Transform-based power system multi-modal data fusion toughness optimization method

Abstract

The invention relates to an electric power system multi-modal data fusion toughness optimization method based on Transform, and the method comprises the steps: carrying out the feature extraction of electric power data, meteorological data and regional data of a to-be-optimized electric power system, and obtaining a first sequence, a second sequence and a third sequence; standardizing and normalizing the first sequence, the second sequence and the third sequence; performing time alignment and interpolation on the three processed sequences; performing denoising processing on the first sequence, and performing smoothing processing on the second sequence; the three processed sequences are coded and spliced into an input vector, the input vector is input into the constructed Transform model, probability distribution of all toughness states is obtained, and the toughness state with the maximum probability serves as the toughness state of the power system to be optimized; and respectively making an optimization strategy for each toughness state, and implementing a corresponding optimization strategy on the to-be-optimized power system based on the toughness state of the to-be-optimized power system.

Description

Technical Field

[0001] This application relates to the technical field of power system resilience optimization, and particularly to a multi-modal data fusion resilience optimization method for power systems based on Transformer. Background Art

[0002] In the current power system field, with the continuous growth of power demand and the continuous expansion of the power grid scale, power systems are facing severe challenges. On the one hand, the operating environment of power systems is becoming increasingly complex. The interweaving of factors such as changes in meteorological conditions, differences in regional characteristics, and fluctuations in power loads has a profound impact on the stability and reliability of power systems. With the frequent occurrence of extreme weather events, problems such as damage to transmission lines and substation failures are caused, leading to large-scale power outages; due to different topographies and economic development levels in different regions, there are significant differences in the supply-demand balance and operating characteristics of power systems in different regions, and they are unevenly distributed in time and space.

[0003] On the other hand, traditional power system analysis and management methods have limitations in dealing with these complex problems. They often rely on a single data source or simple data processing methods, and it is difficult to comprehensively and accurately grasp the complex relationships among multiple multi-modal data in power systems. Only considering historical load data and ignoring the impact of meteorological factors on the load leads to inaccurate load forecasting, which in turn affects the power generation plan and scheduling arrangement of power systems; moreover, they are inefficient in processing large-scale and high-dimensional data, unable to extract valuable information from massive data in a timely and effective manner, and difficult to meet the needs of real-time monitoring and optimization decision-making of power systems. Summary of the Invention

[0004] Based on this, it is necessary to provide a multi-modal data fusion resilience optimization method for power systems based on Transformer, and this method includes: S1: Respectively perform feature extraction on the power data, meteorological data, and regional data of the power system to be optimized, and respectively obtain a first sequence of the time series pattern of the power data after outlier processing, a second sequence of the time series pattern of the meteorological data, and a third sequence of processing the spatial dependence based on the regional data matrix after dimensionality reduction processing; S2: Perform standardization and normalization on the first sequence, the second sequence, and the third sequence; perform time alignment and interpolation on the three processed sequences; perform denoising processing on the first sequence and smoothing processing on the second sequence; S3: Respectively encode the three sequences processed in S2, splice them into an input vector, input the input vector into the constructed Transformer model, obtain the probability distribution of each resilience state, and take the resilience state with the highest probability as the resilience state of the power system to be optimized; S4: Develop optimization strategies for each resilience state, and implement corresponding optimization strategies for the power system to be optimized based on its resilience state.

[0005] Beneficial effects: By performing multi-modal data fusion on the power data, meteorological data, and regional data of the power system, and using the Transformer model for analysis and prediction, this method can comprehensively understand the operating state of the power system, improving the resilience and reliability of the power system. Description of the Drawings

[0006] To more clearly illustrate the technical solutions in the embodiments of the present application or in the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the drawings in the following description are only some embodiments of the present application. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings.

[0007] Figure 1 It is a flowchart of the multi-modal data fusion resilience optimization method for the power system in the embodiments of the present application. Detailed Embodiments

[0008] To make the above objects, features, and advantages of the present application more obvious and understandable, the following will provide a detailed description of the specific embodiments of the present application with reference to the drawings. Many specific details are set forth in the following description to fully understand the present application. However, the present application can be implemented in many other ways different from those described herein. Those skilled in the art can make similar improvements without departing from the connotation of the present application. Therefore, the present application is not limited by the specific embodiments disclosed below.

