Reactive compensation optimization method and system based on AI

Abstract

The invention relates to the technical field of power grid operation and maintenance, and discloses an AI-based reactive compensation optimization method and system, and the method comprises the following steps: integrating the four-dimensional characteristics of traffic flow, power grid load, user preference and equipment health degree, constructing a space-time correlation graph model, achieving the adaptive evolution of a topological structure through dynamic graph convolution and an attention mechanism, and obtaining a reactive compensation optimization model. The method comprises the following steps: fusing spatial-temporal features of a graph neural network with a BiGRU time sequence for modeling, injecting an equipment health degree attenuation factor, generating a charging demand prediction result with uncertainty quantification, expanding an intelligent agent state space, improving a reward function, and embedding equipment health degree constraints in a strategy network and near-end optimization. By sensing the degradation state of the equipment in real time and intelligently adjusting the reactive compensation strategy, the service life of the old equipment is prolonged while the voltage stability is guaranteed, the deviation accumulation effect is effectively inhibited through feedforward-feedback composite control and dynamic rescheduling, and accurate matching of a scheduling instruction and actual output is ensured.

Description

Technical Field

[0001] The present invention relates to the field of power grid operation and maintenance, and more specifically, it relates to an AI-based reactive power compensation optimization method and system. Background Art

[0002] Urban distribution networks play an important role in providing stable and reliable power supply for residential and commercial users. However, with the continuous increase in the number of urban power grid users and the increasingly complex load structure, how to optimize reactive power compensation has become a key issue in the operation and maintenance of distribution networks.

[0003] Traditional reactive power compensation methods usually rely on empirical rules and manual adjustment, which are difficult to adapt to the rapidly changing load characteristics and are prone to low voltage and power loss problems. In recent years, the rapid development of artificial intelligence (AI) technology has provided new ideas for solving reactive power compensation in distribution networks. Summary of the Invention

[0004] The present invention provides an AI-based reactive power compensation optimization method and system to solve the technical problems in related technologies that traditional static scheduling methods rely on empirical rules and manual adjustment, are difficult to adapt to the rapidly changing load characteristics, are prone to low voltage and power loss, especially the impact of power output deviation of old equipment on scheduling.

[0005] The present invention provides an AI-based reactive power compensation optimization method, including the following steps: S100, integrating four-dimensional features of traffic flow, power grid load, user preferences, and equipment health, constructing a spatio-temporal correlation graph model, and realizing the adaptive evolution of the topological structure through dynamic graph convolution and attention mechanism; S200, fusing spatio-temporal features of graph neural networks and BiGRU time series modeling, injecting an equipment health decay factor, and generating a charging demand prediction result with uncertainty quantification; S300, expanding the agent state space and improving the reward function, embedding equipment health constraints in the policy network and proximal optimization, and realizing distributed decision-making considering equipment reliability; S400, performing global resource allocation in the cloud, executing multi-objective optimization at the edge, and realizing cloud-edge collaborative solution through the ADMM algorithm to generate a scheduling instruction with a reliability margin; S500, integrating historical deviation, scheduling instruction, and equipment status data, and using a TCN-GAT hybrid model to predict future execution deviation and evaluate uncertainty; S600, combining feedforward scheduling instructions and PID feedback compensation, dynamically adjusting power output and implementing equipment priority management, and ensuring instruction security through encryption verification; S700, monitoring the cumulative load deviation and triggering a rescheduling mechanism, constructing a multi-objective model considering equipment reliability, and realizing smooth switching between old and new instructions; S800,aggregates real-time execution data, adopts sliding window statistics and anomaly filtering mechanism, recursively updates device health indicators and implements normalization.

[0006] Furthermore, in S100, the following steps are specifically included: S110, node feature initialization: construct a multi-dimensional feature vector for each road section node, covering traffic, power grid, user and equipment status information; S120, edge weight calculation: construct an adjacency matrix based on the traffic flow transfer probability to reflect the dynamic association of the road network; S130, Spatiotemporal Attention Generation: Capturing spatiotemporal correlations through multi-head attention mechanism; S140, dynamic graph convolution update: integrating spatiotemporal features to update graph structure weights; S150, health feature normalization: normalize the equipment health index.

[0007] Furthermore, in S200, the following steps are specifically included: S210, Spatiotemporal Graph Convolution Feature Extraction: Based on Dynamic Graph Perform multi-level spatial feature aggregation; S220, Time Series Modeling: Time series dependency modeling of historical load series; S230, Reliability attenuation factor injection: adjust the prediction weight according to the equipment health; S240, multimodal feature fusion: fusion of spatiotemporal features and attenuation factors to generate the final prediction; The final prediction is calculated as follows: ; in, represents the feature projection matrix, It represents the historical average value for the same period. Indicates time The predicted value of charging demand, Indicates the device At the moment The hidden state of Indicates the total number of devices. Indicates the device Reliability attenuation factor; S250, Uncertainty Quantification: Evaluating confidence intervals for predictions.

[0008] Furthermore, in S300, the following steps are specifically included: S310, state space expansion: incorporating equipment health indicators into the agent's observed state; S320, Reliability-Weighted Reward Design: Introduce a device health penalty term into the basic reward function; S330, Policy Network Improvement: Inject reliability attention into the output layer of the policy network; S340, Proximal Policy Optimization: Policy gradient update with reliability constraints; S350, Distributed Policy Execution: Each agent generates actions based on local observations; The calculation formula for the action is as follows: ; Where, represents the execution action of device i at time t, represents the action space, represents the optimized policy network, represents the current state of device i, represents the health index of device i, represents the temperature coefficient, represents the optimized policy parameters.

