An extreme weather electrical control method based on multi-source sensing fusion

Abstract

The application discloses an extreme weather electrical control method based on multi-source sensing fusion, which comprises the following steps: S1, collecting multi-source sensor data deployed in a power system to construct an original sensing data set; S2, performing time synchronization and frequency unification processing on the original sensing data set to generate a multi-source alignment data sequence; S3, extracting a fusion feature vector; S4, inputting the fusion feature vector into an Autoformer model to capture long-term time sequence dependence between multi-source features and output an extreme weather grade evaluation result and an electrical operation risk score vector; S5, inputting the risk score into an improved Gumbel-Softmax strategy selection network to output a control action sequence and parameter configuration; and S6, finally, converting the control strategy into a device control instruction and issuing it in real time. The method realizes risk prediction and adaptive control of an electrical system under extreme conditions.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of reinforcement learning and policy optimization control, and particularly relates to an extreme weather electrical control method based on multi-source sensor fusion. BACKGROUND

[0002] With the frequent occurrence of extreme weather events, power facilities face significant operational risks under complex weather conditions such as freezing, thunderstorms, and strong winds, and there is an urgent need for intelligent control methods with real-time sensing, rapid response, and adaptive regulation capabilities. However, existing electrical control systems mostly rely on preset rules or static threshold judgments, lack deep fusion modeling of multi-source sensor data and dynamic decision-making capabilities, resulting in lagging risk identification and rigid control strategies, making it difficult to meet the high requirements for electrical system stability and safety in extreme environments.

[0003] Currently, most control methods only collect single-type or low-frequency sensor data, failing to utilize multi-source heterogeneous information such as electrical parameters, visual images, infrared thermal imaging, and meteorological data, resulting in insufficient modeling dimensions for electrical equipment operating conditions and external risks, leading to one-sided evaluation results and low control strategy accuracy. At the same time, existing time series modeling methods often use general RNN or LSTM structures, which are difficult to capture long-term dependencies between multi-modal features, prone to trend neglect and time series drift problems, affecting the stability and reliability of risk prediction results.

[0004] In addition, in the control strategy generation process, discrete action mapping or static strategy tables are mostly used, lacking continuous strategy optimization means based on derivable sampling mechanisms, making it difficult to form smooth transitions and fine scheduling in complex control action spaces. In terms of strategy execution feedback, existing methods usually ignore the backflow evaluation and weight correction of historical execution results, and do not establish a perfect control feedback loop, making it difficult to support the continuous evolution and optimization of strategies, severely limiting the adaptability and robustness of the system.

[0005] Therefore, how to provide an extreme weather electrical control method based on multi-source sensor fusion is a problem that needs to be solved by those skilled in the art. SUMMARY

[0006] One purpose of the present application is to propose an extreme weather electrical control method based on multi-source sensor fusion, and the present application constructs a complete closed-loop process from multi-modal feature extraction, extreme weather recognition, risk score generation to intelligent control strategy selection. Through the time series decomposition structure and multi-head attention mechanism of the Autoformer model, the long-term dependence and trend change between the multi-source features such as meteorological, electrical and image are fully captured, and a more forward-looking electrical operation risk score vector is output. Further combined with the improved Gumbel-Softmax strategy selection network with feedback correction mechanism, the derivable sampling selection in the continuous control action set is realized, and the strategy sampling weight is dynamically optimized to improve the adaptability and response accuracy of the control action. The present application has the advantages of wide feature fusion dimension, strong time series modeling capability, high strategy generation precision, perfect feedback correction closed loop, etc., and can significantly improve the perception response efficiency, intelligent control ability and operation safety of the electrical system under extreme weather conditions, effectively solving the key problems such as insufficient perception range, rigid strategy selection and discontinuous response behavior in the existing control method.

[0007] According to an extreme weather electrical control method based on multi-source sensor fusion according to an embodiment of the present application, the following steps are included:

[0008] S1, collect multi-type sensor data sources, and construct an original sensor data set;

[0009] S2, time synchronization preprocessing is performed on the original sensor data set, timestamp information is extracted, a unified time axis is established, interpolation or downsampling processing is performed on different frequency data, and a multi-source aligned data sequence is generated;

[0010] S3, the multi-source aligned data sequence is standardized, numerical data is normalized, and image data is extracted state features through a convolutional neural network, and meteorological, electrical and visual information is fused into a fusion feature vector;

[0011] S4, the fusion feature vector is input into an Autoformer model, the decomposition structure based on the self-attention mechanism is used to capture the long-term time series dependence relationship between the multi-source features, and an extreme weather grade evaluation result and an electrical operation risk score vector are output;

[0012] S5, the electrical operation risk score vector is input into an improved Gumbel-Softmax strategy selection network, a continuous strategy combination is generated in the control action set, a feedback correction mechanism is introduced, the action sampling weight is weighted according to the execution feedback of the historical control action, and a control action sequence and parameter configuration are output;

[0013] S6, the control action sequence and parameter configuration are converted into device control instructions, and are transmitted to the target device in real time through the control bus.

[0014] Optionally, the multiple types of sensors specifically include meteorological sensors collecting temperature, humidity, and rainfall intensity data, electrical sensors collecting voltage, current, load rate, and frequency data, and visual sensors and infrared sensors collecting device surface images and thermal imaging information.