[0009] In addition, the terms "first" and "second" are only used for descriptive purposes and cannot be construed as indicating or implying relative importance or implicitly specifying the quantity of the indicated technical features. Thus, the features defined with "first" and "second" may explicitly or implicitly include at least one of such features. In the description of the present application, "a plurality" means at least two, such as two, three, etc., unless otherwise specifically defined.

[0010] As Figure 1 shown, this embodiment provides a Transformer-based multi-modal data fusion resilience optimization method for a power system, which includes: S1: Respectively perform feature extraction on the power data, meteorological data, and regional data of the power system to be optimized, and respectively obtain a first sequence of the time series pattern of the power data after outlier processing, a second sequence of the time series pattern of the meteorological data, and a third sequence of processing the spatial dependence based on the regional data matrix after dimensionality reduction processing.

[0011] Specifically, the process of obtaining the first sequence includes: The power data includes daily load curve data, new energy data, and thermal power daily energy supply power curve data. One-day data of the three types of data is collected at 15-minute intervals to form a power data sequence; Calculate the mean and standard deviation of the power data sequence, and construct an outlier detection criterion based on the mean and standard deviation of the power data sequence; Regard the data points in the power data sequence that meet the outlier detection criterion as outliers; For the outliers, use the interpolation method of adjacent data points for correction to obtain the processed power data sequence; Input the processed power data sequence into the LSTM model to output the first sequence in the time series pattern.

[0012] In this embodiment, the feature extraction of the power data of the power system to be optimized further includes: Convert the power data sequence in the time domain into a frequency domain sequence through Fourier transform; Calculate the amplitude of the frequency domain sequence, and synthesize the amplitudes of the frequency domain sequences of each frequency to obtain a spectrogram; Find the frequencies with amplitudes greater than the set amplitude threshold; Calculate the periods corresponding to the selected frequencies based on the length of the power data sequence and the selected frequencies; Count the number of times each period appears in the period set, and regard the period with the most occurrences as the periodic component of the power data; Initialize the long-term trend sequence as an empty sequence, set the sliding window size, and calculate the first window sum according to the length of the power data sequence; Based on the current window size and the first window sum, calculate the average value of the power data corresponding to the current window, and add the average value to the long-term trend sequence; Slide the current window to the right, recalculate the first window sum and the average value corresponding to the window, and add all the calculated average values to the long-term trend sequence until the power data sequence is traversed to obtain the long-term trend sequence of the power data.

[0013] The calculation formula of the long-term trend sequence is: , k represents each window position, add all the power data values from k - W1 + 1 to k within the window, k = W1, W1 + 1,..., n, which is the index variable of the data within the window, where W1 represents the window size, represents the power data sequence The power data value at the time point i + j in, j is the index variable, and the value range is -W1 + 1 to 0, which is used to traverse each data point within the current window. T elec( k ) is the k-th element in the calculated long-term trend value sequence, T elec ( k ) is a one-dimensional array. Each element in the array is a floating-point number with the unit of megawatt (MW), which is consistent with the actual physical dimension of the power data. However, due to measurement device failures, electromagnetic interference, or other sudden factors, some data points may deviate from the normal range. If these outliers are not processed, they will interfere with the subsequent accurate understanding and analysis of the operation rules of the power system.

[0014] Furthermore, the process of obtaining the second sequence includes: The meteorological data includes temperature, humidity, and wind speed. Data for one day is collected at 15-minute intervals for these three types of data to form a meteorological data sequence; The temperature, humidity, and wind speed at the same moment in the meteorological data sequence are combined into a meteorological vector, and the meteorological vectors at different moments are connected in sequence to form a meteorological data sequence; In the meteorological data sequence, data for one day is collected at 15-minute intervals for temperature, humidity, and wind speed; The meteorological data sequence is input into the LSTM model, and a second sequence in the time series pattern is output.