[0009] Furthermore, in S400, it specifically includes the following steps: S410, Cloud Global Resource Allocation: Cross-regional power capacity allocation and transmission loss optimization; S420, Edge Multi-objective Optimization: Balance user cost and grid stability under device reliability constraints; S430, Constraint Consistency Coordination: Achieve cloud-edge optimization decoupling through the ADMM algorithm; S440, Robust Margin Calculation: Dynamically adjust the power margin according to the device health; S450, Scheduling Instruction Generation: Generate a scheduling instruction set with reliability identification; The calculation formula for the scheduling instruction set is as follows: ; Where: ; Where, represents the scheduling instruction set, represents the total number of devices, represents the final scheduling power of device k, represents the health index of device k, represents the safety margin of device k, represents the locally calculated power value, represents the cloud-calculated power value, represents the local power weight coefficient.

[0010] Furthermore, in S500, it specifically includes the following steps: S510, Multi-source data fusion: Integrate device historical data, scheduling instructions, and health indicators; S520, Temporal feature extraction and representation: Use deep TCN to capture long-range dependence patterns; S530, Spatial association modeling and representation: Capture the spatial correlation between devices through graph attention networks; S540, Cross-modal fusion prediction: Fusion of spatio-temporal features to generate deviation predictions; The calculation formula for deviation prediction is as follows: ; Where, represents the fusion projection matrix, represents the feature splicing operation, represents the bias term, represents the predicted deviation at time t+1, represents the layer normalization operation, represents the temporal feature, represents the spatial feature; S550, Uncertainty quantification: Evaluate the probability distribution characteristics of the predicted values.

[0011] Furthermore, in S600, it specifically includes the following steps: S610, Generation of feedforward control term: Generate a reference control quantity based on robust scheduling instructions; S620, Calculation of feedback compensation term: Dynamically adjust the control quantity according to the real-time execution deviation; S630, Dynamic amplitude limiting processing: Combine device reliability indicators to constrain the control output; S640, Reallocation of device priorities: Dynamically adjust the control order according to the health status; S650, Instruction security verification: Ensure the integrity of control instructions based on digital signatures; The calculation formula for the final encrypted instruction is as follows: ; Where, represents the timestamp, represents the ECC elliptic curve signature, represents the SHA-256 hash function, represents the XOR operation, represents the adjusted control instruction, represents the device health indicator, represents the device priority, represents the final encrypted instruction.

[0012] Furthermore, in S700, it specifically includes the following steps: S710, Deviation Cumulative Monitoring: Real-time detection of the integral effect of grid load deviation; S720, Rescheduling Trigger Judgment: Trigger the rescheduling mechanism based on multiple thresholds; S730, Robust Rescheduling Modeling: Construct a multi-objective optimization model considering equipment reliability; S740, Distributed Collaborative Solving: Implement cloud-edge collaborative solving using an improved ADMM algorithm; S750, Instruction Dynamic Switching: Achieve seamless transition between old and new scheduling instructions; The calculation formula for the final power instruction is as follows: ; ; Wherein, represents the dynamic switching coefficient, represents the attenuation coefficient of the old instruction, represents the switching rate parameter, represents the starting moment of switching, represents the current moment, represents the final power instruction, represents the new power instruction, represents the original power instruction.

[0013] Furthermore, in S800, it specifically includes the following steps: S810, Multi-dimensional Data Aggregation: Integrate real-time execution data and historical health records; S820, Sliding Window Statistical Analysis: Calculate the deviation distribution characteristics of equipment within the window period; S830, Abnormal Data Filtering: Eliminate abnormal measurement values based on the improved 3σ criterion; S840, Health Degree Recursive Update: Adopt exponential weighted update with an adaptive forgetting factor; S850, Health Degree Normalization Processing: Adopt improved Min-Max normalization processing.

[0014] The present invention also proposes an AI-based reactive power compensation optimization system for performing the steps of an AI-based reactive power compensation optimization method as described above, including: Dynamic Graph Processing Module: Real-time integrate traffic flow, grid load, user behavior, and equipment health data, construct a dynamic graph with spatio-temporal correlation characteristics, and continuously optimize the network topology expression through adaptive graph convolution; Spatio-temporal Prediction Module: Integrate graph neural network and bidirectional time series model, combine with the equipment reliability attenuation factor, generate future period charging demand prediction, and quantify the uncertainty of the prediction result; Strategy Optimization Module: Embed the device health constraint in the multi-agent reinforcement learning framework, dynamically adjust the reward function and the policy network structure, and achieve reliability-aware distributed decision optimization; Cloud-Edge Scheduling Module: Achieve the coordination of global resource allocation and local multi-objective optimization through a hierarchical optimization architecture, and use a distributed algorithm to solve the scheduling scheme with reliability margin; Deviation Prediction Module: Predict the device execution deviation based on a temporal-spatial hybrid model, and evaluate the credibility of the prediction results by combining historical data and real-time status; Execution Control Module: Integrate the feed-forward scheduling instruction and the feedback compensation mechanism, dynamically adjust the power output and implement device priority management, and ensure the security of control instructions through encryption verification; Dynamic Scheduling Module: Monitor the system operation deviation and trigger the rescheduling mechanism, and achieve the dynamic optimization and smooth transition of resource allocation on the premise of ensuring device reliability; Health Management Module: Continuously collect device operation data, update the health indicators through sliding window statistics and anomaly filtering, and provide dynamic device status evaluation for the whole system.

