A power system dynamic resource allocation method, system, terminal and storage medium

Abstract

The application discloses a kind of power system dynamic resource allocation method, system, terminal and storage medium, belong to power system intelligent allocation technical field, the method is: based on the multi-source perception data of each allocation node constructs power system full-dimension state perception network, the fault situation characteristic extraction of multi-source perception data is carried out, and the real-time state of power system and corresponding resource demand data are obtained;Build multi-agent reinforcement learning model, configure independent agent in each allocation node, and generate target resource allocation strategy according to the real-time state of power system and each independent agent;Target resource allocation strategy is converted into the allocation task of each independent agent based on rule knowledge base, to allocate dynamic resource of power system.Therefore, the application can generate resource allocation decision and execute corresponding resource allocation task according to full-dimension perception result when facing emergency situation, improve power grid operation reliability.

Description

Technical Field

[0001] This invention relates to the field of intelligent power system dispatching technology, and in particular to a method, system, terminal and storage medium for dynamic resource dispatching in a power system. Background Technology

[0002] As power system grid structures become increasingly complex and the scale of new energy integration continues to expand, the types of faults faced by power grids during operation are becoming more diverse. Sudden risks such as physical shocks and cyber-physical coupling faults are currently the main threats restricting the safe and stable operation of the power grid. Under the traditional power management and control model, the dimensions of state perception are limited and the fault location response is lagging, making it difficult to quickly grasp the true state of the power grid and complete resource scheduling in emergency situations. This can easily lead to the expansion of the fault's impact range and cannot adapt to the development requirements of high-reliability operation of the power grid in the new era.

[0003] Currently, existing technologies primarily employ a centralized decision-making architecture combined with human experience and static scheduling rules to generate resource allocation schemes. While gradually integrating IoT sensing and AI algorithms for status monitoring and resource allocation optimization, these technologies face challenges in handling emergencies. The sensing system struggles to effectively identify cyber-physical coupling faults and lacks accurate location tracking. Centralized decision-making suffers from high computational complexity and slow emergency response. Furthermore, the lack of efficient collaboration mechanisms among allocation nodes easily leads to imbalances where regional resource backlogs and shortages coexist. In emergencies, rapid resource replenishment and coordinated handling are impossible. Therefore, existing technologies lack a flexible resource scheduling mechanism to meet the operational requirements of the distribution network, resulting in poor network reliability. Summary of the Invention

[0004] This invention proposes a dynamic resource allocation method, system, terminal, and storage medium for power systems. It addresses the problem of poor grid reliability caused by the lack of flexible resource scheduling mechanisms in existing technologies when facing emergencies, which fails to meet the operational requirements of the distribution network. This invention enables each allocation node to flexibly generate resource allocation decisions and execute corresponding resource allocation tasks based on comprehensive sensing results during sudden emergencies, thereby improving grid operational reliability.

[0005] To achieve the above objectives, embodiments of the present invention provide a method for dynamic resource allocation in a power system, comprising: constructing a power system full-dimensional state perception network based on pre-collected multi-source sensing data from each allocation node; extracting fault state features from the multi-source sensing data based on the power system full-dimensional state perception network to obtain real-time power system state and corresponding resource demand data; constructing a multi-agent reinforcement learning model based on the multi-source sensing data and the resource demand data; configuring corresponding independent agents in each allocation node based on the multi-agent reinforcement learning model, and generating a target resource allocation strategy according to the real-time power system state and each independent agent through a preset near-end decision optimization algorithm; converting the target resource allocation strategy into several allocation tasks corresponding to each independent agent based on a preset rule knowledge base, and executing the corresponding allocation tasks in each independent agent to dynamically allocate power system resources.

[0006] This invention proposes a dynamic resource allocation method for power systems. It establishes a comprehensive power system state perception network to accurately perceive various types of fault characteristics and predict resource demand data. Then, a multi-agent distributed decision-making architecture is constructed, configuring corresponding independent agents in each allocation node. A pre-defined near-end decision optimization algorithm is used to generate target resource allocation strategies, transforming resource allocation decision-making from centralized to distributed, reducing computational complexity and shortening allocation response time. Finally, a rule transformation mechanism converts the target resource allocation strategy into allocation tasks for each independent agent, completing the dynamic resource allocation of the power system and ensuring the feasibility of the allocation tasks. Thus, the entire process from perception and decision-making to execution optimizes the ability to respond to emergencies. This enables each allocation node to flexibly generate resource allocation decisions and execute corresponding resource allocation tasks based on comprehensive perception results when facing sudden emergencies, improving the reliability of power grid operation.

[0007] Furthermore, the construction of a power system full-dimensional state perception network based on the pre-collected multi-source sensing data of each dispatch node includes: preprocessing the multi-source sensing data of each dispatch node and dividing the preprocessed multi-source sensing data into real-time monitoring sequences and historical correlation sequences; building a real-time sensing layer based on the real-time monitoring sequences; building a trend sensing layer based on the historical correlation sequences; and setting up a two-way information interaction channel between the real-time sensing layer and the trend sensing layer based on a preset two-way information flow mechanism to obtain the power system full-dimensional state perception network.

[0008] In the aforementioned scheme, multi-source sensing data is preprocessed and time-series segmented to construct a real-time sensing layer and a trend sensing layer, establishing a two-way information exchange channel. This constitutes a full-dimensional state sensing network for the power system, improving the comprehensiveness and accuracy of fault detection. The two-layer sensing structure captures instantaneous state fluctuations and long-term fault evolution patterns of the power grid, respectively, and the two-way information exchange mechanism enables mutual correction and supplementation of the sensing results from both layers. Therefore, this approach ensures both the sensitivity of real-time monitoring of sudden faults and the accuracy of fault development trend prediction. It allows the full-dimensional state sensing network of the power system to cover more fault types, providing more comprehensive and reliable state data support for subsequent resource allocation decisions, and contributing to improved power grid operational reliability.

[0009] Furthermore, based on the power system's full-dimensional state perception network, fault state features are extracted from the multi-source sensing data to obtain the real-time state of the power system and corresponding resource demand data. This includes: based on the power system's full-dimensional state perception network, fault state features are extracted from the real-time monitoring sequence through the real-time perception layer to obtain instantaneous state features, and these instantaneous state features are transmitted to the trend perception layer through the bidirectional information interaction channel to update the parameters of the trend perception layer, thereby optimizing the trend perception layer; based on the power system's full-dimensional state perception network, fault state features are extracted from the historical correlation sequence through the trend perception layer to obtain fault evolution features, and these fault evolution features are embedded into the input of the real-time perception layer. An optimized real-time perception layer is obtained. Based on the optimized trend perception layer and the optimized real-time perception layer, the power system full-dimensional state perception network is optimized and trained using a preset multi-scale loss function to obtain an optimized power system full-dimensional state perception network. The optimized real-time perception layer output result and the optimized trend perception layer output result are obtained through the optimized power system full-dimensional state perception network. The optimized real-time perception layer output result and the optimized trend perception layer output result are integrated according to a preset integration algorithm to obtain the initial fault perception result and the corresponding initial resource demand data. The initial fault perception result and the corresponding initial resource demand data are calibrated according to the preset power system network topology constraints to obtain the real-time state of the power system and the corresponding resource demand data.

