A low-frequency load shedding intelligent optimization method for local power grid based on reinforcement learning and graph neural network

Abstract

This invention discloses an intelligent optimization method for low-frequency load shedding in local power grids based on reinforcement learning and graph neural networks, belonging to the field of power system frequency control and intelligent load shedding technology. It includes defining evaluation indicators, determining indicator weights using the entropy weight method, optimizing low-frequency load shedding cycles, defining low-frequency load shedding operation constraints, generating initial feeder combinations that satisfy load shedding constraints, performing zoned load shedding balance optimization, and a low-frequency load shedding decision-making method based on graph neural networks and multi-agent reinforcement learning. This invention also integrates a local power grid low-frequency load shedding modeling method that considers feeder net load, load importance, distributed generation output, and network topology characteristics; an optimization method that intelligently generates load shedding cycle schemes under low-frequency disturbance scenarios while meeting frequency and voltage safety constraints; and an intelligent optimization method for low-frequency load shedding in local power grids that balances important load protection, total network load shedding matching, and zoned load shedding balance.

Description

Technical Field

[0001] This invention belongs to the field of power system frequency control and intelligent load shedding technology, specifically involving an intelligent optimization method for low-frequency load shedding in local power grids based on reinforcement learning and graph neural networks. It can be used for dynamic optimization of load shedding cycles and frequency security assurance in power systems with high penetration of new energy and low inertia. Background Technology

[0002] Existing local power grid low-frequency load shedding technologies mainly fall into three categories: The first category is the fixed-cycle load shedding method, which typically pre-sets the operating frequency, load shedding capacity, and shedding sequence for each cycle based on offline tuning results. While simple to implement, it has weak adaptability to changes in operating modes and fluctuations in renewable energy sources. The second category is the adaptive load shedding method, which corrects the load shedding amount based on real-time frequency change rate, power deficit estimation results, or online measurement information. It offers better dynamism than the fixed-cycle method, but still primarily relies on total quantity indicators. The third category is based on optimization or predictive control methods. These methods use power flow calculations, heuristic algorithms, or model predictions to solve the load shedding combination in a rolling manner, which can balance load shedding amount and safety constraints to a certain extent, but requires high model accuracy and computational efficiency.

[0003] Compared with the existing technologies mentioned above, current research on low-frequency load shedding in local power grids still has the following shortcomings: 1. Insufficient characterization of the comprehensive attributes of feeder loads. Although some methods have considered the importance of loads, most still remain at the level of static priority or single weight, failing to uniformly quantify residential loads, industrial loads, feeder net loads, and the reverse support capacity of distributed generation, making it difficult to meet the refined load reduction needs of local power grids.

[0004] 2. Insufficient utilization of local power grid topology and regional differences. Fixed-cycle and partially adaptive methods typically cut off loads in a predetermined order, making it difficult to reflect the topological location of feeders in the network, regional power supply relationships, and power flow changes after the integration of new energy sources. This can easily lead to problems such as excessive load shedding in local areas or incorrect shedding of critical branches.

[0005] 3. Insufficient solution efficiency and global coordination capability under the premise of meeting safety constraints. Methods based on heuristic search or individual optimization have a large computational load when there are many candidate feeders, and are prone to local optima, making it difficult to simultaneously achieve multiple objectives such as frequency recovery, protection of critical loads, and regional load balancing.

[0006] Therefore, existing technologies urgently need a new technical solution to address the above problems. Summary of the Invention

[0007] The purpose of this invention is to overcome the shortcomings of existing technologies and provide a local power grid low-frequency load shedding intelligent optimization method based on reinforcement learning and graph neural networks. This method also integrates feeder net load, load importance, distributed power output, and network topology characteristics into a local power grid low-frequency load shedding modeling method. Furthermore, it provides an optimization method for low-frequency disturbance scenarios that can intelligently generate load shedding rounds while meeting frequency and voltage safety constraints. Finally, it is a local power grid low-frequency load shedding intelligent optimization method that considers important load protection, total load shedding matching across the entire network, and regional load shedding balance.

