SF6 pressure early warning method based on self-adaptive dynamic combination

Abstract

The invention discloses an SF6 pressure early warning method based on self-adaptive dynamic combination, which overcomes the problems of low SF6 pressure prediction precision, large calculation amount and inapplicability to real-time monitoring in the prior art, and comprises the following steps: acquiring and processing SF6 gas pressure data to obtain a time sequence; dividing the time sequence into a plurality of time periods, mapping the original pressure data of each time period to a high-dimensional space through linear transformation, and adding position codes to obtain a feature mapping result; mapping results of each time period are interacted through adaptive dynamic multi-channel attention, an attention head combination mode is dynamically adjusted, and SF6 pressure prediction is carried out; and comparing a prediction result with standard reference data, and performing early warning according to a comparison result. By introducing an adaptive mechanism, the combination mode of the attention head is dynamically adjusted, the adaptive capacity and flexibility of diversified modes in input data are enhanced, and the gas pressure data processing efficiency and prediction accuracy are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of power equipment detection, and particularly relates to an SF6 pressure early warning method based on adaptive dynamic combination. Background Art

[0002] Sulfur hexafluoride (SF6), as an excellent insulating and arc extinguishing medium widely used in power equipment, has stable and superior performance. In substations, sulfur hexafluoride SF6 gas is crucial. With its high insulation and arc extinguishing performance, it improves the efficiency of power equipment and significantly reduces the volume of equipment. However, as a potent greenhouse gas, once SF6 leaks, it will not only cause non-negligible pollution to the environment but also directly threaten the health and safety of personnel. Its colorless and odorless characteristics increase the difficulty of leakage detection. Therefore, developing an efficient and accurate SF6 early warning analysis algorithm is of great significance for ensuring the stable operation of the power system.

[0003] Many existing prediction models (such as ARIMA, LSTM, etc.) have high complexity and require a large amount of computing resources and time for training and prediction, which are not suitable for real-time monitoring and early warning. Since the SF6 pressure sensor data in existing power equipment is often affected by environmental factors such as temperature and humidity, as well as sensor self-faults, etc., resulting in problems such as low data quality, many missing values, and high noise, traditional time series prediction methods perform limitedly when dealing with these complex data.

[0004] For example, the Chinese Patent Office published a patent on February 18, 2022: CN114065667A, a method for predicting the gas pressure of SF6 equipment based on the Prophet-LSTM model. The seasonal and trend components in the SF6 gas pressure data are extracted by the Prophet model, and this information is input into the LSTM to further capture the long-term and short-term dependencies of the data, which helps to judge the possible overheating faults and leakage faults in the equipment through the gas pressure value and play a role in early warning. Although the Prophet-based model can effectively capture the trends and seasonal components in the time series, its ability is relatively limited when dealing with long-term dependencies and high-dimensional data, which will affect the requirements of real-time prediction. Although LSTM performs excellently when dealing with long data sequence data, its performance highly depends on a large amount of historical data. In the case of insufficient or low-quality SF6 data in the power industry, the prediction effect will be greatly reduced. Combining the two models can, to a certain extent, make up for their respective deficiencies, but still lacks the comprehensive processing ability for complex time series data. Especially in the face of long-term dependencies and multi-dimensional data, its prediction accuracy is still not high, and further increases the calculation amount and reduces the data processing efficiency. Summary of the Invention

[0005] The purpose of the present invention is to solve the problems of large computational complexity, low data processing efficiency and low prediction accuracy of the SF6 pressure warning analysis algorithm in the prior art, and provide a SF6 pressure warning method based on adaptive dynamic combination, adopt an efficient time series prediction model PatchTST, and design an adaptive dynamic multi-channel attention mechanism ADMCA, which can flexibly adjust the combination strategy of each attention head according to the unique characteristics of the input data, focus on capturing specific time series features or complex associations across time periods, thereby achieving more sophisticated and efficient information processing, and improving the SF6 gas pressure data processing efficiency and the accuracy of SF6 gas pressure prediction.