[0010] In the above scheme, a two-way information exchange channel is used to transmit and optimize the parameters of the two-layer sensing network. A multi-scale loss function is used to complete the optimization training of the sensing network. Then, the dual-output results are integrated and calibrated by power system network topology constraints to finally obtain the real-time status and resource demand data of the power grid. Thus, adaptive optimization of the sensing network is achieved through error gradient feedback and evolutionary feature embedding. The multi-scale loss function is used to adapt to the sensing training requirements of different time scales. Combined with topology constraint calibration, the fault location results are ensured to conform to the physical operation logic of the power grid. Finally, reliable real-time status of the power system and corresponding resource demands are output, providing accurate data input for subsequent dispatch decisions and helping to improve the reliability of power grid operation.

[0011] Furthermore, constructing a multi-agent reinforcement learning model based on the multi-source perception data and the resource demand data includes: constructing state vectors for each allocation node based on the multi-source perception data and the resource demand data; constructing an adjacency matrix based on a preset topology between the allocation nodes; constructing action vectors for each allocation node based on the multi-source perception data; constructing a global activation function value and several local activation function values ​​based on the multi-source perception data and the resource demand data, and obtaining a multi-objective activation function based on pre-constructed target weight coefficients, the global activation function value, and the several local activation function values; and constructing a multi-agent reinforcement learning model corresponding to each allocation node based on the state vector, adjacency matrix, action vector, and multi-objective activation function corresponding to each allocation node.

[0012] In the above scheme, a corresponding state vector, adjacency matrix, and action vector are constructed for each dispatch node. Then, multi-source sensing data and resource demand data are used to construct a global incentive function and several local incentive function values ​​to build a multi-objective incentive function. Finally, a multi-agent reinforcement learning model is constructed by combining the state vector, adjacency matrix, action vector, and multi-objective incentive function of each dispatch node. Thus, by adapting a single dispatch node as an independent agent to the distributed physical architecture of the power grid, and by using an adjacency matrix to limit the physical boundaries of resource dispatch to ensure the physical feasibility of the decision, the constructed multi-objective incentive function simultaneously considers the local dispatch efficiency of nodes and the global resource balance of the power grid. This allows the dispatch strategy generated by the multi-agent reinforcement learning model to simultaneously meet the core objectives of local fault handling and global resource optimization, ensuring the reliability of power grid operation.

[0013] Furthermore, based on the multi-source sensing data and the resource demand data, a global incentive function value and several local incentive function values ​​are constructed. Based on the pre-constructed target weight coefficients, the global incentive function value, and the several local incentive function values, a multi-objective incentive function is obtained, including: classifying fault levels based on the multi-source sensing data to determine the emergency resource allocation benchmark corresponding to each fault level; calculating the standardized resource supply-demand deviation based on the emergency resource allocation benchmark corresponding to each fault level and the resource demand data; calculating the resource balance incentive value based on the standardized resource supply-demand deviation and a preset time decay factor; calculating the immediate handling efficiency, stage handling efficiency, and overall handling efficiency of each allocation node based on the multi-source sensing data; and calculating the fault handling efficiency incentive value based on the immediate handling efficiency, the stage handling efficiency, and the overall handling efficiency; based on the... The process involves calculating the consumption cost of multi-source sensing data and then calculating the allocation consumption incentive value based on the consumption cost. It also involves calculating the resource allocation difference ratio between adjacent allocation nodes and constructing a difference matrix. Based on the multi-source sensing data, it calculates the distance-delay weight coefficient for each adjacent allocation node and calculates the global assistance incentive value based on the difference matrix and the distance-delay weight coefficient. The process uses the resource balance incentive value, fault handling efficiency incentive value, and allocation consumption incentive value as several local incentive function values, and the global assistance incentive value as the global incentive function value. Initial weight coefficients are set for the several local incentive function values ​​and the global incentive function value, and weight constraints are set on the initial weight coefficients to obtain target weight coefficients. Based on the target weight coefficients, a weighted sum is performed on the global incentive function value and the several local incentive function values ​​to obtain a multi-objective incentive function.

[0014] In the aforementioned scheme, fault levels are classified and corresponding emergency resource allocation benchmarks are determined. Incentive function values ​​are constructed from multiple dimensions, including resource balance, response efficiency, allocation consumption, and global assistance. The weights of each incentive function value are dynamically adjusted based on the degree of performance improvement to form a multi-objective incentive function. Thus, by matching emergency resource allocation benchmarks to fault levels, the uniformity of response standards is ensured. Multi-dimensional incentive indicators are calculated to cover all performance requirements of resource allocation. Furthermore, a dynamic weight adjustment mechanism allows the objective incentive function to adapt to the optimization emphasis of different fault scenarios, guiding the agent to generate resource allocation strategies that better meet actual operational needs, thereby contributing to improved power grid reliability.

[0015] Furthermore, based on the multi-agent reinforcement learning model, corresponding independent agents are configured in each of the allocation nodes, and a target resource allocation strategy is generated according to the real-time state of the power system and each of the independent agents through a preset proximal decision optimization algorithm. This includes: configuring corresponding independent agents for each allocation node based on the multi-agent reinforcement learning model corresponding to each allocation node, and constructing a local sharing area, a cooperative sharing area, and a global sharing area in each independent agent; training each of the independent agents through a preset proximal decision optimization algorithm to generate several state transition samples; calculating the importance score of several state transition samples based on a preset importance scoring algorithm; and storing several state transition samples in corresponding... The system comprises a local shared area, a cooperative shared area, or a global shared area; calculating the sampling probability of several state transition samples, and selecting several training samples from the local shared area, the cooperative shared area, and the global shared area according to the sampling probability; updating the decision network parameters of each independent agent based on the several training samples to obtain the independent agents with updated decision networks; constructing a weight matrix based on the historical decision data of the independent agents with updated decision networks and the preset topology; optimizing the decision networks of the independent agents with updated decision networks according to the weight matrix and the preset near-end decision optimization algorithm to obtain several target independent agents; and generating a target resource allocation strategy through the several target independent agents based on the real-time state of the power system.