[0008] To address the aforementioned technical problems, this invention provides a smart optimization method for low-frequency load shedding in local power grids based on reinforcement learning and graph neural networks, comprising the following steps, performed sequentially: Step 1: Establish a comprehensive load importance index model and topological centrality, and use the entropy weight method to determine the weights of the index model; Step 2: Construct optimization objectives and low-frequency load reduction operation constraints based on the preset low-frequency load reduction cycle; Step 3: For the target cycle, form an initial load shedding feeder combination according to the comprehensive characteristics of the feeder load, and verify it to obtain an initial feeder combination scheme that meets the load shedding ratio constraints of the entire network; Step 4: Based on the initial feeder combination scheme that satisfies the load shedding ratio constraint of the entire network, calculate the load shedding ratio of each zone in the current round, and calculate the load shedding ratio deviation of each zone. Based on the degree of deviation between the load shedding ratio of each zone and the balance target, adjust and optimize the feeder configuration corresponding to each zone. Step 5: Based on the joint optimization algorithm of graph neural network and multi-agent reinforcement learning, dynamically adjust the feeder switching scheme of each partition, and output the low-frequency load reduction cycle scheme after optimization and verification.

[0009] The comprehensive load importance index model mentioned in step one is as follows:

[0010] In the formula: For the feeder load; To support the people's livelihood; For non-residential use; This is the number of the distributed power source; For distributed power sources Power generation capacity; considerations for the arrangement of the low-frequency load shedding schedule. >0 feeder; Define topological centrality :

[0011] In the formula: a is the partition number, and N is the number of partitions. This is the proportion of the load in the r-th round to the total load.

[0012] The optimization objectives described in Step Two are as follows: 1. To minimize the sum of the comprehensive attribute indicators of the feeders corresponding to the feeders cut off in each round, so that the load reduction scheme prioritizes avoiding important loads, residential loads, and feeders that provide frequency support; 2. To minimize the deviation of the load reduction ratio in each zone, so that the load reduction ratio undertaken by each zone is balanced, avoiding excessive or insufficient load reduction in local areas; 3. Under the premise of meeting the system frequency recovery requirements, to make the actual load reduction ratio of the entire network close to the target load reduction ratio corresponding to the power deficit during the accident.

[0013] The constraint in step two is: Network-wide load shedding ratio constraint: For the r-th round of load reduction scheme, the corresponding network-wide load shedding ratio is... Then it should satisfy ,in and These are preset lower and upper limits, respectively; frequency safety constraints: after each round of load reduction, the system frequency must not fall below the preset safety threshold, and the recovered frequency after the disturbance should return to the operating range, satisfying the following conditions. ,in and These are preset lower and upper limits; voltage safety constraints: after each round of load shedding, the voltage at the voltage center and the voltage at key nodes remain within the fluctuation range. ,in and These are the preset lower limit and the preset upper limit, respectively.

[0014] The method for adjusting and optimizing the feeder configuration for each partition as described in step four is as follows: For partitions with a load shedding ratio higher than the target value, reduce the portion of feeders already included in the current cycle in that partition; for partitions with a load shedding ratio lower than the target value, increase the portion of feeders not included in the current cycle in that partition; for partitions with bidirectional adjustment space, select to increase or decrease feeders based on the optimization results.

[0015] Step 5 describes the joint optimization algorithm of graph neural networks and multi-agent reinforcement learning, which includes the following steps: Step 501: Abstract the power grid system into a weighted heterogeneous graph, where nodes represent loads, distributed power sources and key equipment, and edges represent feeders and power transmission relationships. Describe the electrical attributes and operating status through node features and edge features. Step 502: Use graph neural networks to extract features from the power grid topology and electrical state. Through the neighborhood information propagation mechanism, obtain the structural importance and operational correlation features of each node or feeder in the whole network, and form a low-dimensional embedded representation for subsequent decision input. Step 503: Model each power grid zone as an independent intelligent agent. Each intelligent agent learns the corresponding feeder switching decision strategy based on the state of its zone and global embedded features. Step 504: Through the collaborative learning mechanism among multiple agents, each partition ensures local rationality, performs coordinated optimization across the entire network, and gradually converges to obtain the optimal load reduction strategy.