[0006] In order to achieve the above object, the present invention adopts the following technical solutions: A GIS equipment SF6 pressure early warning method based on adaptive dynamic combination includes the following steps: S1: Collect and process SF6 gas pressure data to obtain a time series; S2: Divide the time series into several time periods, map the original pressure data of each time period into a high-dimensional space through linear transformation, and add position encoding to obtain feature mapping results; S3: The mapping results of each time period are interacted through adaptive dynamic multi-channel attention, and the combination of attention heads is dynamically adjusted to predict SF6 pressure; S4: Compare the prediction results with the standard reference data, analyze the SF6 change trend, and issue early warning.

[0007] The present invention utilizes ADMCA to remove the fixed attention head binding restriction in MHA, and introduces an adaptive mechanism to dynamically adjust the combination of attention heads so as to adaptively select the best combination of attention heads according to the input features, so that each attention head can be adaptively combined and weighted according to different data features, thereby enhancing the adaptability and flexibility to diversified patterns in the input data and improving the accuracy of SF6 pressure prediction.

[0008] Preferably, S3 includes: using a weight matrix to convert the feature mapping results of each time period into a query vector, a key vector and a value vector; using a gating mechanism and a unit matrix to calculate an adaptive weight matrix; using an adaptive weight matrix to dynamically adjust the combination of the query vector, the key vector and the value vector; calculating an attention weight matrix based on the similarity between the query and the key and the dimension of the query and the key; converting the attention weight matrix into a probability distribution, and performing weighted summation on the values to obtain a dynamically adjusted weight combination.

[0009] Preferably, the calculating the adaptive weight matrix by using the gating mechanism and the identity matrix includes: mapping the time-series input data to a gating signal by using a neural network; multiplying the gating signal by the identity matrix to obtain the adaptive weight matrix.

[0010] Preferably, S2 includes: given a set of multivariate time-series samples of length L, each element in the set being a vector of dimension d at time step t; dividing each input univariate time series into N time periods to obtain a time period sequence; mapping the time period sequence to a Transformer latent representation space of dimension D through linear projection and additive positional encoding to obtain the feature mapping result of each time period, and using it as the input of the Transformer layer.

[0011] Preferably, S1 includes adding historical weather data to the time series: adding the SF6 historical weather data to the original pressure data table as a new feature column, and then aligning the timestamps of the historical weather data with the hourly timestamps of the time-merged data, so that each time point corresponds to a complete input feature.

[0012] Preferably, according to the SF6 pressure standard value at 20°C, the adiabatic equation of the gas is used to calculate the pressure value at the target temperature to obtain the standard pressure reference data; calculating the difference between the standard pressure reference data and the SF6 pressure prediction result, and if the difference is greater than a preset deviation threshold, a warning is issued.

[0013] Preferably, the mean squared error loss is selected to measure the difference between the SF6 pressure prediction value and the true value, the loss of each channel is calculated, averaged over M time series to obtain the overall objective loss, and finally a flattening operation is performed through a fully connected layer to obtain the SF6 pressure prediction result.

[0014] Preferably, S1 includes data processing and enhancement of the pressure data: combining the nearest neighbor interpolation method, the linear interpolation method and the random forest interpolation method, and using a comprehensive filling strategy to fill the missing values; grouping the collected original pressure data at intervals of t time, merging them to the nearest hourly time point, and calculating the average value as the representative value to obtain the time series; performing standardization processing on the pressure data in the time series, calculating the coefficient of variation of the pressure data, and if the coefficient of variation is greater than the outlier detection threshold, using the nearest neighbor difference method for replacement.

[0015] Preferably, the comprehensive filling strategy includes: if the time interval between the missing value and the front and rear observed values is less than the first threshold, the nearest neighbor interpolation method is used to fill the missing value; if the missing interval is less than the second threshold and the data is stable or linearly changing, the linear interpolation method is used to fill the missing value; if the data fluctuation is greater than the third threshold, the random forest interpolation method is used to fill the missing value.

[0016] Preferably, the sigmoid activation function is used to map the time-series input data into a gating signal.

[0017] Therefore, the present invention has the following beneficial effects: by dividing the SF6 pressure data into several time periods (patches), mapping the original data of each patch into a high-dimensional space through linear transformation, and introducing positional encoding to retain the order information in the time series; then, the mapping results of each patch are interacted through the Adaptive Dynamic Multi-Channel Attention (ADMCA). The ADMCA introduces an adaptive mechanism to dynamically adjust the combination mode of attention heads, so as to adaptively select the best combination of attention heads according to the input features, enabling each attention head to perform adaptive combination and weighting according to different data features, thereby enhancing the adaptability and flexibility of the model to diverse patterns in the input data. Description of the Drawings

[0018] Figure 1 It is the overall step flow chart of the SF6 pressure warning method based on adaptive dynamic combination in the first embodiment.

