention layer to obtain attention features, correcting these attention features using a time decay mask, and then compressing the attention features using an output layer to obtain a state feature vector, including: A preset multi-head self-attention layer is used to determine the query vector, key vector, value vector, relative position encoding matrix, and preset scaling factor corresponding to the relative position encoding value. Then, the preset multi-head self-attention layer is used, and based on the dimension of the multi-head attention layer, the query vector, the key vector, the value vector, the relative position encoding matrix, and the preset scaling factor, the attention score is determined. Determine the timestamp corresponding to each of the relative position encoding values, and then use an exponential function and a preset decay rate hyperparameter and each of the timestamps to determine a time decay coefficient, so as to determine a time decay mask using the time decay coefficient; The temporal decay mask and the attention score are multiplied pixel-by-pixel using the preset multi-head self-attention layer to obtain the corrected features. The corrected features are then compressed using the output layer based on a preset sequence dimension to obtain a state feature vector. The state feature vector is a feature vector that includes communication dispersion and event weights. The preset multi-head self-attention layer is used to focus on the communication quality, event type, temporal continuity, and error association of the relative position encoding value.

5. The equipment status assessment method for a power distribution station according to claim 1, characterized in that, The process involves using a preset encoder to extract and reconstruct trend features from the historical test data to obtain a result to be processed, and then using a preset decoder to reconstruct features from the result to obtain a reconstructed feature vector, including: The historical test data is temporally dependently restored using the long short-term memory network layer corresponding to the preset encoder in the preset long short-term memory network structure, resulting in a temporal dependency restoration result including long-term dependencies and hidden state vectors. The temporal dependency restoration result is then mapped using the fully connected layer corresponding to the preset encoder to obtain the result to be processed. The long-term dependencies include capacitance decay rate and insulation degradation rate. The hidden state vector is a compressed feature vector used to characterize the historical health status of the secondary equipment. The total number of historical pretest data, the historical reconstructed feature vector, and the true value of the historical pretest data are determined. Then, a preset decoder and mean squared error loss function are used to process the result to be processed based on the total number of historical pretest data, the historical reconstructed feature vector, and the true value of the historical pretest data to obtain the reconstructed feature vector.

6. The equipment condition assessment method for a power distribution station according to claim 1, characterized in that, The process involves weighting the state feature vector and the reconstructed feature vector using the activation functions and hyperbolic tangent activation function in a pre-defined gating attention mechanism, and then fusing the state feature vector and the reconstructed feature vector based on the weighting results to obtain a fused feature vector, including: The hyperbolic tangent activation function is used to process the preset matrix, the preset set of real numbers and the preset bias term to obtain the intermediate state result, and the feature dimension and task contribution corresponding to the state feature vector and the reconstructed feature vector are determined. The weights corresponding to the state feature vector and the reconstructed feature vector are determined by using the activation function and based on the intermediate state result, the feature dimension and the task contribution, and the number of feature types corresponding to each state feature vector and the reconstructed feature vector is determined. Based on the number of feature types and each of the weights, the state feature vector and the reconstructed feature vector are fused to obtain the fused feature vector.

7. The equipment condition assessment method for a substation according to any one of claims 1 to 6, characterized in that, The step of using a preset recurrent neural network to perform state prediction based on the fused feature vector to obtain the equipment state evaluation result corresponding to the secondary equipment includes: The fused feature vector is input into a preset recurrent neural network to obtain the device status assessment result corresponding to the secondary device; the device status assessment result includes the health score and device status corresponding to the secondary device; The actual device state corresponding to the secondary device is determined. Then, the absolute error function is used to determine the mean absolute deviation based on the device state evaluation result and the actual device state. The device state evaluation result is then processed using a preset recall rate determination formula to obtain the recall rate determination result. Based on the mean absolute deviation and the recall rate determination result, the parameters corresponding to the preset recurrent neural network are adjusted to obtain a new recurrent neural network. The new recurrent neural network is then used to evaluate the equipment status of the secondary equipment to obtain a new equipment status evaluation result.

8. A device for assessing the condition of equipment in a power distribution station, characterized in that, include: The mapping result determination module is used to preprocess the equipment data corresponding to the secondary equipment of the power distribution station, including historical pre-test data, real-time data of the secondary equipment and environmental data, to obtain the message to be processed. Then, it uses a preset event embedding layer and based on the discrete message type corresponding to the message to be processed, it maps the data to be processed into a continuous vector to obtain the mapping result. Finally, it uses a preset relative position encoding layer to perform position encoding on the mapping result to obtain the relative position encoding value corresponding to each message to be processed. The feature compression module is used to extract features from the relative position encoding value using a preset multi-head self-attention layer to obtain attention features, correct the attention features using a time decay mask, and then compress the attention features using an output layer to obtain a state feature vector. The reconstructed feature vector determination module is used to extract and reconstruct trend features from the historical test data using a preset encoder to obtain the result to be processed, and to reconstruct the features of the result to be processed using a preset decoder to obtain the reconstructed feature vector. The equipment status assessment result determination module is used to assign weights to the state feature vector and the reconstructed feature vector using the activation function and hyperbolic tangent activation function in the preset gating attention mechanism, and to fuse the state feature vector and the reconstructed feature vector based on the weight assignment result to obtain a fused feature vector, so as to use a preset recurrent neural network to perform state prediction based on the fused feature vector to obtain the equipment status assessment result corresponding to the secondary equipment.

9. An electronic device, characterized in that, include: Memory, used to store computer programs; A processor for executing the computer program to implement the equipment status assessment method for a power distribution station as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, Used to store a computer program, wherein the computer program, when executed by a processor, implements the equipment status assessment method for a power distribution room as described in any one of claims 1 to 7.

Images