[0016] The beneficial effects of this invention are: 1) Protect critical loads: Optimize rotation based on comprehensive load characteristic indicators; 2) Global collaborative and efficient optimization: Distributed MARL reduces the centralized computing load and achieves network-wide optimization; 3) Scalability and portability: Supports migration of power grids or regions of different sizes.

[0017] Furthermore, this invention employs a joint optimization algorithm based on Graph Neural Networks (GNN) and Multi-Agent Reinforcement Learning (MARL) to dynamically adjust the feeder switching schemes for each zone while satisfying the total load shedding constraint. Under complex topologies and renewable energy access conditions, it achieves global coordinated optimization of load reduction strategies, avoiding the local optima problem of traditional rule-based or heuristic methods in multi-constraint, multi-objective scenarios. Compared to traditional methods, the GNN+MARL algorithm used in this invention has the following necessity and advantages: 1) Traditional methods struggle to characterize complex topological relationships due to the strong topological characteristics of power grids, while GNNs can effectively integrate node and adjacency structure information to improve the ability of load reduction decisions to perceive the network structure.

[0018] 2) Applicable to multi-constraint, multi-objective optimization problems. Load reduction problems involve multiple objectives such as frequency safety, load importance and regional balance. Reinforcement learning methods can achieve unified modeling of multiple objectives through reward functions, avoiding the bias caused by human weight setting.

[0019] 3) Supports partitioned distributed decision-making. Using a multi-agent architecture, complex network-wide problems can be decomposed into multiple partitioned sub-problems, improving computational efficiency and enhancing the scalability of the method in large-scale power grids. Attached Figure Description

[0020] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0021] Figure 1 This is a diagram illustrating the overall framework of the GNN+MARL algorithm, a local power grid low-frequency load shedding intelligent optimization method based on reinforcement learning and graph neural networks, as described in this invention. Figure 2 This invention provides a flowchart of a local power grid low-frequency load shedding intelligent optimization method based on reinforcement learning and graph neural networks. Figure 3This is a schematic diagram of a 33-node distribution network structure including new energy sources, illustrating a specific implementation of the intelligent optimization method for low-frequency load shedding in local power grids based on reinforcement learning and graph neural networks according to the present invention. Detailed Implementation

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

[0023] like Figures 1-3 As shown, a smart optimization method for low-frequency load shedding in local power grids based on reinforcement learning and graph neural networks is as follows: Step 1: Definition of Evaluation Indicators The overall load importance index is defined as follows: in, For the feeder load; To support the people's livelihood; For non-residential use; The number of the distributed power source; For distributed power sources The power generation capacity. Arrangement considerations for the low-frequency load shedding schedule. >0 feeder; when When the frequency is less than 0, the feeder will supply power back to the grid because it is connected to distributed power sources, which is beneficial for stabilizing the frequency. Therefore, it cannot be included in the low-frequency load shedding cycle.

[0024] Topological centrality Defined as follows Where 'a' is the partition number and 'N' is the number of partitions. This represents the proportion of the load in the r-th round to the total load. Topology centrality reflects the structural location and interconnectivity of the feeder in the network, while the load shedding ratio for each zone characterizes the share of load reduction borne by each zone. Based on the principle of load shedding distribution, the dispersion of the load shedding ratio for each zone should be minimized to make the overall deviation or standard deviation of each zone as small as possible.