[0019] Figure 2 It is the data map of the leakage point position of the SF6 gas pressure. Detailed Embodiments

[0020] The present invention will be further described in detail below in conjunction with the drawings and specific embodiments: First Embodiment: This embodiment provides a SF6 pressure warning method based on adaptive dynamic combination, as Figure 1 shown. Its operation process is as follows: Step 1, collect and process the SF6 gas pressure data to obtain a time series; Step 2, divide the time series into several time periods, map the original pressure data of each time period into a high-dimensional space through linear transformation, and add positional encoding to obtain a feature mapping result; Step 3, interact the mapping results of each time period through the Adaptive Dynamic Multi-Channel Attention to dynamically adjust the combination mode of attention heads for SF6 pressure prediction; Step 4, integrate the obtained calibrated data source code file in the controller to obtain the change point of the calibrated data value.

[0021] The SF6 pressure early warning method based on adaptive dynamic combination provided in this embodiment is committed to improving the processing efficiency of SF6 gas pressure data and the prediction accuracy of the model in the substation. By comprehensively integrating historical data records, real-time monitoring information, and various environmental factors, an outlier elimination strategy based on variability analysis is innovatively introduced. Further, this embodiment adopts an efficient time series prediction model - PatchTST, and designs an adaptive dynamic multi-channel attention mechanism ADMCA to construct an SF6 gas early warning analysis algorithm based on the ADMCA-PatchTST model. This algorithm deeply mines the hidden laws in the data and closely combines with cutting-edge machine learning technologies to achieve forward-looking early warning of SF6 gas leakage risks and accurate location determination, building a solid technical defense line for the stable and reliable operation of the power system.

[0022] Next, through specific examples and specific application scenarios, the technical solutions and technical effects of the present invention will be further described. The following examples are explanations of the present invention, and the present invention is not limited to the following examples.

[0023] Specifically, as Figure 1 shown, an SF6 pressure early warning method for GIS equipment based on adaptive dynamic combination specifically includes the following steps: The first step: Collect and process SF6 gas pressure data to obtain a time series.

[0024] In this embodiment, the collected SF6 gas pressure data is processed and enhanced to obtain a time series. The data processing and enhancement mainly include: (1) Combine the nearest neighbor interpolation method, linear interpolation method, and random forest interpolation method, and use a comprehensive filling strategy to fill in the missing values.

[0025] In the power industry, data quality faces severe challenges, especially the problem of missing values. In this embodiment, a comprehensive filling strategy is adopted, combining the nearest neighbor interpolation, linear interpolation, and random forest interpolation methods to fill in the missing values.

[0026] The comprehensive filling strategy is specifically manifested as: when the time interval between the missing value and the front and back observed values is less than a preset first threshold, the nearest neighbor interpolation method is preferentially used; for the case where the missing interval is less than a preset second threshold and the data is stable or linearly changing, the linear interpolation method is used; for the area where the data fluctuation is greater than a preset third threshold, the random forest interpolation method is introduced to accurately fill in complex missing values by using its powerful prediction ability.

[0027] (2) Time merging to obtain an initial time series.

[0028] To improve the efficiency and accuracy of data processing, this embodiment integrates the time merging technology. Specifically: the collected original pressure data is grouped at intervals of t (t is half an hour in this embodiment), merged to the nearest whole-hour time point, and the average value is calculated as the representative value to obtain the initial time series. This not only simplifies the data scale, but also improves the stability and reliability of the data through averaging. For the missing values generated after time merging, the filling strategy in step (1) is also adopted.

[0029] (3) Add historical weather data to the initial time series to obtain a time series.

[0030] To more comprehensively capture the factors affecting the change of SF6 pressure data, this embodiment innovatively introduces historical weather data as feature input. Historical weather data, mainly including temperature, humidity, rainfall, etc., has a significant impact on equipment operation status, energy production, and SF6 pressure change.