[0015] In this embodiment, feature extraction of the meteorological data of the power system to be optimized further includes: In the meteorological data sequence, the mean, standard deviation, and maximum and minimum values are calculated for the temperature, humidity, and wind speed respectively; A time window is set, the second window sum is calculated based on the length of the meteorological data sequence, and the average feature of the temperature / humidity / wind speed corresponding to the current time window is calculated based on the current time window size and the second window sum.

[0016] Taking temperature as an example, the calculation formula is: ; where represents the average feature of the temperature in the i-th window, W3 represents the time window size, represents the temperature value at the j-th time point in the meteorological data sequence, j is an index variable, and its value range is from i - W3 + 1 to j, is the window sum, which is the sum of the temperature values within the current window.

[0017] In this embodiment, the LSTM model adopted is a commonly used model in the art, and its working process is the same as that of a conventional LSTM model.

[0018] Even further, the process of obtaining the third sequence includes: The regional data of each region includes the population quantity, economic activity indicators, geographical features, distribution of power infrastructure, electricity load density, and power generation resource reserves within the corresponding region; Represent the regional data of each region in matrix form, where each row represents the regional data of the corresponding region, and each column represents the population quantity / economic activity indicators / geographical features / distribution of power infrastructure / electricity load density / power generation resource reserves within the corresponding region, to obtain a regional data matrix; Calculate the covariance matrix of the regional data matrix; Solve the characteristic equation of the covariance matrix to obtain the eigenvalues and eigenvectors of the covariance matrix; Sort the eigenvalues and the corresponding eigenvectors in descending order of eigenvalues; Select the eigenvectors corresponding to the top k eigenvalues in order to form a projection matrix; Multiply the regional data matrix by the projection matrix to obtain the principal component matrix after dimensionality reduction; Take the principal component matrix after dimensionality reduction corresponding to each region as the nodes in the graph neural network, and use the power transmission relationship / geographical adjacency relationship / degree of economic connection between regions as the edges; The connection between regions is reflected by the edge to represent the power transmission relationship, geographical adjacency relationship, or degree of economic connection between region i and region j. Let the data of the edge be expressed as , where represents the value related to the power transmission relationship, represents the value related to the geographical adjacency relationship, represents the value related to the degree of economic connection. For the power transmission relationship, if there is a transmission line connection between region i and region j , then set the transmission capacity to (megawatts), and the formula is , where is the total transmission capacity of the system, and if there is no connection, then In terms of the geographical adjacency relationship, when adjacent, , and when not adjacent, , is the geographical distance between region i and j , in kilometers; in terms of the degree of economic connection, let the trade volume of region i with region j be (ten thousand yuan), and the formula is / , is the total trade volume of region i . In summary, it can provide a comprehensive data basis for GNN operations.

[0019] In the graph neural network, the spatial dependence relationship between nodes is mined through the message passing mechanism; Based on the spatial dependence relationship between the node and its adjacent nodes and the self-update function, the spatial dependence features of the node are obtained; the self-update function is + ; where represents the weight matrix in the self-update function, represents the new feature vector obtained after the node i is calculated by the self-update function, represents the feature vector of node i in the graph neural network, represents the bias vector in the self-update function.

[0020] The third sequence is obtained by integrating the spatial dependence features of all nodes.

[0021] S2: Standardize and normalize the first sequence, the second sequence, and the third sequence; perform time alignment and interpolation on the three processed sequences; perform denoising processing on the first sequence and smoothing processing on the second sequence.

[0022] In this embodiment, the processes of standardization and normalization are both common processes in the art.

[0023] Further, the time alignment and interpolation of the three processed sequences include: In the third sequence, when the time interval during regional data collection is inconsistent with the time interval during power data / meteorological data collection, approximate matching of the time points of the regional data is performed according to the statistical period of the power data / meteorological data, otherwise no special processing is performed to ensure that the power data, meteorological data, and regional data are collected at the same time step; When there are missing values or inconsistent time steps in the power data sequence / meteorological data sequence / regional data matrix, the linear interpolation method is used to calculate the estimated values of the missing time points based on the known time point values and time intervals.