[0016] In the above scheme, a three-layer shared area system (local, collaborative, and global) is constructed to store state transition samples. Samples are stored hierarchically based on importance scores. Training samples are selected from each shared area by calculating sampling probabilities to update the decision network parameters. A weight matrix is ​​then constructed based on the topology and historical decision data. The decision network is then optimized at multiple scales using a near-end decision optimization algorithm to obtain the target independent agent and generate the target resource allocation strategy. Thus, the hierarchical shared area system enables the classification and reuse of decision-making experience across different scopes. The importance scoring mechanism prioritizes the retention of high-value training samples, balancing individual decision-making, neighborhood collaboration, and global optimization during hierarchical parameter updates. Furthermore, the construction of the weight matrix and the use of the near-end decision optimization algorithm to optimize the updated decision network accelerate the convergence speed of the multi-agent reinforcement learning model within the independent agent, improve the multi-objective optimization effect and scenario adaptability of the allocation strategy, and enhance the reliability of power grid operation.

[0017] Furthermore, based on a preset rule knowledge base, the target resource allocation strategy is converted into several allocation tasks corresponding to independent intelligent agents, and the corresponding allocation tasks are executed in each of the independent intelligent agents to allocate dynamic resources of the power system. This includes: parsing the target resource allocation strategy to obtain an initial allocation scheme; performing constraint verification on the initial allocation scheme based on the preset rule knowledge base and calculating the rule matching degree score of the initial allocation scheme; if the initial allocation scheme does not meet the preset constraint verification requirements or the rule matching degree score of the initial allocation scheme does not meet the preset matching degree requirements, then the initial allocation scheme is modified until it is completely modified. The initial allocation scheme after correction satisfies the preset constraint verification requirements, and the rule matching degree score of the modified initial allocation scheme satisfies the preset matching degree requirements, thus obtaining a modified allocation scheme. Based on the pre-acquired historical allocation execution data, the execution risk probability of the modified allocation scheme in different preset time windows is calculated, and an execution priority is generated according to the execution risk probability of the modified allocation scheme in different preset time windows. The modified allocation schemes are sorted according to the execution priority to obtain allocation tasks. The allocation tasks are distributed to each of the independent intelligent agents so that each of the independent intelligent agents executes the corresponding allocation task to dynamically allocate resources in the power system.

[0018] In the above scheme, the target resource allocation strategy is parsed into an initial allocation plan. Constraint verification and automatic correction are performed using a pre-set rule knowledge base. Execution risks are assessed based on historical execution data, and execution priorities are generated. These are then converted into allocation tasks and distributed to each intelligent agent for execution. Thus, by verifying power industry rules, the allocation plan ensures compliance with physical constraints and business management specifications. Risk prioritization improves the reliability of allocation task execution, thereby enhancing the reliability of power grid operation.

[0019] This invention also provides a dynamic resource allocation system for a power system, including a network construction module, a feature extraction module, a model training module, a strategy generation module, and a resource allocation module. The network construction module is used to construct a full-dimensional state perception network for the power system based on pre-collected multi-source sensing data from each allocation node. The feature extraction module is used to extract fault state features from the multi-source sensing data based on the full-dimensional state perception network to obtain the real-time state of the power system and corresponding resource demand data. The model training module is used to construct a multi-agent reinforcement learning model based on the multi-source sensing data and the resource demand data. The strategy generation module is used to configure corresponding independent agents in each allocation node based on the multi-agent reinforcement learning model, and generate a target resource allocation strategy according to the real-time state of the power system and each independent agent using a preset near-end decision optimization algorithm. The resource allocation module is used to convert the target resource allocation strategy into several allocation tasks corresponding to each independent agent based on a preset rule knowledge base, and execute the corresponding allocation tasks in each independent agent to allocate dynamic resources of the power system.

[0020] This invention proposes a dynamic resource allocation system for power systems. It establishes a comprehensive power system state perception network to accurately perceive various types of fault characteristics and predict resource demand data. Then, it constructs a multi-agent distributed decision-making architecture, configuring corresponding independent agents in each allocation node. A pre-defined near-end decision optimization algorithm is used to generate target resource allocation strategies, transforming resource allocation decision-making from centralized to distributed, reducing computational complexity and shortening allocation response time. Finally, a rule transformation mechanism converts the target resource allocation strategy into allocation tasks for each independent agent, completing the dynamic resource allocation of the power system and ensuring the feasibility of the allocation tasks. Thus, the entire process from perception and decision-making to execution optimizes the ability to respond to emergencies. This enables each allocation node to flexibly generate resource allocation decisions and execute corresponding resource allocation tasks based on comprehensive perception results when facing sudden emergencies, improving the reliability of power grid operation.

[0021] A power system dynamic resource allocation terminal includes a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor. When the processor executes the computer program, it implements the power system dynamic resource allocation method.

[0022] A storage medium for a dynamic resource allocation method in a power system is provided. The storage medium includes a stored computer program, wherein the computer program controls the device where the storage medium is located to execute the dynamic resource allocation method in a power system during runtime. Attached Figure Description

[0023] To more clearly illustrate the technical solution of this application, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0024] Figure 1 This is a flowchart illustrating the steps of a dynamic resource allocation method for a power system according to a certain embodiment of the present invention. Figure 2 This is a schematic diagram of the module structure of a dynamic resource allocation system for a power system according to a certain embodiment of the present invention. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below with reference to the accompanying drawings of the embodiments. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0026] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains; the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the application; the terms “comprising” and “having”, and any variations thereof, in the specification, claims, and foregoing description of the drawings are intended to cover non-exclusive inclusion.

[0027] In the description of the embodiments of this application, technical terms such as "first" and "second" are used only to distinguish different objects and should not be construed as indicating or implying relative importance or implicitly specifying the number, specific order, or primary and secondary relationship of the indicated technical features. In the description of the embodiments of this application, "multiple" means two or more, unless otherwise explicitly defined.

[0028] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0029] In the description of the embodiments in this application, the term "and / or" is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, and B existing alone. Additionally, the character " / " in this document generally indicates that the preceding and following related objects have an "or" relationship.

[0030] In the description of the embodiments of this application, the term "multiple" refers to two or more (including two), similarly, "multiple sets" refers to two or more (including two sets), and "multiple pieces" refers to two or more (including two pieces).

[0031] In the description of the embodiments of this application, unless otherwise expressly specified and limited, technical terms such as "installation," "connection," "joining," and "fixing" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral part; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal communication of two components or the interaction between two components. For those skilled in the art, the specific meaning of the above terms in the embodiments of this application can be understood according to the specific circumstances.