[0031] Introducing historical weather data as feature input provides more dimensional information for subsequent SF6 pressure prediction, which helps to capture the complex relationship between pressure data change and external environmental factors. Add SF6 historical weather data to the original data table as a new feature column, and then align the time stamps of historical weather data with the whole-hour time stamps merged in step (2), so that each time point corresponds to a complete input feature, and a time series is obtained.

[0032] (4) For the time series obtained in step (3), perform outlier processing based on variability analysis.

[0033] In response to the pressure data fluctuations caused by sensor failures or other external factors, this embodiment proposes an outlier processing method based on variability analysis. Specifically, it includes: (4.1) Data standardization processing.

[0034] For multi-dimensional SF6 pressure data, a standardization method is adopted to ensure that the SF6 pressure warning analysis model is not affected by the dimension difference during the training process.

[0035] The standardization process is calculated using the following formula: In the formula, x i is the original data point, μ is the mean of the data in this dimension, and σ is the standard deviation.

[0036] (4.2) Calculate the coefficient of variation.

[0037] Calculate the coefficient of variation of the data after standardization processing for measurement analysis. The coefficient of variation CV is the ratio of the standard deviation σ to the mean μ, which reflects the relative volatility of the data.

[0038] In time series data, the coefficient of variation can quantify the difference in volatility over different time periods, and the calculation method is as follows: When the coefficient of variation is greater than the preset outlier detection threshold, it indicates that the volatility of this segment of data is significantly higher than the normal level, and there may be anomalies.

[0039] According to the result of the variability measurement, this embodiment sets the outlier detection threshold CV th = 0.5. If the coefficient of variation exceeds the outlier detection threshold, it is considered that there is abnormal fluctuation in the data of this time period, and the data of this time period is replaced by using adjacent interpolation.

[0040] The outlier rejection method based on variability analysis combines the characteristic fluctuations of SF6 pressure data and can effectively identify and reject abnormal fluctuations. By dynamically measuring the standard deviation or coefficient of variation of the data, it can accurately reject the outliers caused by external interference or equipment failures, which not only improves the data quality, but also improves the prediction accuracy and robustness of the SF6 pressure warning algorithm.

[0041] The second step: Divide the time series into several time periods, map the original pressure data of each time period to a high-dimensional space through linear transformation, and add positional encoding to obtain the feature mapping result.

[0042] This embodiment provides an SF6 pressure warning method based on adaptive dynamic combination, constructs a time series prediction model, divides the historical SF6 pressure time series obtained by collection and data processing into a training set, a test set and a validation set according to the ratio of 7:2:1, and trains the time series prediction model.

[0043] In this embodiment, the time series prediction model mainly includes a data processing layer, a Transformer layer and a prediction head.

[0044] The second step is the work carried out in the data processing layer, mainly including slicing the time series, linear projection and positional encoding. First, divide the SF6 pressure data into several time periods (patches), map the original pressure data of each time period to a high-dimensional space through linear transformation, and introduce positional encoding to retain the order information in the time series.

[0045] The specific working process is as follows: Given a set of multivariate time series samples of length L: (x1, x2,..., x L ), where each x t ∈R d is a vector with dimension d at time step t, and the goal is to predict the future value at time T (xL+1 ,...,x L+T ).

[0046] Each input univariate time series is first divided into N patch chunks:

[0047] where P is the slice length and S is the non - overlapping region between two consecutive patches. The patching process generates a patch sequence x p (i) ∈R P×N , i = 1,..., N.

[0048] With the patching operation, the number of tokens input to the Transformer architecture is reduced from L to L / S, while reducing memory usage and computational complexity.

[0049] Through a trainable linear projection W p ∈R D×P and a learnable additive positional encoding W pos ∈R D×N , the patch sequence is mapped to a Transformer latent representation space of dimension D: x D (i) = W p x p (i) + W pos . Where x p (i) is the i - th patch, W p is the mapping matrix, W pos represents the positional encoding matrix, and x D (i) is the feature mapping result of the i - th patch, serving as the input to the Transformer layer.

[0050] Step 3: Interact the mapping results of each time period through adaptive dynamic multi - channel attention to dynamically adjust the combination method of attention heads.