[0015] Optionally, S2 specifically includes:

[0016] S21, read each type of sensor data in the original sensor data set, extract the time stamp information corresponding to each data record, uniformly convert the time stamp into coordinated universal time format, determine the sampling frequency of all sensors, arbitrarily select a sensor as a reference channel, and establish a unified time axis based on the time interval of the reference channel;

[0017] S22, perform linear interpolation processing on sensor data with a sampling frequency lower than the reference channel, calculate the interpolation value between adjacent time points to fill in missing data, and perform sliding average downsampling on sensor data with a sampling frequency higher than the reference channel, replacing the average value of multiple consecutive sampling points with a representative value of the time period;

[0018] S23, sort image data according to time stamp, extract frame images closest to the time points of the unified time axis, and form a key frame sequence after time synchronization;

[0019] S24, combine all sensor data after interpolation, downsampling, and frame alignment processing in chronological order to generate a multi-source aligned data sequence under a unified time reference.

[0020] Optionally, S3 specifically includes:

[0021] S31, perform normalization processing on numerical data in the multi-source aligned data sequence, scale all numerical values to between 0 and 1 using the min-max normalization method, and preserve the relative change relationship of the original data;

[0022] S32, extract frame images in chronological order for image data and input them into a convolutional neural network, sequentially perform convolution, activation, and pooling operations with a preset size and step, and extract the state feature vector corresponding to the image;

[0023] S33, concatenate the normalized meteorological sensor data, electrical sensor data, and the extracted state feature vector corresponding to the image to construct a fusion feature vector corresponding to each time step.

[0024] Optionally, S4 specifically includes:

[0025] S41, arrange the fusion feature vector corresponding to each time step in chronological order to form a two-dimensional sequence matrix, set the window size, and perform sliding window slicing on the two-dimensional sequence matrix to obtain continuous input segments;

[0026] S42, perform a moving average difference operation on each feature dimension in each input segment in the time sequence direction, calculate the average change between adjacent time points, and decompose each sequence into a set of trend component sequences and residual component sequences;

[0027] S43, input the residual component sequence into the multi-head attention mechanism to perform attention distribution calculation of the preset weight between time steps, output the weighted sequence vector representing the periodic dependent feature, and reserve the trend component sequence as a bias channel;

[0028] S44, add the periodic weighted sequence vector and the trend component sequence by time step, and input them into the fully connected feedforward network to output the prediction vector after activation function and linear mapping;

[0029] S45, the end time step result in the current window prediction vector is used as the trend component sequence input for the next time period, and the trend superposition operation is recursively performed for continuous prediction of multiple time steps;

[0030] S46, the vector at the end time in the prediction vector output by the Autoformer model is divided into dimensions according to the task target structure in the training stage, the first part is mapped to the extreme weather grade by the softmax function, the value range is between 0 and 1, and the score value is divided into normal, general risk, high risk and very high risk four risk grades according to the preset interval;

[0031] S47, the second part is a one-dimensional electrical operation risk score vector, which corresponds to eight types of electrical risk factors: line voltage fluctuation, current peak overrun, power grid frequency deviation, power bus load imbalance, device surface icing, isolating switch contact aging, main transformer temperature rise anomaly and ground resistance anomaly. Each score is a real number between 0 and 1, representing the evaluation intensity of the risk type at the current time, which is used as the input basis for control action generation.

[0032] Optionally, the S5 specifically comprises:

[0033] S51, input the electrical operation risk score vector into the embedding layer of the improved Gumbel-Softmax strategy selection network to map the dimensions and generate a preset length of representation vector, match it with each action category vector in the predefined control action set, calculate the logarithmic distribution parameters of the sampling probability of each type of control action, and output the probability distribution parameters corresponding to the control action set;

[0034] S52, perform a derivable improved Gumbel-Softmax strategy selection network sampling operation on the sampling probability distribution of each type of action to generate continuous strategy vectors under each action category, and combine multiple strategy vectors to form a control action sequence.

[0035] S53, generating corresponding execution parameter configuration values for the control action sequence, the parameter values being floating-point numbers or integers related to the control action semantics, constituting a complete control action sequence and being output, the control actions including load limiting control, circuit breaker selection control, standby power supply enabling instructions, deicing heating device scheduling, or remote risk warning triggering.

[0036] Optionally, the improved Gumbel-Softmax policy selection network specifically comprises:

[0037] The electrical operation risk score vector is sent into a policy embedding layer in order of electrical risk factor categories to form an input vector, a nonlinear transformation and dimensionality operation are performed on each dimension of the risk score to map a single numerical value to a higher dimensional space through a set of pre-trained linear transformation weights, an embedding vector containing multiple dimensions is formed, a policy score matrix consistent with the size of the control action set is generated, each row of the matrix represents a control action, and each column is an embedded mapping value of the corresponding risk factor score in the policy space;

[0038] A standard Gumbel-Softmax sampling operation is performed on the policy score matrix, an independent random disturbance term is added to each control action, each random disturbance term is obtained by sampling once from a Gumbel distribution, the score corresponding to each control action is added to the disturbance term, a normalization operation is performed on each group of disturbed scores, the position score of each control action is converted into a continuous probability value, and a continuous policy distribution is constituted;