[0032] To address the problem of poor grid reliability caused by the lack of flexible resource scheduling mechanisms in existing technologies when facing emergencies, thus failing to meet the operational requirements of the distribution network, see [reference needed]. Figure 1 , Figure 1 This is a flowchart illustrating the steps of a dynamic resource allocation method for a power system according to a certain embodiment of the present invention. Figure 1 As shown in the figure, this embodiment of the invention proposes a method for dynamic resource allocation in a power system, including steps 101 to 105, each step of which is as follows: Step 101: Construct a full-dimensional state perception network for the power system based on the pre-collected multi-source sensing data from each dispatch node; Step 102: Based on the power system full-dimensional state perception network, extract fault state features from the multi-source perception data to obtain the real-time state of the power system and the corresponding resource demand data. Step 103: Construct a multi-agent reinforcement learning model based on the multi-source perception data and the resource demand data; Step 104: Based on the multi-agent reinforcement learning model, configure corresponding independent agents in each of the allocation nodes, and generate a target resource allocation strategy according to the real-time status of the power system and each of the independent agents through a preset near-end decision optimization algorithm. Step 105: Based on a preset rule knowledge base, the target resource allocation strategy is converted into several allocation tasks corresponding to each independent intelligent agent, and the corresponding allocation tasks are executed in each independent intelligent agent to allocate dynamic resources of the power system.

[0033] One possible implementation involves collecting multi-source sensing data, including power equipment operation data, historical fault handling data, real-time communication link status data, and cyber-physical coupling correlation data covering all nodes. The collected multi-source sensing data undergoes sequential data cleaning, outlier correction, and standardization to eliminate interference caused by differences in data volume and quality. The pre-processed data is then divided into real-time monitoring sequences and historical correlation sequences according to their time scale and business attributes. Real-time sensing layers for the corresponding real-time monitoring sequences and trend sensing layers for the corresponding historical correlation sequences are then constructed using these sequences. Simultaneously, a bidirectional information exchange channel is established between the two sensing networks, enabling bidirectional transmission of feature information and optimization signals, ultimately forming a full-dimensional state sensing network for the power system.

[0034] In a full-dimensional state perception network, a real-time perception layer extracts features from the input real-time monitoring sequence to capture fault characteristics in the instantaneous operation of the power grid. A trend perception layer performs in-depth analysis of the input historical correlation sequence to uncover long-term situational characteristics of fault evolution and cyber-physical coupling. During feature extraction, the perception error gradient output by the real-time perception layer is passed to the parameter update stage of the trend perception layer, driving it to optimize its evolutionary analysis capabilities based on real-time operating conditions. Simultaneously, the fault evolution features extracted by the trend perception layer are embedded into the input of the real-time perception layer using existing time-scale alignment and feature dimension concatenation methods, enabling the real-time perception layer to... It can perform predictive sensing based on fault development trends, achieve collaborative optimization of two-layer sensing networks through a two-way information exchange channel, and then use a multi-scale loss function to iteratively optimize the sensing network. The loss term is dynamically weighted according to the time scale characteristics of the two-layer network to adapt the loss function to the optimization objectives of different sensing levels. After the network training is completed, the output results of the two-layer sensing network are integrated and fused to obtain the initial fault sensing and location results. Then, the initial results are calibrated in combination with the physical constraints of the power system network topology to ensure that the fault location results conform to the physical logic of power grid operation. Finally, the real-time status of the power system and the corresponding resource demand data are output.

[0035] Then, a corresponding state vector is constructed for each power dispatch node to fully represent the operation and resource status of a single node. Next, an adjacency matrix is ​​constructed based on the power transmission channel connection and communication link status between dispatch nodes. The adjacency matrix represents the physical connection relationship and cooperation priority between nodes, limits the physical boundary of agent interaction and resource dispatch, and ensures that the subsequent generated decisions are physically feasible. At the same time, a corresponding action space is constructed for each independent agent, and a multi-objective incentive function including several local incentive dimensions and global incentive dimensions is constructed. Finally, the state vector, adjacency matrix, action space and multi-objective incentive function are combined to construct a multi-agent reinforcement learning model.

[0036] Next, based on the multi-agent reinforcement learning model, each allocation node is configured with a corresponding independent agent. Within each independent agent, a hierarchical sharing region consisting of a local sharing region, a cooperative sharing region, and a global sharing region is constructed. This hierarchical sharing region is used to hierarchically store the state transition samples generated during training. Then, the importance of each state transition sample is calculated, and samples are stored in the corresponding hierarchical sharing region according to their importance. The sampling probability of each sample is then calculated by combining the timeliness of the sample with the number of decision updates. High-value, high-timeliness samples are preferentially selected from the local, cooperative, and global sharing regions as training samples. Subsequently, state transition samples from different levels of sharing regions are used to update the decision network parameters in different ranges. During training, a weight matrix is ​​also constructed, and the objective function of the near-end decision optimization algorithm is decomposed into immediate optimization terms, periodic optimization terms, and long-term optimization terms. The weight coefficients of each optimization term are calculated by combining the weight matrix with the real-time state information of the power system, achieving decision optimization at multiple time scales. After completing the decision network training, the real-time state of the power system is input into the trained decision network to generate a target resource allocation strategy that takes into account the requirements of multi-objective optimization.

[0037] Finally, the abstract target resource allocation strategy is parsed into an initial allocation plan, including allocation node pairs, emergency resource types, allocation quantities, and allocation time windows. A preset rule knowledge base is then invoked to perform multi-dimensional constraint verification on the initial allocation plan. Simultaneously, the rule matching degree of the initial plan is calculated, and allocation plans that fail constraint verification or whose matching degree is insufficient are automatically corrected by adjusting the allocation sequence, resource configuration, or execution path to meet the rule requirements. After correction, based on historical allocation execution data of the power system, the execution risk of the corrected plan in different time windows is assessed. Combining the risk assessment results with fault handling priorities, the execution priority of the plan is generated. After sorting the allocation plans according to execution priority, they are transformed into allocation tasks that clearly include allocation objects, allocation quantities, allocation time, and execution entities. The allocation tasks are then distributed to the independent computing units of each allocation node, and each node executes the corresponding tasks sequentially according to priority.

[0038] It is worth mentioning that during task execution, if an abnormality in the status perception result or a resource allocation conflict is detected, the faulty node automatically initiates an emergency assistance request to the adjacent allocation node. The node receiving the request returns response information based on its own resource status and collaboration history. An emergency handling plan is reached through local negotiation between the nodes. At the same time, the operating status and response latency of the participating negotiation nodes are continuously monitored. When a node is detected to be operating abnormally or the response times out, the unfinished negotiation tasks of the abnormal node are obtained. The task reception adaptability of each node is calculated based on the resource status, response capability and transmission distance of the remaining nodes. The unfinished tasks are redistributed under the premise of meeting resource constraints.

[0039] This invention proposes a dynamic resource allocation method for power systems. It establishes a comprehensive power system state perception network to accurately perceive various types of fault characteristics and predict resource demand data. Then, a multi-agent distributed decision-making architecture is constructed, configuring corresponding independent agents in each allocation node. A pre-defined near-end decision optimization algorithm is used to generate target resource allocation strategies, transforming resource allocation decision-making from centralized to distributed, reducing computational complexity and shortening allocation response time. Finally, a rule transformation mechanism converts the target resource allocation strategy into allocation tasks for each independent agent, completing the dynamic resource allocation of the power system and ensuring the feasibility of the allocation tasks. Thus, the entire process from perception and decision-making to execution optimizes the ability to respond to emergencies. This enables each allocation node to flexibly generate resource allocation decisions and execute corresponding resource allocation tasks based on comprehensive perception results when facing sudden emergencies, improving the reliability of power grid operation.