[0051] The third step is the working process carried out in the Transformer layer, mainly including the Adaptive Dynamic Multi-Channel Attention of the Transformer encoder. The mapping results of each time period (patch) are interacted through the Adaptive Dynamic Multi-Channel Attention (ADMCA). Different from the traditional Multi-Head Attention (MHA), ADMCA introduces an adaptive mechanism to dynamically adjust the combination method of attention heads, so as to adaptively select the best combination of attention heads according to the input features. Specifically, ADMCA releases the fixed attention head binding limit in MHA, enabling each attention head to be adaptively combined and weighted according to different data features, thereby enhancing the model's adaptability and flexibility to diverse patterns in the input data.

[0052] The specific operations are as follows: First, convert the patch feature mapping results into a query vector Q i , a key vector K i and a value vector V i : Q i = x D (i) W Q , K i = x D (i) W K , V i = x D (i) W V .

[0053] Among them, W Q represents the trainable weight matrix of the query vector, W K represents the trainable weight matrix of the key vector, and W V is the trainable weight matrix of the value vector. Through these matrices, the representations of the query, key, and value for each patch input can be obtained d i,k is the dimension of each query and key.

[0054] In ADMCA, in this embodiment, an adaptive weight matrix W adaptive is calculated through a gating mechanism and an identity matrix, enabling the model to dynamically adjust the combination method between the query, key, and value according to the input features. The gating mechanism is used to control the weights between different attention heads, allowing the model to focus on different temporal features or cross-patch relationships. In the construction of the adaptive weight matrix, the range and influence of weight adjustment are controlled through the identity matrix to ensure the stability of the model.

[0055] Specifically manifested as: Step (1): First, map the time-series input data to a gating signal g through a neural network i , and in this embodiment, the activation function sigmoid is adopted, and its mapping relationship is as follows: g i = f(x i ) = σ(W g ·X i + b g ) Among them, X i represents the input feature vector with a dimension of D; W g ∈R D×1 is a trainable weight matrix used to linearly project the input feature X i into the scalar space; b g is used to translate the output of the linear transformation; σ is the activation function sigmoid that constrains the output value to a specific range [0, 1]. During the training process of the time-series prediction model, the network learns how to adjust the gating signal according to the input features so as to apply different weights to different attention heads.

[0056] Step (2): Use the obtained gating signal g i to dynamically adjust the weights of the query, key, and value. Specifically, multiply the gating signal g i by the identity matrix to obtain the adaptive weight matrix W adaptive (i), and the calculation formula is as follows: W adaptive (i) = g i ·I D +(1 + g i )·W learned Among them, I D is the identity matrix with a dimension of D, which is required to synchronize the dimension of the Transformer latent representation space. W learned is a trainable matrix in the model that can adaptively learn the optimal weight combination during the training process to enhance the model's ability to understand and express data.

[0057] Step (3): Through the combination of the gating mechanism and the identity matrix, the adaptive weight matrix W adaptive (i) can dynamically adjust the combination method of the query, key, and value: Q' i = W adaptive (i)Q i , K' i = W adaptive (i)K i , V' i = W adaptive (i)Vi 。

[0058] Step (4): Calculate the attention weight matrix:

[0059] where the similarity between the query and the key is calculated, d i，k is the dimension of the query and the key, and A is the attention weight matrix, representing the dependency relationship between each position and other positions.

[0060] Step (5): Apply the softmax function to transform the calculated attention weight A into a probability distribution, and perform a weighted sum on the value V′ to obtain the updated representation: where α i is the weight of the i-th element in the attention matrix. The process of dynamic weighting enables each attention head to focus on different key parts of the input data, thereby enhancing the expressive power of the model.

[0061] In ADMCA, the key point lies in adaptive dynamic combination. By introducing a dynamic combination weight matrix, the time series prediction model can dynamically adjust the combination method of attention heads according to the characteristics of the input data in each training step, so as to optimize the combination of different queries, keys, and values, thereby improving the expressive power of the time series prediction model.

[0062] Fourth step: Perform a flattening operation through a fully connected layer to obtain the SF6 pressure prediction result.

[0063] The fourth step is the working process carried out in the prediction head, mainly including predicting the SF6 pressure.