[0039] According to the continuous policy distribution, a derivable policy weight combination is generated, each control action corresponds to a floating-point value between 0 and 1, representing the possibility of the action being selected under the current risk state, and a continuous policy combination vector at the current time point is constituted;

[0040] The continuous policy combination vector is paired with the preset control action list item by item to form an action-weight pair set and perform a threshold judgment, if the weight value is higher than the set threshold, the action is included in the current candidate control action set, and the first preset number of control actions with the largest weight value are selected as the control plan of the current period, for each selected control action, the corresponding parameter configuration is queried to form a control action sequence and a corresponding parameter configuration list;

[0041] After each control period ends, the control action sequence and the corresponding feedback data that have been issued are collected, including: device response delay time, state change amount of the controlled object, execution stability index, and system alarm flag, after the original feedback data is normalized, a feedback feature vector of a set dimension is formed according to the control action number, and is one-to-one corresponding to the action sampling probability in the Gumbel-Softmax network;

[0042] The feedback feature vector is input into a scoring function, and the scores of the four types of indexes are calculated in turn according to the preset scoring rules: the response timeliness score of the response time, the execution effectiveness score of the state change amount, the execution stability score, and the safety score of the non-triggered protection mechanism.

[0043] The effect score of the single action is weighted and synthesized according to the preset weight to form a score vector, which is input into a feedback correction mechanism, the score value is normalized to form a correction coefficient, and the correction coefficient is multiplied with the sampling probability of the corresponding action in the original sampling probability vector to generate a new probability distribution vector, update the sampling weight of the next round of policy selection of the Gumbel-Softmax network, and synchronously update the forward propagation parameters, complete the policy adjustment and correction.

[0044] Optionally, the S6 specifically comprises:

[0045] S61, the control action sequence and parameter configuration are converted into device control instructions, which are real-time issued to the target device through the control bus, and each control action sequence and the corresponding parameter configuration item are combined into a standardized instruction package, the instruction package containing device number, action type, execution parameter and timestamp information;

[0046] S62, the instruction package is encoded and checked and then sent to the target terminal through the control bus, each device executes the corresponding action immediately after receiving the instruction, and returns the execution state and sensing feedback signal.

[0047] The beneficial effects of the application are:

[0048] The application fuses the Autoformer model and the improved Gumbel-Softmax policy selection network, and constructs an intelligent control process of an electrical system under extreme weather conditions, which has multiple beneficial effects such as strong multi-source sensing capability, continuous adjustable control strategy and high response efficiency.

[0049] Firstly, based on the multi-source sensing fusion mechanism, multiple types of data sources such as meteorological sensors, electrical measurement and control units and visual monitoring equipment deployed around power facilities are fully integrated to construct a unified sensing data system with multiple modalities, multiple frequencies and time sequence alignment, which effectively enhances the early identification capability of the system to extreme weather signals, improves the sensing breadth and data robustness, and provides high-quality data support for subsequent control strategy formulation.

[0050] Secondly, the introduced Autoformer model can model the long-term time sequence dependency relationship between different modal data through the decomposition block structure and the self-attention mechanism, can accurately predict the extreme weather grade change trend and the potential risk score of the electrical operation state, and effectively alleviate the information decay and feature loss problems of the traditional model in processing long sequence and heterogeneous data. The improved Gumbel-Softmax strategy selection network introduces a feedback correction mechanism on the basis of the original derivable sampling mechanism, can dynamically adjust the action sampling weight according to the execution result of the historical control action, realize the adaptive selection and real-time optimization of the continuous control strategy in different scenes, and improve the accuracy of the strategy output and the feasibility of the execution landing.

[0051] In addition, the control action sequence is converted into standardized control instructions and issued to the electrical equipment for execution, and the control closed loop is formed by collecting the equipment execution feedback, which can be further used for strategy iteration and model precision fine tuning, realizes the self-evolution of the system, and enhances the stability and robustness of the system operation.

[0052] In summary, the present application realizes the whole process closed loop management from multi-source sensing, intelligent prediction to continuous control strategy generation and optimization in the extreme weather scene, effectively solves the problems of insufficient information fusion, rigid control response and opaque strategy of the traditional method, and significantly improves the safety and intelligence level of the electrical system. BRIEF DESCRIPTION OF DRAWINGS

[0053] The accompanying drawings are included to provide a further understanding of the present application, and constitute a part of the specification, together with the embodiments of the present application, to explain the present application, and do not constitute a limitation on the present application. In the drawings:

[0054] Fig. 1 A flowchart of an extreme weather electrical control method based on multi-source sensing fusion is provided for the present application;

[0055] Fig. 2 A flowchart of multi-source time sequence feature decomposition and extreme weather risk level evaluation based on the Autoformer model is provided for the present application;

[0056] Fig. 3 A control action sampling and feedback correction optimization flowchart based on the improved Gumbel-Softmax strategy selection network is provided for the present application. DETAILED DESCRIPTION

[0057] The present application will now be described in further detail with reference to the accompanying drawings. These drawings are all simplified schematic diagrams, and only illustrate the basic structure of the present application in a schematic manner, and therefore only show the components related to the present application.