[0025] Step 2: Determine the index weights using the entropy weight method The entropy weight method is used to determine the weights of indicators, helping to reduce subjectivity and more objectively determine the relative importance of indicators. Based on the concept of information entropy, the entropy weight method calculates the entropy value of each indicator to reflect the information richness between them, thereby determining the weight of each indicator. The main idea of ​​the entropy weight method is to determine the weight of each indicator based on the magnitude of its variability. The specific process is as follows: The data is standardized according to the following formula to obtain , : in, For the j-th indicator, .

[0026] The information entropy of each indicator is calculated using the following formula. : in, This represents the weight of the j-th indicator in the i-th scheme relative to the j-th indicator in all schemes; n represents the number of schemes.

[0027] The weight of each indicator is calculated according to the following formula. : Step 3: Optimization Target for Low-Frequency Load Reduction Wheel Schedule For the preset low-frequency load shedding cycles, the following optimization objectives are established: First, to minimize the sum of the comprehensive attribute indicators of the feeders corresponding to the feeders cut off in each cycle, so that the load shedding scheme prioritizes avoiding important loads, residential loads, and feeders with frequency support functions; Second, to minimize the deviation of the load shedding ratio in each zone, so that the load shedding ratio undertaken by each zone is more balanced, avoiding excessive or insufficient load shedding in some areas; Third, under the premise of meeting the system frequency recovery requirements, to make the actual load shedding ratio of the entire network as close as possible to the target load shedding ratio corresponding to the power deficit during the accident. By considering the above objectives simultaneously, the resulting load shedding scheme balances the protection of important loads, regional fairness, and frequency security.

[0028] Step 4: Constraints for Low-Frequency Load Reduction Operation To ensure that the load shedding scheme meets the safety operation requirements of the local power grid, the following constraints are established: Overall grid load shedding ratio constraint: For the r-th round of load shedding scheme, the corresponding overall grid load shedding ratio is... Then it should satisfy ,in and These are preset lower and upper limits, respectively; frequency safety constraints: after each round of load reduction, the system frequency must not fall below the preset safety threshold, and the recovered frequency after the disturbance should return to the allowable operating range, satisfying the following conditions. ,in and These are preset lower and upper limits, respectively; Voltage safety constraints: After each round of load shedding, the voltage at the voltage center point and the voltage at key nodes should remain within the allowable fluctuation range. ,in and These are the preset lower limit and the preset upper limit, respectively.

[0029] Step 5: Generate an initial feeder combination that satisfies the load shearing constraint. For the target cycle, candidate feeders are sorted from low to high according to the comprehensive load characteristics index of the feeders, and the feeders with the highest ranking are selected in turn to form the initial load reduction feeder combination.

[0030] The overall load shedding ratio corresponding to the initial load shedding feeder combination is checked: when the overall load shedding ratio is within a preset range, the feeder combination is used as the initial scheme to meet the load shedding requirements; when the overall load shedding ratio is lower than the preset lower limit, appropriate feeders are added to the candidate feeders that have not been included in this round to increase the load shedding ratio; when the overall load shedding ratio is higher than the preset upper limit, appropriate feeders are removed from the feeders that have been included in this round to reduce the load shedding ratio.

[0031] During the above-mentioned correction process, the current combination can be adjusted by adding feeders, deleting feeders, or replacing feeders with feeders of equal capacity. The feeder combination should be selected from low to high based on the comprehensive attribute index of the feeders, so as to obtain an initial feeder combination scheme that meets the load shedding ratio constraint of the entire network.

[0032] Step Six: Perform Partition Load Balancing Optimization After obtaining the initial feeder configuration that satisfies the overall network load shedding constraints, the load shedding ratio of each zone in the current round is further calculated, and the deviation of the load shedding ratio for each zone is also calculated. Based on the degree of deviation between the load shedding ratio of each zone and the equilibrium target, the feeder configuration corresponding to each zone is adjusted, specifically including: For partitions where the load shedding ratio is higher than the target value, reduce the portion of feeders already included in the current cycle in that partition; for partitions where the load shedding ratio is lower than the target value, increase the portion of feeders not included in the current cycle in that partition; for partitions with bidirectional adjustment space, select to increase or decrease feeders based on the optimization results.