[0015] The beneficial effects of the present invention are as follows: By integrating dynamic spatio-temporal perception and multi-level intelligent decision-making, the present invention effectively improves the response ability of the urban distribution network to complex reactive power compensation requirements, and realizes precise control with self-adaptive device health; The system can autonomously coordinate grid resources, significantly enhance the stability of power supply quality, reduce harmonic interference and power loss, and at the same time extend the service life of key devices; Through the closed-loop mechanism of prediction-compensation-optimization, effectively prevent the fault risk caused by voltage fluctuations, and provide an intelligent and highly reliable reactive power compensation solution for high-density load scenarios; By real-time sensing the device degradation state, intelligently adjusting the reactive power compensation strategy, while ensuring voltage stability, extend the service life of old equipment, and effectively suppress the deviation accumulation effect through feed-forward-feedback composite control and dynamic rescheduling, ensuring the precise matching of scheduling instructions and actual output; Brief Description of the Drawings

[0016] Figure 1 is a flowchart of a reactive power compensation optimization method based on AI proposed by the present invention; Figure 2 is a structural block diagram of a reactive power compensation optimization system based on AI proposed by the present invention.

[0017] In the figure: 101, dynamic graph processing module; 102, spatio-temporal prediction module; 103, policy optimization module; 104, cloud-edge scheduling module; 105, deviation prediction module; 106, execution control module; 107, dynamic scheduling module; 108, health management module. Detailed implementation manners

[0018] The subject matter described herein will now be discussed with reference to example embodiments. It should be understood that discussing these embodiments is only to enable those skilled in the art to better understand and thus implement the subject matter described herein, and changes can be made to the functions and arrangements of the elements discussed without departing from the scope of protection of the content of this specification. Each example can omit, substitute, or add various processes or components as needed. Additionally, the features described for some examples can also be combined in other examples.