[0058] REFERENCE Figs. 1-3The application discloses an extreme weather electrical control method based on multi-source sensor fusion, and comprises the following steps:

[0059] S1, collecting multi-type sensor data sources, constructing a multi-modal, multi-frequency, time-stamped original sensor data set;

[0060] S2, time synchronization preprocessing of the original sensor data set, extracting timestamp information, establishing a unified time axis, interpolating or down-sampling different frequency data, and generating a multi-source alignment data sequence;

[0061] S3, standardizing the multi-source alignment data sequence, normalizing numerical data, and extracting state features from image data through a convolutional neural network, and fusing meteorological, electrical and visual information into a fusion feature vector;

[0062] S4, inputting the fusion feature vector into an Autoformer model, capturing long-term time sequence dependence between multi-source features based on the decomposition structure of the self-attention mechanism, and outputting an extreme weather grade evaluation result and an electrical operation risk score vector;

[0063] S5, inputting the electrical operation risk score vector into an improved Gumbel-Softmax strategy selection network, generating a continuous type policy combination in a control action set, simultaneously introducing a feedback correction mechanism, weighting the action sampling weight according to the execution feedback of the historical control action, and outputting a control action sequence and parameter configuration, wherein the control action at least includes: load limiting control, circuit breaker selection control, standby power supply enabling instruction, deicing heating equipment scheduling or remote risk early warning triggering;

[0064] S6, converting the control action sequence and parameter configuration into device control instructions, and issuing them to the target device in real time through the control bus, and collecting the execution feedback for subsequent policy iteration and control optimization.

[0065] The embodiment can significantly improve the electrical control response efficiency and strategy intelligence under extreme weather scenarios. By collecting multiple types of sensor data sources, a multi-modal data set with timestamps is constructed, and frequency alignment and standardization processing are performed to ensure that meteorological, electrical and visual information can be uniformly modeled and time series processed. On this basis, the self-attention decomposition structure of the Autoformer model is used to accurately model the time series dependence of various features, output the extreme weather level and electrical risk score, and effectively improve the pre-identification ability of sudden risks. Further, the risk score result is input to the improved Gumbel-Softmax strategy selection network, and a continuous control action combination is generated through differentiable sampling, covering key response operations such as limiting, circuit breaker control, backup power activation, deicing heating, and introducing a weight adjustment mechanism based on historical feedback to continuously optimize the action sampling strategy. Through this mechanism, the control strategy can be adaptively evolved dynamically under complex and variable extreme weather, significantly improving the robustness and emergency response capability of the system in the face of sudden load impact, equipment freezing, and voltage fluctuations. The embodiment has the advantages of strong multi-source data fusion, high risk identification accuracy, high intelligent degree of control strategy generation, and strong response stability, and has good practical deployment value and engineering applicability.

[0066] In the embodiment, the multiple types of sensors specifically include meteorological sensors collecting temperature, humidity and rainfall intensity data, electrical sensors collecting voltage, current, load rate and frequency data, and visual sensors and infrared sensors collecting device surface images and thermal imaging information.

[0067] In the embodiment, S2 specifically includes:

[0068] S21, read various sensor data in the original sensor data set, extract the timestamp information corresponding to each data record, and uniformly convert the timestamp to coordinated universal time format, determine the sampling frequency of all sensors, and select any sensor as a reference channel, and establish a unified time axis with the time interval of the reference channel as the standard;

[0069] S22, performing linear interpolation processing on sensor data with a sampling frequency lower than the reference channel, calculating interpolation values between adjacent time points to fill in missing data, and performing sliding average downsampling on sensor data with a sampling frequency higher than the reference channel, replacing the average value of multiple consecutive sampling points with a representative value for the time period;

[0070] S23, sorting image data according to timestamps, extracting frame images closest to the time points of the unified time axis, and forming a key frame sequence after time synchronization;

[0071] S24, combine all sensor data after interpolation, down-sampling and frame alignment processing in time sequence to generate multi-source aligned data sequence under unified time reference.

[0072] In the embodiment, the S3 specifically includes:

[0073] S31, perform normalization processing on numerical data in the multi-source aligned data sequence, and scale all numbers to 0-1 using the minimum-maximum normalization method to preserve the relative change relationship of the original data;

[0074] S32, input frame images of image data in time sequence into a convolutional neural network, and sequentially perform convolution, activation and pooling operations with preset size and step length to extract state feature vectors corresponding to the images, wherein the state features are used to reflect the visual state of the device surface;

[0075] S33, splice the normalized meteorological sensor data, electrical sensor data and the extracted state feature vectors corresponding to the images to construct a fusion feature vector corresponding to each time step as the input of the subsequent model.