[0033] During the process of adding or removing feeders, adjustments should be made first according to the ranking results of the comprehensive load characteristics index of the feeders, so as to ensure that the risk of cutting off important loads is minimized as much as possible during the balancing optimization process.

[0034] Step 7: A Low-Frequency Load Reduction Decision-Making Method Based on Graph Neural Networks and Multi-Agent Reinforcement Learning Based on the completion of the initial load shedding feeder combination generation and the satisfaction of the overall network load shedding ratio constraint, in order to further realize the partition load shedding balance optimization and dynamic decision optimization, this invention introduces a joint optimization algorithm based on graph neural network (GNN) and multi-agent reinforcement learning (MARL).

[0035] This algorithm is mainly used to solve the following problems: dynamically adjusting the feeder switching scheme of each zone under the premise of satisfying the total load shedding constraint; realizing global coordination and optimization of load reduction strategy under complex topology and new energy access conditions; and avoiding the local optimum problem of traditional rule or heuristic methods in multi-constraint and multi-objective scenarios.

[0036] This method generally adopts a technical approach of "graph representation + feature extraction + distributed decision learning", and its basic principle is as follows: Power grid diagram modeling and feature representation The power grid system is abstracted into a weighted heterogeneous graph structure, where nodes represent loads, distributed power sources and key equipment, and edges represent feeders and power transmission relationships. The electrical attributes and operating status are described through node features and edge features.

[0037] Graph Neural Network Feature Embedding Graph neural networks are used to extract features from the power grid topology and electrical state. Through the neighborhood information propagation mechanism, the structural importance and operational correlation features of each node (or feeder) in the whole network are obtained, thereby forming a low-dimensional embedded representation for subsequent decision input.

[0038] Multi-agent reinforcement learning modeling Each power grid zone is modeled as an independent intelligent agent. Each intelligent agent learns the corresponding feeder switching decision strategy based on the state of its zone and global embedded features.

[0039] Collaborative optimization and strategy iteration Through a collaborative learning mechanism among multiple agents, each partition can achieve coordinated optimization across the entire network while ensuring local rationality, gradually converging to obtain the optimal load reduction strategy.

[0040] Step 8: Review safety constraints and output the final solution. After completing the partitioned load balancing optimization, the adjusted feeder combination was re-verified for the overall network load balancing ratio, system frequency, and voltage center point voltage.

[0041] When the adjusted feeder combination simultaneously meets the above constraints, the feeder combination is output as the final low-frequency load shedding scheme for the target cycle. When the adjusted feeder combination does not meet any constraint, the corresponding combination is discarded, and the feeder adjustment is re-executed until a load shedding scheme that meets all constraints is obtained. The final low-frequency load shedding cycle scheme can be used for low-frequency load shedding execution control under grid fault disturbances.

[0042] One specific implementation method: To verify the effectiveness of the proposed intelligent optimization method for low-frequency load shedding in local power grids based on reinforcement learning and graph neural networks in distribution networks containing renewable energy sources, an improved IEEE 33-node distribution system was selected as the simulation object. This system is a typical radial distribution network structure, divided into five zones based on the characteristics of local power grid operation, with distributed photovoltaic and wind power sources connected at nodes 7, 16, 21, 24, and 30. The system's renewable energy installed capacity was set to account for 31.6% of the total active load, the total load factor was 100%, and the voltage operating range of each node was set to [0.95, 1.05] pu. In the simulation, a power disturbance was introduced, causing an active power deficit of approximately 11.2%, resulting in a rapid drop in system frequency. A round of low-frequency load shedding was triggered when the frequency dropped to 49Hz.

[0043] The algorithm employs a "weighted heterogeneous graph modeling + GNN feature embedding + partitioned multi-agent reinforcement learning" approach for training and solving. State variables include frequency deviation in each partition, node voltage level, load importance index, and real-time output of distributed generation; action variables are the switching decisions for candidate feeders in each partition. For feeders exhibiting reverse power flow characteristics, the algorithm identifies them based on the feeder's net injected power and distributed generation output, and explicitly constrains such feeders from being disconnected to avoid erroneous disconnection of branches containing renewable energy backflow.