[0019] As Figure 1 - Figure 2 shown, an AI-based reactive power compensation optimization method includes the following steps: S100, Dynamic Graph Construction (DGNN): Integrate four-dimensional features of traffic flow, grid load, user preferences, and equipment health, construct a spatio-temporal correlation graph model, and achieve adaptive evolution of the topological structure through dynamic graph convolution and attention mechanism; In an embodiment of the present invention, it specifically includes the following steps: S110, Node Feature Initialization: Construct a multi-dimensional feature vector for each road segment node covering traffic, grid, user, and equipment status information; Formula: ; Wherein, represents the traffic flow of road segment at time, represents the grid node load associated with road segment , represents the charging preference index of the user of road segment , , represents the standard deviation of the historical power deviation of the charging pile of road segment , that is, the health index, represents the feature vector of road segment at time , , represents the set of all road segment nodes in the road network; S120, Edge Weight Calculation: Construct an adjacency matrix based on the vehicle flow transfer probability to reflect the dynamic association of the road network; Formula: ; Among them, represents the road segment similarity coefficient, represents the traffic flow from road segment to road segment ; represents the set of adjacent road segments of road segment ; represents the edge weight from road segment to ; represents the traffic flow from road segment i to road segment k at time t, represents the total out-flow traffic flow of road segment i at time t; S130, Spatio-Temporal Attention Generation: Capturing spatio-temporal correlations through the multi-head attention mechanism; Formula: ; ; Among them, respectively represent the query / key transformation matrices, represents the output projection of the -th attention head, represents the dimension of a single attention head, represents the number of attention heads, represents the spatio-temporal attention matrix of the -th attention head, represents the fused spatio-temporal attention matrix, represents the attention head concatenation operation, represents the normalization function, represents the matrix transpose, road segment at time ; S140, Dynamic Graph Convolution Update: Fusing spatio-temporal features to update the graph structure weights; Formula: ; Time difference operator: ; ; Among them, represents the dynamic graph convolution kernel parameter, represents the tensor concatenation operation, represents the graph weight matrix at time ; represents the graph weight matrix at the next time, represents the traffic flow matrix at time ; represents the traffic flow matrix at the previous moment, represents the time difference of the traffic flow, represents the time difference of the health degree, represents the Hadamard product (element-wise product), represents the dynamic graph convolutional network operation, represents the road section at the current moment the standard deviation of the historical power deviation of the charging pile; S150, normalization of the health degree feature: standardize the device health degree index; Formula: ; where, represents the mean value of the health degree index, represents the standard deviation of the health degree index, represents the expected value of calculating the standard deviation of the power deviation of all devices, represents the standard deviation of calculating the standard deviation of the power deviation of all devices, represents the standard deviation of the historical power deviation of the charging pile at the current moment after standardization, that is, the standardized health degree index; Output product: the standardized dynamic graph structure; Output product: the standardized dynamic graph structure; ; wherein: ; where, represents the standardized dynamic graph structure, represents the standardized node features, represents the standardized edge weights, represents the standardized graph weight matrix; S200, demand prediction: fuse the spatio-temporal features of the graph neural network and the BiGRU time series modeling, inject the device health degree decay factor, and generate the charging demand prediction result with uncertainty quantification; In one embodiment of the present invention, it specifically includes the following steps: S210, spatio-temporal graph convolutional feature extraction: based on the dynamic graph perform multi-level spatial feature aggregation; Formula: ; ; where, represents the initial node feature matrix, represents the adjacency matrix with self-connections added, represents the trainable parameters, Indicates The node feature matrix of the layer, express The identity matrix, Indicates the total number of nodes, Represents the rectified linear unit activation function, Represents the power of the degree matrix; S220, Time Series Modeling: Time series dependency modeling of historical load series; formula: ; in, Represents history Charging requirements during the period, represents the final graph convolution output, Indicates time The hidden state output of represents the time step, Indicates the size of the historical time window, Indicates the length of historical data, represents a bidirectional gated recurrent unit network; S230, Reliability attenuation factor injection: adjust the prediction weight according to the equipment health; formula: ; in, represents the health decay coefficient, represents the standardized health index, Indicates the device reliability attenuation factor, Represents the S-type activation function; S240, multimodal feature fusion: fusion of spatiotemporal features and attenuation factors to generate the final prediction; formula: ; in, Represents the feature projection matrix, It represents the historical average value for the same period. Indicates time charging demand forecast value, Indicates the device At the moment The hidden state of Indicates the total number of equipment; S250, Uncertainty Quantification: Evaluating confidence intervals of prediction results; formula: ; in, Represents the standard deviation of the prediction result, Represents the gradient of the predicted value with respect to the hidden state, Represents the uncertainty of the hidden state, Represents the base uncertainty constant, Represents the total number of devices; S300, Reliability-Enhanced MAPPO: Expand the agent state space and improve the reward function, embed device health constraints in the policy network and proximal optimization, and achieve distributed decision-making considering device reliability; In one embodiment of the present invention, it specifically includes the following steps: S310, State Space Expansion: Incorporate device health metrics into the agent's observation state; Formula: ; Wherein, Represents the state vector of device i at time t, Represents the normalized node feature vector, Represents the predicted value of the charging demand, Represents the prediction uncertainty, Represents the device reliability decay factor; S320, Reliability-Weighted Reward Design: Introduce a device health penalty term into the base reward function; Formula: ; Wherein, Represents the value of the new reward function, Respectively represent the weight coefficients of the base reward terms for operating revenue, user cost, and grid stability metrics, Represents the operating revenue, Represents the user cost, Represents the grid stability metric, Represents the health penalty coefficient, Represents the health metric of device k, Represents the historical execution deviation, Represents the total number of devices; S330, Policy Network Improvement means injecting reliability attention into the output layer of the policy network; Formula: ; Wherein, Represents the action probability output by the policy network, Represents the weight matrix of the policy network, Represents the bias term of the policy network, Represents the state of the hidden layer representation, represents the reliability scaling factor, represents the health index of device i, represents the action space, represents the candidate action, the corresponding hidden layer representation; S340, Proximal Policy Optimization: Policy gradient update with reliability constraints; Formula: ; Constraint condition: ; where, represents the updated policy parameters, represents the current policy parameters, represents the action probability under the new policy, represents the action probability under the old policy, represents the advantage function, represents the KL divergence penalty coefficient, represents the health penalty coefficient, represents the KL divergence, represents the KL divergence constraint threshold, represents the health constraint threshold, represents the KL divergence between the new and old policies, represents the expected value of the health indices of all devices; S350, Distributed Policy Execution: Each agent generates an action based on local observations; Formula: ; where, represents the action executed by device i at time t, represents the action space, represents the optimized policy network, represents the current state of device i, represents the health index of device i, represents the temperature coefficient, represents the optimized policy parameters; S400, Hierarchical Scheduling Decision: The cloud performs global resource allocation, the edge executes multi-objective optimization, and cloud-edge collaborative solution is achieved through the ADMM algorithm to generate scheduling instructions with reliability margins; In one embodiment of the present invention, it specifically includes the following steps: S410, Cloud Global Resource Allocation: Cross-regional power capacity allocation and transmission loss optimization; Formula: ; ; ; ; Among them, represents the total number of regions, represents the operating revenue of region n, represents the transmission loss weight coefficient, represents the transmission loss of region n, represents the grid power of region n, represents the maximum grid capacity, represents the system-level health index, represents the power upper and lower limits of region n; S420, Edge multi-objective optimization: Balance user cost and grid stability under device reliability constraints; Formula: ; ; ; Among them, represents the total number of users, represents the cost of user u, represents the preference coefficient of user u, represents the attenuation factor of user u, represents the power deviation penalty coefficient, represents the power prediction deviation, represents the maximum power of device k, represents the health index of device k, represents the optimized policy parameter, represents the local policy parameter, represents the policy deviation threshold; S430, Constraint consistency coordination: Achieve cloud-edge optimization decoupling through the ADMM algorithm; Formula: ; Among them, represents the step size, represents the penalty factor, represents the cloud constraint space, represents the auxiliary