[0076] In the embodiment, the S4 specifically includes:

[0077] S41, arrange the fusion feature vector corresponding to each time step in time sequence to form a two-dimensional sequence matrix, set a window size, and perform sliding window slicing on the two-dimensional sequence matrix to obtain continuous input segments;

[0078] S42, perform a moving average difference operation on each feature dimension in each input segment in time sequence direction to calculate the average change between adjacent time points, and decompose each sequence into a group of trend component sequences and residual component sequences;

[0079] S43, input the residual component sequence into a multi-head attention mechanism to perform attention distribution calculation with a preset weight between time steps, output a weighted sequence vector representing periodic dependent features, and separately reserve the trend component sequence as a bias channel;

[0080] S44, add the periodic weighted sequence vector and the trend component sequence by time step, and input them into a fully connected feedforward network to output a prediction vector after activation function and linear mapping;

[0081] S45, use the last time step result in the current window prediction vector as the trend component sequence input for the next time period, and recursively perform trend superposition operation for continuous prediction of multiple time steps;

[0082] S46, the vector of the last time in the prediction vector finally output by the Autoformer model is dimensionally divided according to the task target structure in the training stage, the first part is mapped to the extreme weather grade by the softmax function, the value range is 0 to 1, and the score value is divided into four risk grades of normal, general risk, higher risk and extremely high risk according to a preset interval;

[0083] S47, the second part is a one-dimensional electrical operation risk score vector, corresponding to eight types of electrical risk factors: line voltage fluctuation, current peak overrun, power grid frequency deviation, power bus load imbalance, equipment surface icing, disconnecting switch contact aging, main transformer temperature rise anomaly and ground resistance anomaly, each score is a real number between 0 and 1, indicating the evaluation intensity of the risk type at the current time, which is used as the input basis for generating control actions.

[0084] The embodiment introduces an Autoformer time series prediction model based on trend and cycle decomposition mechanism, effectively enhancing the forward-looking evaluation capability of extreme weather grade and electrical operation risk. Through sliding window slicing and moving average difference operation, the input feature sequence is decomposed into trend and residual components, and further the multi-head attention mechanism is used to model the time-dependent features, improving the capture effect of periodic risk changes. In the recursive prediction process, the trend superposition realizes the continuous prediction output of multiple time steps, enhancing the stability and extensibility of the model. The model can synchronously output the extreme weather grade evaluation result and the refined electrical risk score vector, covering eight typical indicators such as line voltage, current peak, power grid frequency, load balancing, equipment icing, contact aging, temperature rise anomaly and ground resistance. Through risk grade quantification and risk factor sub-item scoring, precise input basis is provided for subsequent control action generation and early warning response, significantly improving the timeliness and operability of meteorological anomaly and power safety hazard identification.

[0085] In the embodiment, the S5 specifically includes:

[0086] S51, input the electrical operation risk score vector into the embedding layer of the improved Gumbel-Softmax strategy selection network for dimension mapping, generate a preset length of representation vector, match with each action category vector in the pre-defined control action set, calculate the logarithmic distribution parameters of the sampling probability of each type of control action, and output the probability distribution parameters corresponding to the control action set;

[0087] S52, perform a derivable improved Gumbel-Softmax strategy selection network sampling operation on the sampling probability distribution of each type of action to generate continuous strategy vectors under each action category, and combine multiple strategy vectors to form a control action sequence;

[0088] S53, generating corresponding execution parameter configuration values for the control action sequence, the parameter values being floating-point numbers or integers related to the control action semantics, constituting a complete control action sequence and being output, the control actions including load limiting control, circuit breaker selection control, standby power supply enabling instruction, deicing heating device scheduling, or remote risk warning triggering.

[0089] The embodiment significantly improves the accuracy of control policy generation and the ability to correct feedback in an electrical risk situation by introducing an improved Gumbel-Softmax policy selection network. For the risk score vector, the network first completes high-dimensional embedding expression and constructs an action-risk matching score matrix, generates continuous policy combinations combined with the Gumbel sampling mechanism, and effectively realizes the parallel generation of multiple control actions. The system can execute action filtering and parameter configuration according to the policy weight, adapt to the multi-strategy linkage control demand in complex scenarios. At the same time, the feedback feature vector is formed by converting the execution result of each round, fusing the response timeliness, execution effectiveness, stability and safety, etc. The multi-dimensional score results form a correction coefficient feedback to the sampling probability distribution, realizing the dynamic adaptive optimization of the control policy. This mechanism not only maintains the differentiability of policy selection, which is suitable for deep network iterative training, but also has good feedback loop and parameter self-updating ability in control behavior adjustment, which fully improves the control response accuracy and robustness in extreme weather, electrical abnormalities and other complex states.

[0090] In the embodiment, the improved Gumbel-Softmax policy selection network specifically includes:

[0091] The electrical operation risk score vector is sent to the strategy embedding layer as an input vector in the order of electrical risk factor categories. Nonlinear transformation and dimensionality operation are performed on each dimension of the risk score to map a single numerical value to a higher dimensional space through a set of pre-trained linear transformation weights, forming an embedding vector containing multiple dimensions. A policy score matrix consistent with the size of the control action set is generated, with each row representing a control action and each column representing the embedded mapping value of the corresponding risk factor score in the policy space.

[0092] The standard Gumbel-Softmax sampling operation is performed on the policy score matrix, and an independent random disturbance term is added to each control action. Each random disturbance term is obtained by sampling once from the Gumbel distribution. The score corresponding to each control action is added to the disturbance term, and normalization is performed on each group of disturbed scores to convert the position score of each control action to a continuous probability value, forming a continuous policy distribution.