[0040] In a preferred embodiment, the construction of a power system full-dimensional state perception network based on pre-collected multi-source sensing data from each dispatch node includes: preprocessing the multi-source sensing data from each dispatch node and dividing the preprocessed multi-source sensing data into real-time monitoring sequences and historical correlation sequences; building a real-time sensing layer based on the real-time monitoring sequences; building a trend sensing layer based on the historical correlation sequences; and setting up a bidirectional information interaction channel between the real-time sensing layer and the trend sensing layer based on a preset bidirectional information flow mechanism to obtain the power system full-dimensional state perception network.

[0041] For example, the data collected covers power equipment operation data, historical fault handling data, real-time communication link status data, and cyber-physical coupling correlation data at each node as multi-source sensing data. A unified preprocessing operation is performed on the multi-source sensing data, which sequentially completes data cleaning, missing value imputation, outlier correction, and data standardization. The preprocessing operation can be implemented using existing technologies and will not be elaborated here. Then, the preprocessed multi-source sensing data is divided into real-time monitoring sequences and historical correlation sequences. The real-time monitoring sequences include instantaneous sensing data such as short-cycle updated equipment operation status and real-time communication link status, which are used to capture instantaneous state fluctuations and sudden anomalies in the power system. The historical correlation sequences include long-term accumulated historical fault records, physical impact handling records, and cyber-physical coupling fault evolution data, which are used to analyze fault development patterns and coupling correlation characteristics.

[0042] A real-time perception layer is constructed using the segmented real-time monitoring sequences. This layer employs a long short-term memory network structure with an attention mechanism, enabling it to dynamically adjust the duration of the sensing and monitoring window based on multi-source sensing data. For example, a 5-minute monitoring window is set for fault sensing of transmission lines, while a 1-minute monitoring window is set for fault sensing of core substation equipment. Then, a trend perception layer is constructed using historical correlation sequences. This layer employs a bidirectional temporal memory network structure with a memory enhancement module. The memory enhancement module stores and preserves long-term historical features such as fault evolution patterns and cyber-physical coupling correlation laws, enabling the discovery of coupling evolution laws between equipment faults and communication link anomalies under extreme operating conditions and physical impact scenarios.

[0043] After completing the construction of the two-layer sensing network, a two-way information interaction channel is established between the real-time sensing layer and the trend sensing layer according to the preset two-way information flow mechanism. The sensing error gradient output by the real-time sensing layer is transmitted back to the parameter update stage of the trend sensing layer through the two-way information interaction channel. This drives the trend sensing layer to dynamically optimize its own evolution analysis model in combination with the actual real-time operation, thereby improving the fit between the situation assessment results and the actual operating conditions. Then, the fault evolution features extracted by the trend sensing layer are forward-embedded into the input of the real-time sensing layer, enabling the real-time sensing layer to conduct predictive sensing in combination with the fault development trend and identify potential fault evolution risks in advance. Finally, the real-time sensing layer and the trend sensing layer are integrated to construct a full-dimensional state sensing network for the power system.

[0044] In the aforementioned scheme, multi-source sensing data is preprocessed and time-series segmented to construct a real-time sensing layer and a trend sensing layer, establishing a two-way information exchange channel. This constitutes a full-dimensional state sensing network for the power system, improving the comprehensiveness and accuracy of fault detection. The two-layer sensing structure captures instantaneous state fluctuations and long-term fault evolution patterns of the power grid, respectively, and the two-way information exchange mechanism enables mutual correction and supplementation of the sensing results from both layers. Therefore, this approach ensures both the sensitivity of real-time monitoring of sudden faults and the accuracy of fault development trend prediction. It allows the full-dimensional state sensing network of the power system to cover more fault types, providing more comprehensive and reliable state data support for subsequent resource allocation decisions, and contributing to improved power grid operational reliability.

[0045] A preferred embodiment involves extracting fault state features from the multi-source sensing data based on the power system's full-dimensional state perception network to obtain real-time power system status and corresponding resource demand data. This includes: extracting fault state features from the real-time monitoring sequence using a real-time sensing layer based on the power system's full-dimensional state perception network to obtain instantaneous state features; transmitting these instantaneous state features to the trend perception layer via the bidirectional information interaction channel to update the parameters of the trend perception layer, thereby optimizing the trend perception layer; and extracting fault state features from the historical correlation sequence using the trend perception layer based on the power system's full-dimensional state perception network to obtain fault evolution features, embedding these fault evolution features into the input of the real-time sensing layer. The system obtains an optimized real-time perception layer. Based on the optimized trend perception layer and the optimized real-time perception layer, the power system full-dimensional state perception network is optimized and trained using a preset multi-scale loss function to obtain an optimized power system full-dimensional state perception network. The optimized real-time perception layer output result and the optimized trend perception layer output result are obtained through the optimized power system full-dimensional state perception network. The optimized real-time perception layer output result and the optimized trend perception layer output result are integrated according to a preset integration algorithm to obtain initial fault perception results and corresponding initial resource demand data. The initial fault perception results and the corresponding initial resource demand data are calibrated according to preset power system network topology constraints to obtain the real-time state of the power system and the corresponding resource demand data.

[0046] For example, the real-time perception layer extracts fault state features from the input real-time monitoring sequence at each time step. It dynamically adjusts the granularity of the perception and monitoring window for different types of power equipment such as transformers, switchgear, and transmission lines, as well as different fault modes such as short circuits, open circuits, and equipment degradation. It adapts more refined monitoring durations to core substation equipment, accurately captures instantaneous abnormal signals such as equipment operation fluctuations and communication link status changes, and obtains instantaneous state features that can characterize the instantaneous operating state of the power grid. At the same time, the perception error gradient calculated by the real-time perception layer is transmitted back to the trend perception layer through a two-way information interaction channel, participates in the network parameter iterative update process of the trend perception layer, and drives the trend perception layer to correct its own evolutionary analysis logic in combination with the current real-time operating conditions of the power grid, resulting in an optimized trend perception layer with optimized parameters. By performing deep temporal feature mining on the input historical correlation sequence through the trend perception layer, and using the built-in memory enhancement module to store and precipitate the historical evolution patterns and cyber-physical coupling correlation laws of various faults, the long-term coupling evolution characteristics of equipment failure and communication link anomalies under extreme working conditions, physical impact and other scenarios are captured, and fault evolution features that can characterize the fault development trend are obtained. Then, the extracted fault evolution features are positively embedded into the input end of the real-time perception layer through a two-way information interaction channel, so that the real-time perception layer can make predictive perceptions in combination with the fault development trend when conducting instantaneous state analysis, resulting in an optimized real-time perception layer with enhanced features.