[0024] Furthermore, the noise in the power data mainly comes from the errors of measurement devices, electromagnetic interference, and random fluctuations during the operation of the power system. Based on this noise, a linear system state space model is used, and the formula is , is the system state vector at time k, containing key information such as power load, and A is the state transition matrix that characterizes the system dynamics; reflects the influence of the control input; is the process noise with zero mean Gaussian distribution, and the covariance matrix is Q and the observation equation , is the observation vector, i.e., the measured power data, and H is the observation matrix. is the observation noise with zero-mean Gaussian distribution, and the covariance matrix is R. Through this linear system state-space model, the noise in the power data is incorporated into the system for consideration. Meteorological data often has abnormal fluctuations due to extreme weather, measurement errors, and rapid local meteorological changes. The LOESS smoothing method can be used to eliminate the fluctuations. For meteorological data points , is the time, is the meteorological variable value. At each , weights are assigned according to the surrounding points using a tricube function, and the formula is , where h is the bandwidth, which can control the size of the local area; then local weighted least squares regression is performed, starting from the linear function , where x represents the time, y represents the temperature, and the goal is to minimize the weighted sum of squared errors , and the coefficients and are obtained to minimize the weighted sum of squared errors. After obtaining the coefficients, the fitted values are calculated, and the LOESS smoothing is applied to meteorological variables such as temperature, humidity, and wind speed respectively to improve the data quality and enhance the effectiveness in the resilience analysis of the power system.

[0025] S3: Encode the three sequences processed in S2 respectively and splice them into an input vector. Input the input vector into the constructed Transformer model to obtain the probability distribution of each resilience state, and take the resilience state with the maximum probability as the resilience state of the power system to be optimized.

[0026] Specifically, the process of obtaining the input vector includes: Perform word embedding operation on the first sequence processed in S2 to obtain the first matrix; generate the first position encoding for the first matrix through sine and cosine functions; add the first position encoding and the first matrix element by element to obtain the first encoded vector; splice the periodic component, the long-term trend sequence, the first sequence, and the first encoded vector in the column direction to obtain the power data encoding; Perform word embedding operation on the second sequence processed in S2 to obtain the second matrix; generate the second position encoding for the second matrix through sine and cosine functions; add the second position encoding and the second matrix element by element to obtain the second encoded vector; splice the average features of temperature, humidity, wind speed, the second sequence, and the second encoded vector in the column direction to obtain the meteorological data encoding; Perform a spatial embedding operation on the third sequence after S2 processing to obtain a region vector; after converting the dimension-reduced principal component matrix and the spatial dependence features of the nodes to dimensions matching the region vector, splice them with the region vector to obtain a region data encoding; Splice the power data encoding, the meteorological data encoding, and the region data encoding to obtain the input vector.

[0027] Further, the Transformer model includes: a position encoding mechanism, an encoder, and a decoder; The position encoding mechanism includes a sine-cosine function and an adaptive position encoding; the sine-cosine function is used to generate a first position encoding / second position encoding; the adaptive position encoding is used to provide an adaptive weight factor for the first matrix / second matrix during the calculation of the sine-cosine function; The encoder fuses the input vector based on the multi-head attention mechanism to obtain a fused vector; The decoder maps the fused vector through a fully connected layer therein to obtain a fused feature vector; the Softmax function is used to convert the fused feature vector into a probability distribution of different toughness states.

[0028] The training process of the Transformer model includes: Process the collected historical power system data through steps S1 - S2, encoding, and splicing to obtain the input vector for training, and divide it into a training set, a validation set, and a test set according to the ratios of 70%, 15%, and 15% respectively; Before training, ensure that the dataset meets the model input requirements, then build the environment based on the deep learning framework, and enable GPU acceleration calculation to improve efficiency; In each iteration of the core link of model training, randomly extract a batch of samples (batch) from the training set. The first sample data taken out is the input vector, and then forward propagation calculation is performed on the input vector. The specific steps are: calculate the position encoding for each element according to the sine-cosine function formula, then perform adaptive position encoding, and evaluate the importance of power, meteorological, and regional modal data according to the adaptive mechanism to update the adaptive weight factor; After the completion of the position encoding, enter the calculation of the multi-head self-attention mechanism in the fusion layer, and use the linear transformation matrices W Q 、W K 、W V to map the input vector incorporating the position encoding into Q, K, V, and then normalize it through the Softmax function to obtain the original attention distribution weight matrix , then update the weight adjustment factor using the Adam algorithm, and adjust according to the updated factor to obtain Get Then, according to multiple set headers, splice and integrate different header outputs. Finally, through the output layer, map it to the probability distribution of different toughness states through the fully connected layer, and finally take the category with the highest probability as the predicted output of the model for the toughness state of the current sample power system.