[0093] According to the continuous strategy distribution, a derivable strategy weight combination is generated, each control action corresponds to a floating-point value between 0 and 1, representing the possibility of the action being selected in the current risk state, forming a continuous strategy combination vector at the current time point;

[0094] The continuous strategy combination vector is matched with the preset control action list item by item to form an action-weight pair set and perform threshold judgment. If the weight value is higher than the set threshold, the action is included in the current candidate control action set, and the first preset number of control actions with the maximum weight value are selected as the control plan in the current period. For each selected control action, the corresponding parameter configuration is queried to form a control action sequence and a corresponding parameter configuration list;

[0095] After each control period ends, the control action sequence and the corresponding feedback data are collected, including: device response delay time, state change amount of the controlled object, execution stability index and system alarm flag. After the original feedback data is normalized, a feedback feature vector of a set dimension is formed according to the control action number, and is one-to-one corresponding to the sampling probability of each action in the Gumbel-Softmax network;

[0096] Each feedback feature vector is input into a scoring function, and the scores of four types of indexes are calculated in turn according to the preset scoring rules: response time response timeliness score, state change amount execution effectiveness score, execution stability score, and safety score of not triggering the protection mechanism;

[0097] The effect score of a single action is weighted and synthesized according to the preset weight to form a score vector, which is input into the feedback correction mechanism. The score value is normalized to form a correction coefficient, and is multiplied by the sampling probability of the corresponding action in the original sampling probability vector to generate a new probability distribution vector. The sampling weight of the next round of strategy selection of the Gumbel-Softmax network is updated, and the forward propagation parameters are updated synchronously to complete the strategy adjustment and correction.

[0098] In the embodiment, the S6 specifically includes:

[0099] S61, the control action sequence and the parameter configuration are converted into device control instructions, which are real-time issued to the target device through the control bus. Each control action sequence and the corresponding parameter configuration item are combined into a standardized instruction package, which includes device number, action type, execution parameter and timestamp information;

[0100] S62, after the instruction package is encoded and checked, it is sent to the target terminal through the control bus. Each device immediately executes the corresponding action after receiving the instruction, and returns the execution state and the sensing feedback signal, which are used for system monitoring and subsequent strategy optimization, realizing the real-time response process of instruction issuing, execution feedback and control closed loop.

[0101] Example 1

[0102] To verify the feasibility and practical value of the application in optimizing electrical control strategies under extreme weather conditions, the application is deployed in a 110kV regional power distribution station group of a large urban energy security hub center in a certain province. The region includes 6 main substations and more than 40 sets of distributed sensing terminals covering key devices such as transformers, circuit breakers, line protection equipment, and backup power units. There are also more than 100 sets of meteorological micro stations, anemometers, rain gauges, humidity and temperature composite probes, and front-end cameras deployed. The region where the city is located has an average of 18 extreme weather events per year, including thunderstorms, strong winds, short-term heavy rain, high temperature heat waves, and freezing rain and ice, causing electrical equipment to fluctuate, overload, and non-planned switching events, with an average of more than 100 events per year. The traditional control system has a lagging response, rigid strategy, and insufficient early warning, which seriously threatens the stability of power supply and equipment life.

[0103] After the application is deployed, the edge gateway collects multi-modal sensing data from wind speed, rainfall, temperature and humidity, load current, voltage disturbance images, etc., and performs interpolation alignment and normalization processing to generate structured multi-source input vectors. The fusion vector is input into the Autoformer model, which uses its self-attention decomposition structure to capture long-term time series dependencies and predict extreme weather grade indicators such as wind grade and electrical risk score within the next 30 minutes. The risk score vector is input into the improved Gumbel-Softmax strategy selection network. The network uses a differentiable sampling mechanism to output an action combination probability vector and introduces a feedback correction module to adjust the sampling weights based on the error, abnormal number, and action effect of the past 5 rounds of strategy execution results, achieving dynamic generation of optimal control combination strategies.

[0104] In actual operation tests, three high-risk months were selected for comparative evaluation of 19 typical extreme weather processes, covering thunderstorms, heavy rain, high temperatures, and strong winds. Evaluation indicators include electrical load peak shaving effect, action response time, backup power source calling accuracy, abnormal voltage control success rate, and control instruction false trigger rate. Table 1 below is a performance comparison table of the application method and existing dispatching control strategies in actual operation, including detailed indicators and real values:

[0105] Table 1 Performance comparison table of the application method and the traditional control system

[0106]

[0107] According to the comparison data in Table 1, it can be seen that the electrical control method based on Autoformer and improved Gumbel-Softmax strategy selection network proposed in the present application has significant advantages in electrical response control under extreme weather conditions compared with the traditional strategy. In terms of control action response time, the present application can complete action execution in an average of only 12.3 seconds, which is much lower than the 38.4 seconds of the traditional method, achieving nearly three times faster response, especially in high-risk scenarios such as thunderstorms and short-term heavy rain. In terms of control action hit rate, the average of the present application method is more than 93%, compared with the traditional average of about 68%, significantly improving the specificity and accuracy of the control strategy.