variable at the k-th iteration, represents the objective function, represents the augmented Lagrangian function, represents the Lagrange multiplier, represents the grid power, represents the local power prediction value, Represents a projection operator; S440, Robust margin calculation: Dynamically adjust the power margin according to the device health; Formula: ; Wherein, Represents the safety margin of device k, Represents the maximum power of device k, Represents the predicted power of device k, Represents the health index of device k, Represents the inverse function of the standard normal distribution, Represents the rectified linear unit function; S450, Scheduling instruction generation: Generate a set of scheduling instructions with reliability identification; Formula: ; Wherein: ; Wherein, Represents the set of scheduling instructions, Represents the total number of devices, Represents the final scheduled power of device k, Represents the health index of device k, Represents the safety margin of device k, Represents the locally calculated power value, Represents the cloud-computed power value, Represents the local power weight coefficient; S500, Execution deviation prediction: Integrate historical deviations, scheduling instructions, and device status data, and use a TCN-GAT hybrid model to predict future execution deviations and evaluate uncertainties; In an embodiment of the present invention, it specifically includes the following steps: S510, Multi-source data fusion: Integrate device historical data, scheduling instructions, and health indicators; Formula: ; Wherein, Represents the historical execution deviation, Represents the final scheduling instruction, Represents the device health, Represents the attenuation factor, Represents the length of the historical data window, Represents the current time, Represents the device index; S520, Temporal feature extraction: Use deep TCN to capture long-range dependence patterns; Formula: ; wherein, denotes the dilated convolution kernel, denotes the temporal feature output, denotes the input data sequence, denotes the dilated convolution operation with the weight matrix ; " " denotes the standard one-dimensional causal convolution operation, ensuring that the current output only depends on historical inputs; S530, Spatial correlation modeling means capturing the spatial correlation between devices through a graph attention network; Formula: ; ; wherein, denotes the topological neighbor of device k, denotes the node feature embedding, denotes the query / key-value transformation matrix, denotes the value transformation matrix, denotes the attention weight, denotes the spatial feature output, denotes the hidden layer dimension, denotes the multi-head attention concatenation operation; S540, Cross-modal fusion prediction: Fusing spatio-temporal features to generate deviation predictions; Formula: ; wherein, denotes the fusion projection matrix, denotes the feature concatenation operation, denotes the bias term, denotes the predicted deviation at time t+1, denotes the layer normalization operation, denotes the temporal feature, denotes the spatial feature; S550, Uncertainty quantification: Evaluating the probability distribution characteristics of predicted values; Formula: ; wherein, denotes the numerical stability term, denotes the prediction uncertainty, denotes the temporal feature variance, denotes the spatial feature variance; S600, Real-time control with feedback correction: Combining feedforward scheduling instructions with PID feedback compensation, dynamically adjusting power output and implementing device priority management, ensuring instruction security through encrypted verification; In an embodiment of the present invention, it specifically includes the following steps: S610, Generation of feedforward control term: Generating a reference control quantity based on robust scheduling instructions; Formula: ; Wherein, represents the final scheduling instruction value, represents the prediction deviation value, represents the device health index, represents the rated power of the device, represents the feedforward control output, represents the S-shaped activation function; S620, Calculation of feedback compensation term: Dynamically adjusting the control quantity according to the real-time execution deviation; Formula: ; ; Wherein, represents the proportional coefficient, represents the integral coefficient, represents the differential coefficient, represents the integral window length, represents the historical deviation, represents the actual execution power, represents the target power, represents the feedback compensation quantity represents the current moment; S630, Dynamic amplitude limiting processing: Combining the device reliability index to constrain the control output; Formula: ; Wherein, represents the upper and lower limits of the device operating power, represents the device health index, represents restricting x within the interval [a, b], represents the feedforward control quantity, represents the feedback compensation quantity, represents the adjusted control output; S640, Reallocation of device priority means dynamically adjusting the control order according to the health degree; Formula: ; Wherein, Indicates the sensitivity of the reward to power, Indicates the device attenuation factor, Indicates the time decay coefficient, Indicates the device running time, Indicates the device health index, Indicates the priority value of device k; S650, Instruction Security Verification: Ensure the integrity of control instructions based on digital signatures; Formula: ; Wherein, Indicates the timestamp, Indicates the ECC elliptic curve signature, Indicates the SHA-256 hash function, Indicates the exclusive OR operation, Indicates the adjusted control instruction, Indicates the device health index, Indicates the device priority, Indicates the final encrypted instruction; S700, Dynamic Resource Rescheduling: Monitor the cumulative load deviation and trigger the rescheduling mechanism, build a multi-objective model considering device reliability, and achieve smooth switching between old and new instructions; In an embodiment of the present invention, it specifically includes the following steps: S710, Deviation Cumulative Monitoring: Real-time detect the integral effect of the grid load deviation; Formula: ; ; Wherein, Indicates the real-time load deviation, Indicates the actual load value, Indicates the predicted load value, Indicates the time decay coefficient, Indicates the monitoring start time, Indicates the current time, Indicates the total number of load measurement points, Indicates the integral variable; S720, Rescheduling Trigger Judgment: Trigger the rescheduling mechanism based on multiple thresholds; Formula: ; ; ; Wherein, Indicates the integral threshold, Indicates the instantaneous threshold, represents the minimum reliability threshold, represents the reference load value, represents the evaluation time length, represents the reliability factor of device k, represents the load deviation value; S730, Robust Rescheduling Modeling: Construct a multi-objective optimization model considering device reliability; Formula: ; ; ; ; ; where, respectively represent the weight coefficients, represents the power compensation amount, represents the device reliability factor, represents the maximum allowable adjustment amount, represents the load deviation, represents the health index of device k, represents the power adjustment amount, represents the new power command, represents the original power command, respectively represent the upper and lower power limits, represents the new policy parameter, represents the original policy parameter, represents the policy adjustment threshold, represents the total number of devices; S740, Distributed Collaborative Solving: Use the improved ADMM algorithm to achieve cloud-edge collaborative solving; Cloud: ; Edge: ; Update: ; where, represents the inertia coefficient, represents the reliability weight, represents the penalty factor, represents the augmented Lagrangian function, represents the auxiliary variable at the k-th iteration, represents the power variable at the k-th iteration, represents the Lagrange multiplier at the k-th iteration, Denotes the device reliability factor, Denotes the updated auxiliary variable, Denotes the updated power variable, Denotes the updated Lagrange multiplier; S750, Instruction dynamic switching: Achieve seamless transition between old and new scheduling instructions; Formula: ; ; Among them, Denotes the dynamic switching coefficient, Denotes the old instruction attenuation coefficient, Denotes the switching rate parameter, Denotes the starting time of switching, Denotes the current time, Denotes the final power instruction, Denotes the new power instruction, Denotes the original power instruction; S800, Health status online update: Aggregate real-time execution data, adopt a sliding window statistics and anomaly filtering mechanism, recursively update the device health status indicators and perform normalization processing; In an embodiment of the present invention, it specifically includes the following steps: S810, Multi-dimensional data aggregation: Integrate real-time execution data and historical health records; Formula: ; ; Among them, Denotes the real-time execution deviation, Denotes the actual execution power, Denotes the target power instruction, Denotes the predicted deviation value, Denotes the historical health status indicator, Denotes the device priority, Denotes the reliability attenuation factor, Denotes the current time; S820, Sliding window statistical analysis: Calculate the deviation distribution characteristics of the device within the window period; Formula: ; ; ; Among them, Denotes the statistical window size, Denotes the time decay weight, represents the attenuation coefficient, represents the mean deviation, represents the original standard deviation, represents the power deviation at time τ, represents the current time, represents the historical time; S830, Abnormal data filtering: Eliminate abnormal measurement values based on the improved 3σ criterion; Formula: ; where, represents the reliability factor adjustment threshold, represents the adjusted power deviation, represents the original power deviation, represents the mean deviation, represents the original standard deviation; S840, Health degree recursive update: Adopt exponential weighted update with an adaptive forgetting factor; Formula: ; ; where, represents the adaptive forgetting factor, represents the rated power of the device, represents the health degree index at the current time, represents the health degree index at the previous time, represents the adjusted power deviation; S850, Health degree normalization processing: Adopt improved Min - Max normalization processing; Formula: ; where, represents the mean value of the system health degree, represents the stable term of the health degree value, represents the maximum health degree value based on all system devices, represents the normalized health degree index, represents the original health degree index.