[0047] By combining the optimization trend perception layer and the optimization real-time perception layer, end-to-end optimization training of the entire power system's full-dimensional state perception network is performed using a preset multi-scale loss function. In this embodiment, the preset multi-scale loss function includes a real-time perception loss term, a trend perception loss term, and a regularization constraint term, expressed as follows: ; In the formula, This is the total loss function; Loss due to real-time sensing layer; Loss in the trend perception layer; and For dynamically weighted coefficients, satisfying ; The regularization coefficient is used. These are the network model parameters; For regularization constraints.

[0048] In this embodiment, during the training of the perception network, the weight coefficients of each loss term are dynamically adjusted according to the time scale characteristics of the fault development stage and the perception level. For example, the weight of the real-time perception loss term is increased during the fault outbreak stage, and the weight of the trend perception loss term is increased during the fault evolution stage. Through the dynamic weighted multi-scale loss function, collaborative optimization is performed to obtain the optimized power system full-dimensional state perception network that has been trained and converged. The multi-source sensing data to be analyzed is input into the optimized power system full-dimensional state sensing network. The instantaneous sensing results output by the optimized real-time sensing layer and the situation assessment results output by the optimized trend sensing layer are obtained respectively. A preset fusion algorithm is used to weight and fuse the two output results. The fusion weights are assigned based on the confidence level of the two sensing output results to obtain the initial fault perception and location results and the corresponding initial resource demand data. In this embodiment, the initial fault perception result is specifically represented as follows: In the formula, The result of fault detection; This represents the average value of fault perception. For perceived standard deviation; Confidence coefficient; The comprehensive fault perception benchmark value obtained by integrating and optimizing the outputs of the real-time perception layer and the trend perception layer represents the average fault perception value of the location and scale of the power grid fault. For the first h The fusion weights of the road perception outputs are dynamically allocated based on the confidence levels of the two perception results. The higher the confidence level, the greater the weight, and all weights sum to 1. For the first h The original fault output of the layered sensing network. h =1 represents the instantaneous perception result. h =2 represents the situation assessment result. The number of sensing output channels participating in the fusion; For the first h The standard deviation of the road perception output; The calculation process for initial resource requirements data is specifically represented as follows: ; In the formula, Let be the resource demand of node i for resource r at time t; The multi-source sensing input vector for node i; The maximum capacity constraint for node i and resource r; LSTM To perform time series prediction on the input vector based on a Long Short-Term Memory (LSTM) temporal network, the output is an unconstrained original resource demand prediction value; LSTM is used to capture the time series evolution of faults to achieve advance prediction of resource demand. As a truncation function, a range truncation operation is performed to limit the output demand to a reasonable physical range, avoiding negative predicted values ​​or exceeding the node's capacity limit. The lower limit of 0 indicates that the resource demand cannot be negative, eliminating meaningless negative values ​​generated by the model prediction; the upper limit... The maximum capacity constraint of r-type resources that node i can carry / allocate is determined by the physical boundaries of the power grid, such as the substation equipment capacity, transportation upper limit, and reserve scale.

[0049] After obtaining the initial fault perception and location results and the corresponding initial resource requirement data, the initial fault perception results are verified and calibrated by introducing preset power system network topology constraints. Abnormal location results that do not conform to the power transmission topology logic are eliminated. At the same time, fault perception results with confidence intervals are generated by combining the degree of perception deviation, and the demand calculation results of various resources such as emergency communication, power supply and emergency repair are adjusted accordingly to obtain the real-time status of the power system and the corresponding resource requirement data. The real-time status of the power system is reflected by the fault perception results.

[0050] In the above scheme, a two-way information exchange channel is used to transmit and optimize the parameters of the two-layer sensing network. A multi-scale loss function is used to complete the optimization training of the sensing network. Then, the dual-output results are integrated and calibrated by power system network topology constraints to finally obtain the real-time status and resource demand data of the power grid. Thus, adaptive optimization of the sensing network is achieved through error gradient feedback and evolutionary feature embedding. The multi-scale loss function is used to adapt to the sensing training requirements of different time scales. Combined with topology constraint calibration, the fault location results are ensured to conform to the physical operation logic of the power grid. Finally, reliable real-time status of the power system and corresponding resource demands are output, providing accurate data input for subsequent dispatch decisions and helping to improve the reliability of power grid operation.

[0051] A preferred embodiment involves constructing a multi-agent reinforcement learning model based on the multi-source perception data and the resource demand data, comprising: constructing state vectors for each allocation node based on the multi-source perception data and the resource demand data; constructing an adjacency matrix based on a preset topological structure between the allocation nodes; constructing action vectors for each allocation node based on the multi-source perception data; constructing a global activation function value and several local activation function values ​​based on the multi-source perception data and the resource demand data, and obtaining a multi-objective activation function based on pre-constructed target weight coefficients, the global activation function value, and the several local activation function values; and constructing a multi-agent reinforcement learning model corresponding to each allocation node based on the state vectors, adjacency matrix, action vectors, and multi-objective activation functions corresponding to each allocation node.

[0052] For example, based on the collected multi-source sensing data and the output resource demand data, a corresponding state vector is constructed for each power dispatch node, represented as follows: ; In the formula, Let i be the state vector of node i at time t; This indicates the real-time running status. This represents the resource demand vector. This is a vector of available resources; This represents the fault level and location vector.

[0053] The state vector incorporates real-time operating parameters of node equipment, the scale of various emergency resource demands after fault location, the available emergency resource reserves of the node itself, and fault level and fault location information. This allows for a complete representation of the current status of a single dispatch node from multiple dimensions, including operating status, resource supply and demand, and fault attributes, providing comprehensive input for intelligent agent decision-making. The dispatch nodes cover different levels of power resource management nodes, such as regional dispatch centers, substations, and emergency repair sites, ensuring that the state representation is adapted to the functional positioning of different nodes.

[0054] Then, an adjacency matrix is ​​constructed based on the preset network topology between each allocation node. The topology is determined according to the connection status of the power transmission channel and the deployment of the communication link between the nodes. The element values ​​of the adjacency matrix correspond to the assistance priority between nodes. The connection priority between the core allocation node and the grassroots node is higher than the connection priority between the grassroots nodes. The adjacency matrix clearly defines the interaction range between each agent and the physical boundary of resource allocation.

[0055] Subsequently, action vectors for the agent are constructed for each allocation node, specifically as follows: ; In the formula, Let i be the action vector of node i at time t; The amount of resource r allocated from node i to node j; To allocate execution time.