[0108] In terms of abnormal voltage recovery and standby power supply calling accuracy, the present application significantly improves the control effect by feature extraction and risk scoring reasoning on multi-source meteorological and electrical data, with accuracy improved to more than 90%, indicating that the system has stronger response capability in high-risk periods. In addition, the false trigger rate is reduced from more than 10% of the traditional method to less than 2%, effectively avoiding misoperation and resource waste. At the same time, the number of manual interventions is reduced from 7 to 9 times a day to 1 to 2 times in the entire deployment cycle, and the intelligence and autonomous control level of the system is greatly improved.

[0109] Overall, combining the indicators in the table, it can be confirmed that the present application method significantly improves the response speed, action matching degree, control stability and system robustness in actual extreme weather control scenarios, providing effective support for intelligent regulation and fault emergency of power systems. Especially in the extreme scenarios of freezing rain and icing and high temperature heat wave, the feedback correction mechanism of the strategy error shows strong adaptive ability, which is an important supplement and breakthrough to the lack of timeliness and flexibility of traditional control systems.

[0110] The above describes only the preferred specific embodiments of the present application, but the protection scope of the present application is not limited thereto, and any person skilled in the art can make equivalent substitutions or changes to the technical solutions and inventive concepts of the present application within the technical scope disclosed by the present application, which should be covered within the protection scope of the present application.

Claims

1. A multi-source sensor fusion-based extreme weather electrical control method, characterized in that, The method comprises the following steps: S1, collecting multi-type sensor data sources to construct an original sensor data set; S2, performing time synchronization preprocessing on the original sensor data set, extracting timestamp information, establishing a unified time axis, performing interpolation or downsampling processing on different frequency data, and generating a multi-source alignment data sequence; S3, performing standardization processing on the multi-source alignment data sequence, normalizing numerical data, and extracting state features from image data through a convolutional neural network, and fusing meteorological, electrical and visual information into a fusion feature vector; S4, inputting the fusion feature vector into an Autoformer model, capturing long-term time sequence dependency between multi-source features based on the decomposition structure of the self-attention mechanism, and outputting an extreme weather grade evaluation result and an electrical operation risk score vector; S5, inputting the electrical operation risk score vector into an improved Gumbel-Softmax strategy selection network to generate a continuous type policy combination in a control action set, while introducing a feedback correction mechanism to weight the action sampling weight according to the execution feedback of the historical control action, and outputting a control action sequence and parameter configuration; S6, converting the control action sequence and parameter configuration into device control instructions and issuing them to the target device in real time through the control bus; The S4 specifically comprises: S41, arranging the fusion feature vector corresponding to each time step into a two-dimensional sequence matrix in time sequence, setting a window size, and performing sliding window slicing on the two-dimensional sequence matrix to obtain continuous input segments; S42, performing a sliding average difference operation on each feature dimension in each input segment in the time sequence direction to calculate the average change between adjacent time points, and decomposing each sequence into a group of trend component sequences and residual component sequences; S43, inputting the residual component sequence into a multi-head attention mechanism to perform attention distribution calculation of a preset weight between time steps, outputting a weighted sequence vector representing periodic dependency features, and retaining the trend component sequence as a bias channel; S44, adding the periodic weighted sequence vector and the trend component sequence by time step, and inputting them into a fully connected feedforward network to output a prediction vector after activation function and linear mapping; S45, taking the end time step result in the current window prediction vector as the trend component sequence input in the next time period, recursively performing trend superposition operation, and performing continuous prediction of multiple time steps; S46, dividing the vector at the end time in the prediction vector output by the Autoformer model according to the task target structure in the training stage, and dimensionally dividing the first part into an extreme weather grade through a softmax function, with a value range of 0 to 1, and dividing the score value into four risk grades of normal, general risk, higher risk and extremely high risk according to a preset interval. S47, the second part is a one-dimensional electrical operation risk score vector, corresponding to eight types of electrical risk factors: line voltage fluctuation, current peak overrun, power grid frequency deviation, power bus load imbalance, equipment surface icing, isolating switch contact aging, main transformer temperature rise anomaly and ground resistance anomaly, each score is a real number between 0 and 1, indicating the evaluation intensity of the risk type at the current time, which is used as the input basis for generating control actions.

2. The extreme weather electrical control method based on multi-source sensing fusion according to claim 1, characterized in that, The multi-type sensor specifically includes a meteorological sensor for collecting temperature, humidity and rainfall intensity data, an electrical sensor for collecting voltage, current, load rate and frequency data, and a visual sensor and an infrared sensor for collecting equipment surface image and thermal imaging information.

3. The method of claim 1, wherein the method is based on multi-source sensor fusion for extreme weather electrical control. The S2 specifically includes: S21, read various sensor data in the original sensor data set, extract the time stamp information corresponding to each data record, and uniformly convert the time stamp into coordinated universal time format, determine the sampling frequency of all sensors, and select any sensor as a reference channel, and establish a unified time axis with the time interval of the reference channel as the standard; S22, performing linear interpolation processing on sensor data with a sampling frequency lower than the reference channel, calculating the interpolation value between adjacent time points to fill in the missing data, and performing sliding average downsampling on sensor data with a sampling frequency higher than the reference channel, replacing the average value of a plurality of consecutive sampling points with a representative value of a time period; S23, sorting image data according to time stamp, extracting frame images closest to the time points of the unified time axis, and forming a key frame sequence after time synchronization; S24, combining all sensor data after interpolation, downsampling and frame alignment processing in time sequence to generate a multi-source aligned data sequence under a unified time reference.