[0064] In this embodiment, the MSE loss is selected to measure the difference between the predicted value and the true value, and evaluate the performance of the constructed time series prediction model. The loss of each channel is Taking the average over M time series, the overall objective loss is:

[0065] Finally, perform a flattening operation through a fully connected layer to obtain the SF6 pressure prediction result.

[0066] Fifth step: Compare the SF6 pressure prediction result with the standard reference data, analyze the SF6 change trend, and issue a warning.

[0067] According to the SF6 standard pressure value at 20 °C, use the adiabatic equation of the gas to calculate the pressure value at the target temperature: Among them, P1 is the standard pressure value of SF6 gas at 20°C; T1 is 20°C, which is converted to 293.15K; T2 is the target temperature; R is the SF6 gas constant, taking an approximate value of 0.000828 KJ / (kg·K); C p is the value of specific heat capacity at atmospheric pressure. For SF6 gas, an approximate value of 0.930 KJ / (kg·K) is taken.

[0068] Taking the SF6 gas pressure value solved from the above adiabatic equation as the standard reference data, the deviation of the SF6 gas pressure can be expressed as: In the formula, P pred is the predicted value of SF6 gas pressure output by the above time series prediction model. If the deviation of the SF6 gas pressure is greater than the preset deviation threshold, an early warning is given.

[0069] A method for SF6 pressure early warning based on adaptive dynamic combination provided by this embodiment has the following beneficial effects: 1. In the ADMCA (Adaptive Dynamic Multi-Attention Combination) framework, an adaptive dynamic combination is adopted. In each training iteration, ADMCA can flexibly adjust the combination strategy of each attention head according to the unique features of the input data. This mechanism ensures that different attention heads can perform their respective functions, focusing on capturing specific time series features or complex cross-patch correlations, thus achieving more refined and efficient information processing.

[0070] 2. An outlier rejection strategy integrating variability analysis is proposed, which cleverly combines the unique fluctuation patterns of SF6 pressure data and can accurately identify and eliminate those abnormal fluctuations. By flexibly evaluating the standard deviation or coefficient of variation of the data, outliers caused by external environmental interference or equipment failures and other factors can be effectively identified and eliminated. This approach not only significantly improves the purity of the data set but also further enhances the prediction accuracy and robustness of the pressure early warning algorithm.

[0071] Embodiment 2: This embodiment provides a method for SF6 pressure early warning based on adaptive dynamic combination, which is applied to a specific scenario to implement and verify the method for SF6 pressure early warning based on adaptive dynamic combination in Embodiment 1.

[0072] SF6 gas has good insulation and arc extinguishing characteristics and is currently widely used in substation equipment. However, there is a common gas leakage problem in closed GIS (Gas Insulated Substation) equipment. Once leakage occurs, due to the action of high-voltage arcs, SF6 will decompose, generating some toxic substances, which will further cause the air in the GIS room to be oxygen-deficient and poisonous, threatening the safety of personnel. On the other hand, the leakage of SF6 gas will also reduce the insulation and arc extinguishing performance of the equipment. Therefore, it is necessary to take measures to achieve rapid and effective monitoring of the SF6 gas pressure to judge the leakage fault of the equipment.

[0073] Currently, the mainly adopted Transformer-based time series prediction method has great advantages in dealing with long-order time series prediction problems. However, Transformer faces quadratic time complexity and high memory usage problems when dealing with long sequences. Moreover, in the multi-head self-attention of the Transformer architecture, each attention head works independently, which will lead to problems such as low-rank bottlenecks and head redundancy in the attention score matrix.

[0074] Therefore, this embodiment provides an SF6 pressure warning method based on adaptive dynamic combination, aiming to improve the processing efficiency of SF6 gas pressure data and the prediction accuracy of the model in the substation. Through time series prediction technology, it can monitor and predict the pressure change of SF6 gas in GIS equipment in real time, so as to discover potential problems in time and carry out maintenance. Utilize the powerful time series modeling ability and the processing ability of complex time series data of the PatchTST model to achieve accurate prediction of SF6 pressure.

[0075] At the same time, in addition, most faults of primary substation equipment are manifested in the form of heat, and this heat release is all completed through SF6 gas. Inevitably, the change of SF6 gas pressure will be caused during such a heat exchange process. Therefore, real-time monitoring of SF6 gas pressure is also of great significance for analyzing overheating faults of equipment. Power enterprises can achieve real-time monitoring and intelligent warning of SF6 gas through the pressure warning method, so as to effectively manage these potential hazards and ensure the safe and stable operation of the power system.