[0029] Adopt the cross-entropy loss function based on Measure the true value And the predicted value The difference, where n is the number of samples in a batch, calculate the cross-entropy loss for each sample and average it to obtain the batch loss function value. Then, starting from the output layer, use the softmax function to map the feature vector to the probability distribution of the toughness state Calculate the output Of the fully connected layer Gradient And backpropagate it to the previous layer. Then, the gradient propagates to the multi-head self-attention mechanism of the fusion layer. The gradient of the fully connected layer Where Is the weight of the fully connected layer, and it propagates to the output splicing result of the multi-head self-attention mechanism. Let the output after splicing be The output of the h-th head is Then Where Is the matrix that distributes the gradient to the corresponding head output position. For the self-attention calculation process of each head, continue to calculate the gradient according to the chain rule, so as to obtain W Q 、W K 、W V Parameter gradients. Finally, the gradient is passed to the input layer (the initial input vectors of power, meteorological, and regional data encodings), and according to the data preprocessing and encoding operations, the input layer parameter gradients are calculated using the matrix derivative rule and the chain rule to prepare for parameter update.

[0030] After completing the gradient calculation, use the Adam optimization algorithm to update the model parameters according to the gradient. Adjust the weights and biases in the model according to the calculated gradient. At the same time, adopt the learning rate decay strategy, and perform the learning rate decay operation every certain number of training rounds (every 10 epochs). Let the current learning rate be And the decay factor is 0.95. As the training progresses, the learning rate gradually decreases, which enables the model to update the parameters with a finer step size in the later stage of training.

[0031] In the model evaluation stage, accuracy, recall, and F1-score are selected as the key indicators for evaluating the model performance. The accuracy is obtained by calculating the proportion of the number of samples with exactly the same prediction results and true labels in the total number of samples in the test set. Recall focuses on the model's prediction ability for positive samples, and is statistically calculated for high toughness, medium toughness, and low toughness respectively, which is the ratio of the number of samples with the actual category and being correctly predicted to the number of samples actually belonging to this category. The F1-score combines accuracy and recall, and calculates the scores for different toughness status categories according to the formula to provide an important reference basis for the model to be applied to the actual power system.

[0032] However, the toughness status of the power system does not exist in isolation, but is comprehensively affected by a variety of complex conditions. The following shows how different conditions affect the operation of the power system and play a role in the determination of the toughness status: meteorological conditions, power load, regional characteristics.

[0033] S4: Develop optimization strategies for each toughness status, and implement corresponding optimization strategies for the power system to be optimized based on its toughness status.

[0034] Specifically, the toughness status includes high toughness, medium toughness, and low toughness; For a power system with high toughness, since it has characteristics such as a powerful reserve capacity ≥ 20% of the system's maximum load, voltage fluctuations strictly controlled within ±3%, frequency deviation not exceeding ±0.05Hz, power supply reliability above 99.999%, and can quickly self-recover in an extremely short time (step 3.3.3), an optimization and maintenance strategy should be adopted.

[0035] When the toughness status of the power system to be optimized is high toughness, according to the equipment operation parameter rules reflected by the training set data during the training of the Transformer model, develop an inspection plan for the power system to be optimized, and optimize the power dispatching strategy based on the correlation between the power data and meteorological data learned during the training process, and obtain meteorological forecasts in advance based on the source of meteorological data.

[0036] For a power system with medium toughness, considering that although it can meet the basic electricity demand during normal operation, it requires manual intervention and a certain time to recover during complex and severe disturbances, with system voltage fluctuations of ±5% - ±8%, frequency deviation of ±0.1Hz - ±0.15Hz, power supply reliability of 99.5% - 99.9%, and users may experience short-term voltage instability or occasional power outages (step 3.3.3), a promotion and improvement strategy should be implemented.