4. The extreme weather electrical control method based on multi-source sensing fusion according to claim 1, characterized in that, The S3 specifically includes: S31, performing normalization processing on numerical data in the multi-source aligned data sequence, using the minimum-maximum normalization method to scale all numerical values to between 0 and 1, and preserving the relative change relationship of the original data; S32, extracting frame images in time sequence for image data and inputting them into a convolutional neural network, sequentially performing convolution, activation and pooling operations with a preset size and step, and extracting state feature vectors corresponding to the images; S33, concatenating the normalized meteorological sensor data, electrical sensor data and the extracted state feature vectors corresponding to the images to construct a fusion feature vector corresponding to each time step.

5. The method of claim 1, wherein the method is based on multi-source sensor fusion for extreme weather electrical control. The S5 specifically includes: S51, inputting the electrical operation risk score vector into the embedding layer of the improved Gumbel-Softmax policy selection network for dimension mapping to generate a preset length representation vector, matching it with each action category vector in the predefined control action set, calculating the logarithmic distribution parameters of the sampling probability for each type of control action, and outputting the probability distribution parameters corresponding to the control action set; S52, performing a derivable improved Gumbel-Softmax policy selection network sampling operation on the sampling probability distribution of each category action to generate continuous policy vectors under each action category, and combining multiple policy vectors to form a control action sequence; S53、for generating corresponding execution parameter configuration values for the control action sequence, the parameter configuration values being floating-point numbers or integers related to the control action semantics, constituting a complete control action sequence and being output, the control actions including load limiting control, circuit breaker selection control, standby power supply enabling instruction, deicing heating device scheduling, or remote risk warning triggering.

6. The extreme weather electrical control method based on multi-source sensing fusion according to claim 5, characterized in that, The improved Gumbel-Softmax policy selection network specifically comprises: The electrical operation risk score vector is sent to the policy embedding layer in the order of the electrical risk factor categories to form an input vector, a nonlinear transformation and dimensionality operation are performed on each dimension of the risk score to map a single numerical value to a higher dimensional space through a set of pre-trained linear transformation weights, an embedding vector containing multiple dimensions is formed, a policy score matrix consistent with the size of the control action set is generated, each row of the matrix represents a control action, and each column is the embedding mapping value of the corresponding risk factor score in the policy space; A standard Gumbel-Softmax sampling operation is performed on the policy score matrix, an independent random disturbance term is added to each control action, each random disturbance term is obtained by sampling once from a Gumbel distribution, the score corresponding to each control action is added to the disturbance term, and a normalization operation is performed on each group of disturbed scores to convert the position score of each control action to a continuous probability value, thereby forming a continuous policy distribution; According to the continuous policy distribution, a derivable policy weight combination is generated, each control action corresponds to a floating-point value between 0 and 1, representing the possibility of the action being selected under the current risk state, thereby forming a continuous policy combination vector at the current time point; The continuous policy combination vector is paired with the preset control action list item by item to form an action-weight pair set and perform threshold judgment, if the weight value is higher than the set threshold, the action is included in the current candidate control action set, and the first preset number of control actions with the largest weight value are selected as the control plan for the current period, for each selected control action, the corresponding parameter configuration is queried to form a control action sequence and a corresponding parameter configuration list; After each control period ends, the control action sequence and the corresponding feedback data that have been issued are collected, including: device response delay time, state change amount of the controlled object, execution stability index, and system alarm flag, after the original feedback data is normalized, a feedback feature vector of a set dimension is formed according to the control action number, and is one-to-one corresponding to the action sampling probability in the Gumbel-Softmax network; Each feedback feature vector is input into a scoring function, and the scores of the four types of indexes are calculated in turn according to the preset scoring rules: response timeliness score of response time, execution effectiveness score of state change amount, execution stability score, and safety score of not triggering the protection mechanism. The score vector composed of the effect scores of the single actions respectively weighted by preset weights is input into a feedback correction mechanism, the score value is normalized to convert into a correction coefficient, and the correction coefficient is multiplied by the sampling probability of the corresponding action in the original sampling probability vector to generate a new probability distribution vector, update the sampling weight of the next round of policy selection of the Gumbel-Softmax network, and synchronously update the forward propagation parameters, so as to complete the policy adjustment and correction.

7. The method of claim 1, wherein the method further comprises: The S6 specifically includes: S61, converting the control action sequence and parameter configuration into a device control instruction, and issuing the device control instruction to a target device in real time through a control bus, combining each control action sequence with a corresponding parameter configuration item into a standardized instruction package, and the instruction package containing device number, action type, execution parameter and timestamp information; S62, after encoding and checking the instruction package, the instruction package is sent to the target terminal through the control bus, each device executes the corresponding action immediately after receiving the instruction, and returns the execution state and sensing feedback signal.

Images