[0076] Specifically: The SF6 pressure warning method based on adaptive dynamic combination provided by this embodiment first divides the SF6 data into several time periods (patches), maps the original data of each patch to a high-dimensional space through linear transformation, and introduces positional encoding to retain the order information in the time series. Then, the mapping results of each patch are interacted through Adaptive Dynamic Multi-Channel Attention (ADMCA). Different from the traditional multi-head attention (MHA), ADMCA introduces an adaptive mechanism to dynamically adjust the combination method of attention heads, so as to adaptively select the best combination of attention heads according to the input features. Specifically, ADMCA removes the fixed attention head binding restriction in MHA, enabling each attention head to perform adaptive combination and weighting according to different data features, thereby enhancing the adaptability and flexibility of the model to diverse patterns in the input data.

[0077] In the ADMCA, one of the core innovations is to adaptively weight the matrix through a gating mechanism and identity matrix calculation, enabling the model to dynamically adjust the combination of queries, keys, and values according to the input features. The gating mechanism is used to control the weights between different attention heads, allowing the model to focus on different temporal features or cross-patch relationships. In the construction of the adaptive weight matrix, the range and impact of weight adjustment are controlled by the identity matrix to ensure the stability of the model.

[0078] Finally, a flattening operation is performed through a fully connected layer to map the extracted features to the output space, obtaining the SF6 pressure prediction result.

[0079] Using the SF6 pressure warning method based on adaptive dynamic combination provided in this embodiment, the predicted SF6 gas pressure leakage point data is as Figure 2 shown.

[0080] According to the rule of thumb, for GIS equipment, the deviation threshold is set to 0.02 MPa in this embodiment. That is, if the SF6 gas pressure deviation exceeds 0.02 MPa (higher or lower than the standard value), it may indicate a leakage or other problems.

[0081] On the other hand, the prediction results obtained by the time series prediction model provided in this embodiment are compared with the prediction results of existing mainstream SF6 pressure prediction models, and the following data is obtained: model Mean Squared Error (MSE) Mean Absolute Error (MAE) Relative Standard Error (RSE) ARIMA 0.726 0.556 0.638 CNN + LSTM 0.663 0.498 0.612 Autoformer 0.360 0.372 0.550 PatchTST 0.345 0.369 0.538 ours 0.328 0.358 0.512 It can be seen that the mean square error, mean absolute error, and relative standard error of the prediction results obtained by the time series prediction model provided in this embodiment are all smaller than those of other existing models. Therefore, the time series prediction model provided in this embodiment enhances the prediction accuracy and robustness of SF6 pressure data.

[0082] This embodiment also provides a GIS equipment SF6 pressure warning system based on adaptive dynamic combination, including: A data acquisition module that acquires and processes SF6 gas pressure data to obtain a time series.

[0083] A time series prediction module that establishes a time series prediction model based on Adaptive Dynamic Multi-Channel Attention (ADMCA) to predict the SF6 pressure data of GIS equipment.

[0084] Among them, the time series prediction model includes a data processing layer, a Transformer layer, and a prediction head. The data processing layer is used to divide the time series into several time periods, map the original pressure data of each time period to a high-dimensional space through a linear transformation, and add positional encoding to obtain a feature mapping result; the Transformer layer is used to interact the mapping results of each time period through adaptive dynamic multi-channel attention, and dynamically adjust the combination method of attention heads in each training iteration; the prediction head is used to perform a flattening operation through a fully connected layer to obtain the SF6 pressure prediction result.

[0085] The data comparison module calculates the standard reference data, compares the prediction result of the time series prediction module with the standard reference data, analyzes the SF6 change trend, calculates the deviation between the SF6 gas pressure value output by the time series prediction model and the standard reference data, and alarms when the deviation exceeds 0.02 MPa (higher or lower than the standard value).

[0086] The above-described embodiments are only a preferred solution of the present invention, and do not impose any form of limitation on the present invention. There are other variations and modifications without exceeding the technical solutions described in the claims.