[0037] When the resilience state of the power system to be optimized is medium resilience, analyze the equipment performance in the power system to be optimized, and conduct targeted upgrades on key equipment; introduce smart grid technology, and use the trained Transformer model to monitor the power load changes in real time and adjust the power generation and transmission strategies; simulate common fault scenarios to organize emergency drills and strengthen the professional skills training of operation and maintenance personnel.

[0038] A power system with low resilience is extremely fragile. Slight disturbances will cause performance deterioration, serious deterioration of key performance indicators, frequent power outages, long and uncertain recovery times, voltage fluctuations exceeding ±8%, frequency deviations exceeding ±0.15Hz, power supply reliability <99%, frequent power outages, and long and uncertain recovery times (Step 3.3.3), and a reconstruction and transformation strategy is required.

[0039] When the resilience state of the power system to be optimized is low resilience, transform the aging and backward power equipment and power transmission in the power system to be optimized; based on power data and regional data, and according to the regional development plan, promote distributed energy generation technology and construct a microgrid; re-plan the layout of the power system according to regional data and ensure that the transformation work meets regional needs and the power system development plan.

[0040] This method for optimizing the resilience of a power system based on Transformer provided in this embodiment extracts multi-modal data features from the power data, meteorological data, and regional data of the power system. The specific extraction content is the periodic components of the power data, long-term trend extraction, outlier detection and processing, and temporal pattern learning during power load learning. The extraction of meteorological data mainly includes three aspects: the basic source information of the data, the calculation of statistical features based on a time window, and the processing of the temporal characteristics of meteorological data using LSTM, to obtain feature information that can reflect the impact of meteorological factors on the power system. The regional extraction mainly includes three aspects: clarifying the original data, principal component analysis (PCA) dimensionality reduction processing, and using graph neural network (GNN) to process spatial dependence, to obtain information that can reflect the comprehensive characteristics and spatial dependence relationship of the region in the power system. Then, fusion preprocessing is carried out, specifically including standardization and normalization, time alignment and interpolation, denoising and smoothing. Then, a Transformer model is constructed. In the input encoding stage, various types of data are encoded and fused respectively. The fusion layer incorporates a multi-head self-attention mechanism, and the output layer is designed to map to the probability distribution of different resilience states according to the processed input data. After that, the collected historical data is divided into a training set, a validation set, and a test set according to a certain proportion, and then data training and optimization are carried out to enable the model to better master the complex relationships among power, meteorological, and regional data. Finally, based on the model prediction, resilience optimization strategies are implemented, and optimization and maintenance, improvement and enhancement, and reconstruction and transformation measures are taken for high, medium, and low resilience systems respectively to improve the resilience of the power system.

[0041] The technical features of the above-described embodiments can be combined arbitrarily. For the sake of brevity of description, not all possible combinations of the technical features in the above-described embodiments are described. However, as long as there is no contradiction in the combination of these technical features, it should be considered as falling within the scope described in this specification.

[0042] The above-described embodiments only represent several implementation manners of the present application. The description is relatively specific and detailed, but it should not be construed as a limitation on the scope of the patent application. It should be noted that for those of ordinary skill in the art, without departing from the concept of the present application, several modifications and improvements can still be made, and these all belong to the protection scope of the present application. Therefore, the protection scope of the patent of the present application shall be subject to the appended claims.

Claims

1. A Transformer-based multi-modal data fusion resilience optimization method for power systems, characterized in that: include: S1: extracting features from the power data, meteorological data and regional data of the power system to be optimized, respectively, and obtaining the first sequence of the time series pattern of the power data after outlier processing, the second sequence of the time series pattern of the meteorological data, and the third sequence of the spatial dependency based on the regional data matrix after dimensionality reduction processing; S2: standardizing and normalizing the first sequence, the second sequence, and the third sequence; performing time alignment and interpolation on the three processed sequences; performing denoising on the first sequence, and performing smoothing on the second sequence; S3: Encode the three sequences processed by S2 respectively and concatenate them into input vectors, input the input vectors into the constructed Transformer model, obtain the probability distribution of each resilience state, and take the resilience state with the highest probability as the resilience state of the power system to be optimized; S4: Develop optimization strategies for each resilience state, and implement corresponding optimization strategies for the power system to be optimized based on the resilience state of the power system to be optimized.