[0044] The evaluation indicators for the case study include: the overall network load shedding ratio, the load shedding ratio deviation of each zone, and the load composite index (FLCI). Among them, the overall network load shedding ratio is used to measure whether the load reduction meets the emergency power deficit requirements; the load shedding ratio deviation is used to measure the degree of load sharing balance among different zones; the FLCI is used to evaluate the importance level of the shedding load; and the frequency safety score comprehensively considers the lowest frequency, steady-state recovery frequency, and recovery time to reflect the frequency recovery effect after load shedding.

[0045] In the above scenario, the strategy after training convergence is tested offline, and the load shedding orchestration results for the target round are output. Table 1 shows the key performance indicators obtained by the method of this invention under this condition, and Table 2 shows the partition adjustment after satisfying the network-wide load shedding constraint. The test results show that the method can minimize the shedding of important loads while ensuring frequency security, and also take into account the fairness of load shedding in each partition.

[0046] The simulation results are shown in Tables 1 and 2: Table 1 Optimal Solution

[0047] Table 2: Results of Partition Adjustment 1 0.0864 2 3 2 0.1391 4 3 3 0.1217 4 3 4 0.0798 3 4 5 0.1067 2 2 As shown in Table 1, the load shedding ratio of the present invention under the combined disturbance condition of the distribution network containing new energy sources is 11.38%, which basically matches the active power deficit caused by the accident and retains a certain dynamic adjustment margin; the load shedding ratio deviation is 0.0016, indicating that the load reduction distribution among different zones is relatively balanced; the load comprehensive index FLCI corresponding to the cut feeder is 0.0735, indicating that the overall importance of the cut load is low and it can better avoid residential loads and important loads; the frequency safety score is 0.91, indicating that the minimum frequency of the system after load reduction and the recovery process both meet the safety requirements.

[0048] As shown in Table 2, after the initial scheme was optimized by zonal balancing, zones 1 and 4 each added one feeder to compensate for their initially low load shedding ratio; zones 2 and 3 each removed one feeder to reduce the excessive load shedding in some areas. After these adjustments, the load reduction ratios of each zone are more similar, achieving a balance between total load shedding, protection of critical loads, and regional fairness while meeting frequency safety and voltage constraints.

[0049] Comparative analysis with traditional low-frequency load reduction methods: Strong topology awareness: GNN fully integrates structural and electrical features to improve decision-making accuracy; Critical load protection: Avoiding critical loads based on FLCI and priority scoring; Partition balancing: Multi-agent collaboration achieves fair load balancing and avoids local overload; Highly adaptable: It can be adapted to power grids of different sizes and with different new energy penetration rates.

[0050] In summary, the method of this invention achieves the comprehensive goals of "fewer cuts, more even cuts, and stable frequency" in the fault scenario of a distribution network with 33 nodes containing new energy sources. It takes into account both the frequency recovery index and the importance of residential loads, and uses graph neural networks to perceive the topology of the power grid to avoid accidentally cutting feeders with the reverse power transmission characteristics of distributed sources, thereby improving the intelligent optimization level of low-frequency load shedding in distribution networks with high penetration of new energy sources.