Claims

1. An SF6 pressure early warning method based on adaptive dynamic combination, characterized in that, include: S1: Collect and process SF6 gas pressure data to obtain a time series; S2: Divide the time series into several time periods, map the original pressure data of each time period into a high-dimensional space through linear transformation, and add position encoding to obtain feature mapping results; S3: The mapping results of each time period are interacted through adaptive dynamic multi-channel attention, and the combination of attention heads is dynamically adjusted to predict SF6 pressure; S4: Compare the prediction results with the standard reference data and issue an early warning based on the comparison results.

2. The SF6 pressure warning method based on adaptive dynamic combination according to claim 1, wherein The S3 includes: using a weight matrix to convert the feature mapping results of each time period into a query vector, a key vector and a value vector; using a gating mechanism and a unit matrix to calculate an adaptive weight matrix; using an adaptive weight matrix to dynamically adjust the combination of the query vector, the key vector and the value vector; calculating an attention weight matrix based on the similarity between the query and the key and the dimension of the query and the key; converting the attention weight matrix into a probability distribution, and performing weighted summation on the values to obtain a dynamically adjusted weight combination.

3. The SF6 pressure warning method based on adaptive dynamic combination according to claim 2, characterized in that The method of calculating the adaptive weight matrix using the gating mechanism and the unit matrix includes: mapping the time series input data into a gating signal using a neural network; and multiplying the gating signal with the unit matrix to obtain an adaptive weight matrix.

4. A method for SF6 pressure early warning based on adaptive dynamic combination according to claim 1 or 2 or 3, characterized in that, The S2 includes: given a multivariate time series sample set with a length of L, each element in the set is a vector with a dimension of d at a time step t; each input univariate time series is divided into N time periods to obtain a time period sequence; through linear projection and additive position coding, the time period sequence is mapped to a Transformer latent representation space with a dimension of D, and a feature mapping result of each time period is obtained, and it is used as the input of the Transformer layer.

5. The SF6 pressure warning method based on adaptive dynamic combination according to claim 1, wherein, The S1 includes adding historical weather data to the time series: adding the SF6 historical weather data to the original pressure data table as a new feature column, and then aligning the timestamp of the historical weather data and the timestamp of the time merge, so that each time point corresponds to a complete input feature.

6. A method for SF6 pressure early warning based on adaptive dynamic combination according to claim 1 or 2 or 3 or 5, characterized in that According to the standard value of SF6 pressure at 20°C, the adiabatic equation of the gas is used to calculate the pressure value at the target temperature to obtain the standard pressure reference data; the difference between the standard pressure reference data and the SF6 pressure prediction result is calculated, and if the difference is greater than the preset deviation threshold, an early warning is issued.

7. A method for SF6 pressure warning based on adaptive dynamic combination according to claim 1 or 2 or 3 or 5, characterized in that The mean square error loss is selected to measure the difference between the predicted SF6 pressure value and the true value, the loss of each channel is calculated, and the average is taken over M time series to obtain the overall target loss. Finally, a flattening operation is performed through the fully connected layer to obtain the SF6 pressure prediction result.

8. A method for SF6 pressure early warning based on adaptive dynamic combination according to claim 1 or 5, characterized in that, The S1 includes data processing and enhancement of pressure data: combining the nearest neighbor interpolation method, linear interpolation method, and random forest interpolation method, using a comprehensive filling strategy to fill in the missing values; grouping the collected original pressure data at intervals of t time, merging them to the nearest whole point time, and calculating the average value as the representative value to obtain a time series; performing standardization processing on the pressure data in the time series, calculating the coefficient of variation of the pressure data, and if the coefficient of variation is greater than the outlier detection threshold, using the nearest neighbor difference method for replacement.

9. The SF6 pressure warning method based on adaptive dynamic combination according to claim 8, wherein, The comprehensive filling strategy includes: if the time interval between the missing value and the previous and subsequent observed values is less than the first threshold, using the nearest neighbor interpolation method to fill in the missing value; if the missing interval is less than the second threshold and the data is stable or linearly changing, using the linear interpolation method to fill in the missing value; if the data fluctuation is greater than the third threshold, using the random forest interpolation method to fill in the missing value.

10. A method for SF6 pressure early warning based on adaptive dynamic combination according to claim 3, characterized in that, Using the sigmoid activation function to map the time series input data into a gating signal.

Images