2. The method for optimizing resilience of multimodal data fusion in power system according to claim 1, characterized in that: The process of obtaining the first sequence includes: The power data includes daily load curve data, new energy data, and thermal power daily energy supply power curve data. All three types of data are collected at intervals of 15 minutes for one day to form a power data sequence; Calculate the mean and standard deviation of the power data series, and construct an outlier detection standard based on the mean and standard deviation of the power data series; The data points in the power data series that meet the outlier detection criteria are regarded as outliers; The abnormal value is corrected by using an interpolation method of adjacent data points to obtain a processed power data sequence; The processed power data sequence is input into the LSTM model, and a first sequence of time series patterns is output.

3. The method for optimizing resilience of multimodal data fusion in power system according to claim 2, characterized in that: Feature extraction of power data of the power system to be optimized also includes: The power data sequence in the time domain is converted into a frequency domain sequence by Fourier transform; Calculate the amplitude of the frequency domain sequence, combine the amplitude of the frequency domain sequence of each frequency, and obtain the spectrum diagram; Find the frequencies whose amplitude is greater than the set amplitude threshold; Calculate the period corresponding to the filtered frequency based on the length of the power data sequence and the filtered frequency; Count the number of times each cycle appears in the cycle set, and take the cycle with the largest number of occurrences as the periodic component of the power data; Initialize the long-term trend sequence to an empty sequence, set the sliding window size, and calculate the first window sum according to the length of the power data sequence; Based on the current window size and the sum of the first window, calculate the average value of the power data corresponding to the current window, and add the average value to the long-term trend sequence; Slide the current window to the right, recalculate the sum of the first window and the average value corresponding to the window, and add all calculated average values ​​to the long-term trend sequence until the power data sequence is traversed to obtain the long-term trend sequence of the power data.

4. The method for optimizing resilience of multimodal data fusion in power system according to claim 3 is characterized in that: The process of obtaining the second sequence includes: Meteorological data include temperature, humidity, and wind speed. All three types of data are collected at 15-minute intervals for one day to form a meteorological data sequence. The temperature, humidity and wind speed at the same time in the meteorological data sequence are combined into a meteorological vector, and the meteorological vectors at different times are sequentially connected to form a meteorological data sequence; In the meteorological data sequence, the temperature, humidity and wind speed are collected for one day at intervals of 15 minutes; The meteorological data sequence is input into the LSTM model, and a second sequence of time series patterns is output.

5. The method for optimizing resilience of multimodal data fusion in power system according to claim 4 is characterized in that: Feature extraction of meteorological data for the power system to be optimized also includes: In the meteorological data sequence, the mean and standard deviation of the temperature, the humidity and the wind speed are calculated and the maximum value is determined; Set the time window, calculate the second window sum based on the length of the meteorological data sequence, and calculate the average characteristics of temperature / humidity / wind speed corresponding to the current time window based on the current time window size and the second window sum.

6. The method for optimizing resilience of multimodal data fusion in power system according to claim 5, characterized in that: The process of obtaining the third sequence includes: The regional data of each region includes the population, economic activity indicators, geographical features, power infrastructure distribution, power load density, and power generation resource reserves in the corresponding region; The regional data of each region is represented in matrix form, where each row represents the regional data of the corresponding region, and each column represents the population / economic activity index / geographical features / power infrastructure distribution / power load density / power generation resource reserves in the corresponding region, to obtain a regional data matrix; Calculate the covariance matrix of the regional data matrix; Solve the characteristic equation of the covariance matrix to obtain the eigenvalues ​​and eigenvectors of the covariance matrix; Sort the eigenvalues ​​and corresponding eigenvectors in descending order; Select the eigenvectors corresponding to the first k eigenvalues ​​to form a projection matrix; Multiplying the regional data matrix by the projection matrix to obtain a principal component matrix after dimensionality reduction; The principal component matrices after dimensionality reduction corresponding to each region are used as nodes in the graph neural network, and the power transmission relationship / geographic proximity relationship / economic connection degree between regions are used as edges; In graph neural networks, the spatial dependencies between nodes are mined through the message passing mechanism; Based on the spatial dependency relationship between nodes and adjacent nodes and the self-update function, the spatial dependency characteristics of nodes are obtained; The third sequence is obtained by integrating the spatial dependence characteristics of all nodes.