[0020] In an embodiment of the present invention, a reactive power compensation optimization system based on AI is also proposed, which is used to execute the steps in the above - mentioned reactive power compensation optimization method based on AI, and includes the following modules: Dynamic graph processing module 101: Integrate traffic flow, power grid load, user behavior, and device health degree data in real - time, construct a dynamic graph with spatio - temporal correlation characteristics, and continuously optimize the network topology expression through adaptive graph convolution; Space-time prediction module 102: Integrate the graph neural network and the bidirectional time series model, combine with the device reliability decay factor, generate the charging demand prediction for the future period, and quantify the uncertainty of the prediction results; Policy optimization module 103: Embed the device health constraint in the multi-agent reinforcement learning framework, dynamically adjust the reward function and the policy network structure, and achieve the reliability-aware distributed decision optimization; Cloud-edge scheduling module 104: Achieve the coordination of global resource allocation and local multi-objective optimization through a hierarchical optimization architecture, and use a distributed algorithm to solve the scheduling scheme with reliability margin; Deviation prediction module 105: Predict the device execution deviation based on the time series-space hybrid model, and evaluate the credibility of the prediction results by combining historical data and real-time status; Execution control module 106: Integrate the feed-forward scheduling instruction and the feedback compensation mechanism, dynamically adjust the power output and implement the device priority management, and ensure the security of the control instruction through encryption verification; Dynamic scheduling module 107: Monitor the system operation deviation and trigger the rescheduling mechanism, and achieve the dynamic optimization and smooth transition of resource allocation on the premise of ensuring the device reliability; Health management module 108: Continuously collect the device operation data, update the health index through sliding window statistics and anomaly filtering, and provide the dynamic device status evaluation for the whole system.