[0056] The action vector includes two core decision-making elements: the quantity of emergency resources allocated and the timing of allocation. The quantity of allocation corresponds to the scale of allocation of different types of emergency resources, such as emergency communication resources, power supply resources, and repair resources, among nodes. The timing of allocation corresponds to the start and execution time of the resource allocation action. It is worth mentioning that the action space can be discretized during implementation, and action options can be divided according to the resource scale level and the time urgency. This reduces the computational complexity of the model while ensuring decision-making flexibility and adapts to the real-time requirements of emergency allocation in the power system. For example, emergency generators can be divided into levels of 0, 1, 2, and 3 units, and repair personnel can be divided into levels of 10, 20, and 30 people. The timing of allocation is divided into time periods such as immediately, within 30 minutes, within 1 hour, and within 2 hours according to the urgency of the fault handling.

[0057] After defining the states and actions, a multi-objective incentive system is constructed based on multi-source sensing data and resource demand data, including several local incentive function values ​​and global incentive function values. The local incentive function values ​​are based on the operational efficiency of a single dispatch node, corresponding to the degree of resource supply and demand balance, fault handling efficiency, and dispatch cost. The local incentive function values ​​can be represented by resource balance incentive function values, fault handling efficiency incentive function values, and dispatch cost incentive function values, which are used to guide the agent to optimize the resource allocation and handling effect of its own node. The global incentive function value is based on the overall operational efficiency of the power system, combined with the resource allocation differences of adjacent nodes, transmission distance and communication latency, hierarchical cooperation requirements, and fault propagation prevention and control requirements, and is used to guide the agent to take into account both the overall network resource balance and collaborative handling effect.

[0058] Initial target weight coefficients are assigned to each incentive term, and a complete multi-objective incentive function is obtained by weighted summation. This provides a clear optimization direction and evaluation criterion for model training. The multi-objective incentive function is expressed as follows: ; In the formula, Values ​​of multi-objective excitation functions; , , and For adaptive weighting coefficients, satisfying .

[0059] Finally, the state vectors corresponding to each allocation node, the adjacency matrix representing the topological association of the nodes, the action vectors defining the decision output, and the multi-objective activation function guiding the optimization direction are combined to construct a multi-agent reinforcement learning model for each allocation node, with each allocation node as an independent decision agent.

[0060] In the above scheme, a corresponding state vector, adjacency matrix, and action vector are constructed for each dispatch node. Then, multi-source sensing data and resource demand data are used to construct a global incentive function and several local incentive function values ​​to build a multi-objective incentive function. Finally, a multi-agent reinforcement learning model is constructed by combining the state vector, adjacency matrix, action vector, and multi-objective incentive function of each dispatch node. Thus, by adapting a single dispatch node as an independent agent to the distributed physical architecture of the power grid, and by using an adjacency matrix to limit the physical boundaries of resource dispatch to ensure the physical feasibility of the decision, the constructed multi-objective incentive function simultaneously considers the local dispatch efficiency of nodes and the global resource balance of the power grid. This allows the dispatch strategy generated by the multi-agent reinforcement learning model to simultaneously meet the core objectives of local fault handling and global resource optimization, ensuring the reliability of power grid operation.

[0061] A preferred embodiment involves constructing a global incentive function value and several local incentive function values ​​based on the multi-source sensing data and the resource demand data. A multi-objective incentive function is then obtained based on pre-constructed target weight coefficients, the global incentive function value, and the several local incentive function values. This includes: classifying fault levels based on the multi-source sensing data to determine the emergency resource allocation benchmark corresponding to each fault level; calculating a standardized resource supply-demand deviation based on the emergency resource allocation benchmark corresponding to each fault level and the resource demand data; calculating a resource balance incentive value based on the standardized resource supply-demand deviation and a preset time decay factor; calculating the immediate response efficiency, stage response efficiency, and overall response efficiency of each allocation node based on the multi-source sensing data; and calculating a fault response efficiency incentive value based on the immediate response efficiency, the stage response efficiency, and the overall response efficiency. The multi-source sensing data is used to calculate the consumption cost, and the allocation consumption incentive value is calculated based on the consumption cost. The resource allocation difference ratio between adjacent allocation nodes is calculated and a difference matrix is ​​constructed. The distance delay weight coefficient of each adjacent allocation node is calculated based on the multi-source sensing data, and the global assistance incentive value is calculated based on the difference matrix and the distance delay weight coefficient. The resource balance incentive value, fault handling efficiency incentive value, and allocation consumption incentive value are used as several local incentive function values, and the global assistance incentive value is used as the global incentive function value. Initial weight coefficients are set for several local incentive function values ​​and the global incentive function value, and weight constraints are set for the initial weight coefficients to obtain target weight coefficients. Based on the target weight coefficients, the global incentive function value and several local incentive function values ​​are weighted and summed to obtain a multi-objective incentive function.

[0062] For example, fault levels are classified based on historical fault handling records in multi-source sensing data, forming a multi-level fault classification system covering different severity levels. For each fault level, a corresponding emergency resource configuration benchmark is determined, clarifying the standard configuration scale of various emergency communication resources, power supply resources, and repair resources. In one interpretation of this embodiment, combining power industry standards and historical fault handling statistics, power system faults are classified into four levels: general faults, major faults, serious faults, and extremely serious faults. The fault impact range, power outage load scale, equipment damage degree, and communication interruption range are used as classification criteria. For each fault level, historical best handling cases are statistically analyzed, and the configuration benchmark quantities of three core resources—emergency power supply resources, repair personnel and materials, and emergency communication equipment—are calibrated. For example, fault level L∈{1, 2, 3, 4}, where 1=general fault, 2=major fault, 3=serious fault, 4=extremely serious fault; emergency resource type R={r1, r2, r3}, where r1 is emergency power supply resources, r2 is repair resources, and r3 is emergency communication resources; the standard configuration benchmark quantity of resource r corresponding to fault level L. The data is obtained from historical fault handling datasets. Combined with the emergency resource allocation benchmarks corresponding to each fault level and the resource demand data of each allocation node, the standardized resource supply-demand deviation of each node is calculated to quantify the degree of deviation between the actual resource status of the node and the benchmark configuration. Then, a preset time decay factor is introduced and combined with the standardized resource supply-demand deviation to calculate the resource balance incentive value, specifically as follows: ; In the formula, To standardize the supply and demand imbalance of resources, This is a time decay factor, which can be determined based on the duration of the fault. and attenuation coefficient The calculation yields the following result, which is expressed as: attenuation coefficient It can be 0.1; Standardized resource supply and demand imbalance The calculation process is as follows: ; In the formula, To allocate the available amount of resource r at node i; To allocate the actual demand of resource r for node i; Based on the full-process fault handling data collected from various dispatch nodes using multi-source sensing data, handling efficiency indicators are calculated from three time dimensions: immediate, stage, and overall. Immediate handling efficiency reflects the response progress in the initial stage of a fault, stage handling efficiency reflects the task progress in the middle stage of fault handling, and overall handling efficiency reflects the final handling effect throughout the entire fault cycle. Weighting coefficients are assigned to the handling efficiency at the three different time scales, and a weighted sum is obtained to obtain the fault handling efficiency incentive value. The calculation process is expressed as follows: ; In the formula, To optimize the real-time processing efficiency of node i; To adjust the stage processing efficiency of node i; To optimize the overall processing efficiency of node i; , and The fixed weighting coefficients corresponding to the three types of efficiency satisfy the following conditions: .