7. The power system multimodal data fusion resilience optimization method according to claim 1 is characterized in that: The time alignment and interpolation of the three processed sequences comprises: In the third sequence, when the time interval for collecting regional data is inconsistent with the time interval for collecting power data / meteorological data, the time points of regional data are approximately matched according to the statistical period of power data / meteorological data, otherwise no special processing is performed to ensure that power data, meteorological data and regional data are collected at the same time step; When missing values ​​or inconsistent time steps occur in the power data series / meteorological data series / regional data matrix, linear interpolation is used to calculate the estimated values ​​of the missing time points based on the values ​​and time intervals of the known time points.

8. The method for optimizing resilience of multimodal data fusion in power system according to claim 6, characterized in that: The process of obtaining the input vector includes: Perform a word embedding operation on the first sequence processed by S2 to obtain a first matrix; generate a first position code for the first matrix through a sine and cosine function; add the first position code to the first matrix in a position-by-position manner to obtain a first coding vector; splice the periodic component, the long-term trend sequence, the first sequence and the first coding vector in a column direction to obtain a power data code; Perform a word embedding operation on the second sequence processed by S2 to obtain a second matrix; generate a second position code for the second matrix through a sine and cosine function; add the second position code to the second matrix in a position-by-position manner to obtain a second code vector; splice the average feature of temperature, the average feature of humidity, the average feature of wind speed, the second sequence and the second code vector in a column direction to obtain a meteorological data code; Performing a spatial embedding operation on the third sequence processed by S2 to obtain a regional vector; converting the spatial dependency features of the principal component matrix and nodes after dimensionality reduction into dimensions matching the regional vector, and then concatenating them with the regional vector to obtain a regional data code; The power data code, the meteorological data code, and the regional data code are concatenated to obtain the input vector.

9. The power system multimodal data fusion resilience optimization method according to claim 8 is characterized in that: The Transformer model includes: position encoding mechanism, encoder, and decoder; The position coding mechanism includes sine and cosine functions and adaptive position coding; the sine and cosine functions are used to generate the first position coding / the second position coding; the adaptive position coding is used to provide an adaptive weight factor for the first matrix / the second matrix when calculating the sine and cosine functions; The encoder fuses the input vectors based on a multi-head attention mechanism to obtain a fused vector; The decoder maps the fusion vector through a fully connected layer therein to obtain a fusion feature vector; and uses a Softmax function to convert the fusion feature vector into a probability distribution of different toughness states.

10. The power system multimodal data fusion resilience optimization method according to claim 1, characterized in that: The toughness state includes high toughness, medium toughness, and low toughness; When the resilience state of the power system to be optimized is high resilience, an inspection plan is formulated for the power system to be optimized according to the equipment operating parameter rules reflected by the training set data during the training of the Transformer model. The power dispatch strategy is optimized according to the correlation between the power data and meteorological data learned during the training process, and the weather forecast is obtained in advance based on the source of meteorological data. When the resilience state of the power system to be optimized is medium resilience, analyze the performance of the equipment in the power system to be optimized, and upgrade key equipment in a targeted manner; introduce smart grid technology, and use the trained Transformer model to monitor power load changes in real time and adjust power generation and transmission strategies; Organize emergency drills by simulating common failure scenarios and strengthen professional skills training for operation and maintenance personnel; When the resilience state of the power system to be optimized is low resilience, the aging and backward power equipment and transmission in the power system to be optimized will be continuously transformed; based on power data and regional data, and in accordance with regional development plans, distributed energy generation technology will be promoted and microgrids will be constructed; the power system layout will be replanned according to regional data, and it will be ensured that the transformation work meets regional needs and power system development plans.

Images