[0021] Scenario background: A large commercial complex faces the charging management challenge during the holiday peak period: when 300 fast chargers in the underground parking lot operate simultaneously, the reactive power fluctuation caused by the charging harmonics of electric vehicles leads to a sudden drop in the voltage of the distribution room, and at the same time, the start and stop of the air conditioning system in the commercial area exacerbate the load peak-valley difference.

[0022] The process of its reactive power compensation optimization method is as follows: Dynamic perception: The roadside camera and the charging pile upload the traffic flow data and the device status in real time, and the system constructs a dynamic map including the parking lot topology, the charging demand hot zone, and the transformer load rate.

[0023] Demand prediction: The AI model predicts the charging peak during the cinema closing time 2 hours in advance, and identifies that the charging piles in Area C will have a concentrated demand 3 times that of normal days.

[0024] Intelligent scheduling: The cloud coordinates the photovoltaic energy storage system in the commercial area to increase the output, and the edge node adjusts the charging pile group on the B1 floor to the "power flexible mode" and applies for the dynamic capacity increase permission from the power grid.

[0025] Deviation compensation: When the actual output of old charging piles fluctuates, the system automatically reduces the reactive power compensation of adjacent healthy devices for hedging to maintain the stability of the bus voltage.

[0026] Emergency response: Suddenly encountering voltage flicker caused by the startup of the elevator group control system, the system switches to the standby compensation scheme within 200 ms and converts the charging piles in Area D to the voltage support mode.

[0027] Application value: In the actual measurement during the peak period of summer promotion in the commercial area, the system successfully avoided the tripping of the distribution room caused by the impact of charging load. At the same time, through dynamic scheduling, the reactive power loss of the central air-conditioning system was reduced, and the overall energy efficiency level of the shopping mall was significantly improved.

[0028] The operation and maintenance personnel can real-time master the power quality status of each area through the three-dimensional visualization interface, realizing the transformation from passive repair to active prevention.

[0029] The above describes the embodiments of the present invention, but the present invention is not limited to the above specific implementation manners. The above specific implementation manners are only illustrative and not restrictive. Under the inspiration of the present invention, those of ordinary skill in the art can also make many forms, all of which fall within the protection scope of the present invention.

Claims

1. An AI-based reactive power compensation optimization method, characterized in that, It includes the following steps: S100, Integrate four-dimensional features of traffic flow, grid load, user preferences, and equipment health, construct a spatio-temporal correlation graph model, and achieve the adaptive evolution of the topological structure through dynamic graph convolution and attention mechanism; S200, Fuse the spatio-temporal features of the graph neural network and the BiGRU time series modeling, inject the equipment health decay factor, and generate the charging demand prediction results with uncertainty quantification; S300, Expand the agent state space and improve the reward function, embed the equipment health constraint in the policy network and proximal optimization, and achieve distributed decision-making considering equipment reliability; S400, The cloud performs global resource allocation, the edge executes multi-objective optimization, and the cloud-edge collaborative solution is achieved through the ADMM algorithm to generate scheduling instructions with reliability margins; S500, Integrate historical deviation, scheduling instructions, and equipment status data, and use the TCN-GAT hybrid model to predict future execution deviation and evaluate uncertainty; S600, Combine the feedforward scheduling instructions and PID feedback compensation, dynamically adjust the power output and implement equipment priority management, and ensure the safety of the instructions through encryption verification; S700, Monitor the cumulative load deviation and trigger the rescheduling mechanism, construct a multi-objective model considering equipment reliability, and achieve smooth switching between old and new instructions; S800, Aggregate real-time execution data, use the sliding window statistics and anomaly filtering mechanism, recursively update the equipment health indicators and perform normalization processing.

2. The reactive power compensation optimization method based on AI according to claim 1, wherein, In S100, it specifically includes the following steps: S110, Node feature initialization: Construct a multi-dimensional feature vector for each road section node, covering traffic, grid, user, and equipment status information; S120, Edge weight calculation: Construct an adjacency matrix based on the vehicle flow transfer probability to reflect the dynamic association of the road network; S130, Spatio-temporal attention generation: Capture spatio-temporal correlation through the multi-head attention mechanism; S140, Dynamic graph convolution update: Integrate spatio-temporal features to update the graph structure weight; S150, Health feature normalization: Standardize the equipment health indicators.

3. The reactive power compensation optimization method based on AI according to claim 2, wherein In S200, it specifically includes the following steps: S210, Spatio-temporal graph convolution feature extraction: Perform multi-level spatial feature aggregation based on the dynamic graph; S220, Time series modeling: Perform time series dependence modeling on the historical load sequence; S230, Reliability decay factor injection: Adjust the prediction weight according to the equipment health; S240, Multi-modal feature fusion: Integrate spatio-temporal features and decay factors to generate the final prediction; The calculation formula for the final prediction is as follows: ； in, represents the feature projection matrix, It represents the historical average value for the same period. Indicates time The predicted value of charging demand, Indicates the device At the moment The hidden state of Indicates the total number of devices. Indicates the device Reliability attenuation factor; S250, Uncertainty quantification: Evaluate the confidence interval of the prediction results.

4. The reactive power compensation optimization method based on AI according to claim 3, characterized in that In S300, it specifically includes the following steps: S310, State space expansion: Incorporate the equipment health indicators into the agent's observation state; S320, Reliability weighted reward design: Introduce an equipment health penalty term in the basic reward function; S330, Policy network improvement: Inject reliability attention in the output layer of the policy network; S340, Proximal policy optimization: Policy gradient update with reliability constraints; S350, Distributed policy execution: Each agent generates actions based on local observations; The calculation formula for the action is as follows: ； wherein, represents the execution action of device i at time t, represents the action space, represents the optimized policy network, represents the current state of device i, represents the health index of device i, represents the temperature coefficient, represents the optimized policy parameters.