[0051] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

[0052] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A smart optimization method for low-frequency load shedding in local power grids based on reinforcement learning and graph neural networks, characterized in that: Includes the following steps, And the following steps are performed in sequence: Step 1: Establish a comprehensive load importance index model and topological centrality, and use the entropy weight method to determine the weights of the index model; Step 2: Construct optimization objectives and low-frequency load reduction operation constraints based on the preset low-frequency load reduction cycle; Step 3: For the target cycle, form an initial load shedding feeder combination according to the comprehensive characteristics of the feeder load, and verify it to obtain an initial feeder combination scheme that meets the load shedding ratio constraints of the entire network; Step 4: Based on the initial feeder combination scheme that satisfies the load shedding ratio constraint of the entire network, calculate the load shedding ratio of each zone in the current round, and calculate the load shedding ratio deviation of each zone. Based on the degree of deviation between the load shedding ratio of each zone and the balance target, adjust and optimize the feeder configuration corresponding to each zone. Step 5: Based on the joint optimization algorithm of graph neural network and multi-agent reinforcement learning, dynamically adjust the feeder switching scheme of each partition, and output the low-frequency load reduction cycle scheme after optimization and verification.

2. The method of claim 1, wherein the method is based on reinforcement learning and graph neural networks. The comprehensive load importance index model mentioned in step one is as follows: ， In the formula: For the feeder load; To support the people's livelihood; For non-residential use; This is the number of the distributed power source; For distributed power sources Power generation capacity; considerations for the arrangement of the low-frequency load shedding schedule. >0 feeder; Defining a topological centrality : ， In the formula: a is the partition number, N is the number of partitions, is the proportion of the load of the rth round to its total load.

3. The method of claim 1, wherein the method is based on reinforcement learning and graph neural networks for smart optimization of low frequency load shedding in local power grids. The optimization objectives described in Step Two are as follows:

1. To minimize the sum of the comprehensive attribute indicators of the feeders corresponding to the feeders cut off in each round, so that the load reduction scheme prioritizes avoiding important loads, residential loads, and feeders that provide frequency support; 2. To minimize the deviation of the load shedding ratio in each zone, so that the load reduction ratio undertaken by each zone is balanced, avoiding excessive or insufficient load reduction in local areas; 3. Under the premise of meeting the system frequency recovery requirements, to make the actual load shedding ratio of the entire network close to the target load shedding ratio corresponding to the power deficit during the accident.

4. The method of claim 1, wherein: The constraint in step two is: Network-wide load shedding ratio constraint: For the r-th round of load reduction scheme, the corresponding network-wide load shedding ratio is... Then it should satisfy ,in and These are the preset lower limit and the preset upper limit, respectively; Frequency safety constraints: After each round of load reduction, the system frequency must not fall below the preset safety threshold, and the recovered frequency after the disturbance should return to the operating range, satisfying the requirements. ,in and These are the preset lower limit and the preset upper limit, respectively; Voltage security constraint: the voltage of the central point and the voltage of the key node remain in the fluctuation range after each load shedding execution wherein and are the preset lower limit and the preset upper limit, respectively.

5. The method of claim 1, wherein: The method for adjusting and optimizing the feeder configuration for each partition as described in step four is as follows: For partitions with a load shedding ratio higher than the target value, reduce the portion of feeders already included in the current cycle in that partition; for partitions with a load shedding ratio lower than the target value, increase the portion of feeders not included in the current cycle in that partition; for partitions with bidirectional adjustment space, select to increase or decrease feeders based on the optimization results.

6. The method of claim 1, wherein: Step 5 describes the joint optimization algorithm of graph neural networks and multi-agent reinforcement learning, which includes the following steps: Step 501: Abstract the power grid system into a weighted heterogeneous graph, where nodes represent loads, distributed power sources and key equipment, and edges represent feeders and power transmission relationships. Describe the electrical attributes and operating status through node features and edge features. Step 502: Use graph neural networks to extract features from the power grid topology and electrical state. Through the neighborhood information propagation mechanism, obtain the structural importance and operational correlation features of each node or feeder in the whole network, and form a low-dimensional embedded representation for subsequent decision input. Step 503: Model each power grid zone as an independent intelligent agent. Each intelligent agent learns the corresponding feeder switching decision strategy based on the state of its zone and global embedded features. Step 504: Through the collaborative learning mechanism among multiple agents, each partition ensures local rationality, performs coordinated optimization across the entire network, and gradually converges to obtain the optimal load reduction strategy.

Images