5. An AI-based reactive power compensation optimization method according to claim 4, characterized in that In S400, it specifically includes the following steps: S410, Cloud global resource allocation: Cross-regional power capacity allocation and transmission loss optimization; S420, Edge multi-objective optimization: Balancing user costs and grid stability under device reliability constraints; S430, Constraint consistency coordination: Achieving cloud-edge optimization decoupling through the ADMM algorithm; S440, Robust margin calculation: Dynamically adjusting the power margin according to device health; S450, Generation of scheduling instructions: Generating a set of scheduling instructions with reliability identifiers; The calculation formula of the scheduling instruction set is as follows: ； Where: ； Among them, represents the scheduling instruction set, represents the total number of devices, represents the final scheduling power of device k, represents the health index of device k, represents the safety margin of device k, represents the power value of local computing, represents the power value of cloud computing, represents the local power weight coefficient.

6. The method for optimizing reactive power compensation based on AI according to claim 5, wherein In S500, it specifically includes the following steps: S510, Multi-source data fusion: Integrating device historical data, scheduling instructions, and health indicators; S520, Temporal feature extraction and representation: Capturing long-range dependence patterns using deep TCN; S530, Spatial association modeling and representation: Capturing the spatial correlation between devices through graph attention networks; S540, Cross-modal fusion prediction: Fusing spatio-temporal features to generate deviation predictions; The calculation formula of the deviation prediction is as follows: Among them, represents the fusion projection matrix, represents the feature splicing operation, represents the bias term, represents the prediction deviation at time t+1, represents the layer normalization operation, represents the temporal feature, represents the spatial feature; S550, Uncertainty quantification: Evaluating the probability distribution characteristics of predicted values.

7. An AI-based reactive power compensation optimization method according to claim 6, characterized in that, In S600, it specifically includes the following steps: S610, Generation of feedforward control terms: Generating a benchmark control quantity based on robust scheduling instructions; S620, Calculation of feedback compensation terms: Dynamically adjusting the control quantity according to real-time execution deviations; S630, Dynamic amplitude limiting processing: Combining device reliability index constraints to control the output; S640, Reallocation of device priorities: Dynamically adjusting the control order according to health; S650, Instruction security verification: Ensuring the integrity of control instructions based on digital signatures; The calculation formula of the final encrypted instruction is as follows: ； Among them, represents the timestamp, represents the ECC elliptic curve signature, represents the SHA-256 hash function, represents the exclusive OR operation, represents the adjusted control instruction, represents the device health indicator, represents the device priority, represents the final encryption instruction.

8. An AI-based reactive power compensation optimization method according to claim 7, characterized in that In S700, it specifically includes the following steps: S710, Deviation accumulation monitoring: Real-time detection of the integral effect of grid load deviations; S720, Judgment of rescheduling trigger: Triggering the rescheduling mechanism based on multiple thresholds; S730, Robust rescheduling modeling: Constructing a multi-objective optimization model considering device reliability; S740, Distributed collaborative solution: Implementing cloud-edge collaborative solution using an improved ADMM algorithm; S750, Instruction dynamic switching: Achieving seamless transition between old and new scheduling instructions; The calculation formula of the final power instruction is as follows: ； ； Among them, represents the dynamic switching coefficient, represents the attenuation coefficient of the old instruction, represents the switching rate parameter, represents the starting moment of switching, represents the current moment, represents the final power instruction, represents the new power instruction, represents the original power instruction.

9. The reactive power compensation optimization method based on AI according to claim 8, characterized in that In S800, it specifically includes the following steps: S810, Multi-dimensional data aggregation: Integrating real-time execution data and historical health records; S820, Sliding window statistical analysis: Calculating the deviation distribution characteristics of devices within the window period; S830, Abnormal data filtering: Removing abnormal measurement values based on the improved 3σ criterion; S840, Recursive update of health: Exponentially weighted update using an adaptive forgetting factor; S850, Normalization processing of health: Using improved Min-Max normalization processing.

10. An AI-based reactive power compensation optimization system, characterized in that, The steps for executing a reactive power compensation optimization method based on AI as described in any one of claims 1-9 include: Dynamic graph processing module: Real-time integration of traffic flow, grid load, user behavior, and device health data, constructing a dynamic graph with spatio-temporal correlation characteristics, and continuously optimizing the network topology expression through adaptive graph convolution; Space-time prediction module: Integrate graph neural network and bidirectional time series model, combine with the device reliability decay factor, generate the charging demand prediction for future periods, and quantify the uncertainty of the prediction results; Strategy optimization module: Embed the device health constraint in the multi-agent reinforcement learning framework, dynamically adjust the reward function and policy network structure, and achieve reliability-aware distributed decision optimization; Cloud-edge scheduling module: Achieve the coordination of global resource allocation and local multi-objective optimization through a hierarchical optimization architecture, and use a distributed algorithm to solve the scheduling scheme with reliability margin; Deviation prediction module: Predict the device execution deviation based on the time series-space hybrid model, and evaluate the credibility of the prediction results by combining historical data and real-time status; Execution control module: Integrate the feedforward scheduling instruction and feedback compensation mechanism, dynamically adjust the power output and implement device priority management, and ensure the security of control instructions through encryption verification; Dynamic scheduling module: Monitor the system operation deviation and trigger the rescheduling mechanism, and achieve dynamic optimization and smooth transition of resource allocation on the premise of ensuring device reliability; Health management module: Continuously collect device operation data, update the health index through sliding window statistics and anomaly filtering, and provide dynamic device status evaluation for the whole system.

Images