[0063] Then, based on multi-source sensing data, various consumption costs throughout the entire resource allocation process are statistically analyzed, specifically covering three core cost categories: emergency resource transportation consumption, on-site allocation operation consumption, and equipment wear and tear consumption. Corresponding weight coefficients are set for these three types of consumption to calculate the comprehensive allocation consumption level. Then, the allocation consumption incentive value is calculated with a preset allocation benchmark consumption as a reference. The calculation process is expressed as follows: ; In the formula, To allocate emergency resource transportation consumption at node i; Consumption of on-site allocation operations for allocation node i; To manage the equipment wear and tear at node i The total resource allocation baseline consumption for node i; , and The weighting coefficients for the three types of consumption satisfy the following conditions: .

[0064] Then, a global assistance incentive value is constructed. One possible implementation includes: first calculating the resource allocation difference ratio between each adjacent allocation node, and constructing the corresponding resource difference matrix to quantify the degree of resource imbalance between adjacent nodes. For adjacent nodes i and j, the resource difference matrix construction process is expressed as follows: ; Then, based on the physical distance of power transmission and the communication link delay data in the multi-source sensing data, the distance delay weight coefficient of each adjacent node is calculated. The closer the physical distance and the lower the communication delay, the higher the weight of the node. The calculation process of the distance delay weight coefficient is as follows: ; In the formula, The physical distance for power transmission between node i and node j; This refers to the communication link delay data between node i and node j; and This is the comprehensive weighting coefficient for distance and time delay; Simultaneously, a power system-level assistance coefficient is introduced. With fault propagation penalty factor The hierarchical assistance coefficient is determined based on the power grid hierarchy to which the node belongs. Higher-level core nodes correspond to higher cooperation weights. The fault propagation penalty factor decreases accordingly as the fault propagation range expands. Finally, by traversing all adjacent node pairs across the entire network, the distance-delay weight coefficient, resource difference matrix, hierarchical assistance coefficient, and fault propagation penalty factor are combined to calculate the global assistance incentive value, expressed as: In the formula, To obtain the average value of multiple resource types and thus the comprehensive resource differences between nodes, .

[0065] Finally, the resource balance incentive value, fault handling efficiency incentive value, and dispatch consumption incentive value are used as local incentive function values, corresponding to the operational optimization objective of a single dispatch node. The global assistance incentive value is used as the global incentive function value, corresponding to the overall collaborative optimization objective of the power system. Initial weight coefficients are set for the four types of incentive function values, along with weight boundary constraints to ensure that all weight coefficients are within a reasonable range and meet normalization requirements. This yields the target weight coefficients. Then, based on these target weight coefficients, a weighted sum is performed on the global incentive function value and several local incentive function values ​​to obtain the complete multi-objective incentive function for the power system. The multi-objective incentive function is expressed as follows: ; In the formula, Values ​​of multi-objective excitation functions; , , and For adaptive weighting coefficients, satisfying Where, let the initial weight coefficient be... The initial weight coefficients are given boundary constraints and the weights are iterated, as follows: ; In the formula, The values ​​are the fine-tuning amounts for the weight coefficients, n=1,2,3,4. Finally, the iteratively updated weight coefficients are... , , and As the target weight coefficient; As a truncation function, it performs a range truncation operation to limit the output demand to a reasonable physical range, avoiding negative predicted values ​​or exceeding the node's capacity limit.

[0066] In the aforementioned scheme, fault levels are classified and corresponding emergency resource allocation benchmarks are determined. Incentive function values ​​are constructed from multiple dimensions, including resource balance, response efficiency, allocation consumption, and global assistance. The weights of each incentive function value are dynamically adjusted based on the degree of performance improvement to form a multi-objective incentive function. Thus, by matching emergency resource allocation benchmarks to fault levels, the uniformity of response standards is ensured. Multi-dimensional incentive indicators are calculated to cover all performance requirements of resource allocation. Furthermore, a dynamic weight adjustment mechanism allows the objective incentive function to adapt to the optimization emphasis of different fault scenarios, guiding the agent to generate resource allocation strategies that better meet actual operational needs, thereby contributing to improved power grid reliability.

[0067] A preferred embodiment, based on the multi-agent reinforcement learning model, configures corresponding independent agents in each of the allocation nodes, and generates a target resource allocation strategy according to the real-time state of the power system and each of the independent agents through a preset proximal decision optimization algorithm, including: configuring corresponding independent agents for each allocation node based on the multi-agent reinforcement learning model corresponding to each allocation node, and constructing a local sharing area, a cooperative sharing area, and a global sharing area in each independent agent; training each of the independent agents through a preset proximal decision optimization algorithm to generate several state transition samples; calculating the importance score of the several state transition samples based on a preset importance scoring algorithm; and storing the several state transition samples in a target location based on the importance scores of the several state transition samples. The system identifies a local shared area, an assistance shared area, or a global shared area; calculates the sampling probability of several state transition samples, and selects several training samples from the local shared area, the assistance shared area, and the global shared area based on the sampling probability; updates the decision network parameters of each independent agent based on the several training samples to obtain the independent agents with updated decision networks; constructs a weight matrix based on the historical decision data of the independent agents with updated decision networks and the preset topology; optimizes the decision networks of the independent agents with updated decision networks based on the weight matrix and the preset near-end decision optimization algorithm to obtain several target independent agents; and generates a target resource allocation strategy based on the real-time state of the power system through the several target independent agents.

[0068] For example, based on the multi-agent reinforcement learning model corresponding to each dispatch node, independent agents are configured for each level of power dispatch node, such as regional dispatch centers, substations, and emergency repair sites. This maintains the relative independence of each agent in terms of state input, action output, and task execution. Simultaneously, a three-layer sample storage structure—local shared area, collaborative shared area, and global shared area—is built in the training architecture of each agent. The local shared area stores local state transition samples generated by the interaction between the independent agent and the environment; the collaborative shared area stores interaction samples generated by collaborative decision-making between neighboring agents; and the global shared area stores global state samples covering the overall network operation status. Then, a preset near-end decision optimization algorithm is used to iteratively train each independent agent. During training, each agent outputs corresponding resource dispatch actions based on the current node state. After interacting with the power system operating environment, the environment provides feedback with corresponding incentive values ​​and the node's operating state at the next moment. The current state, resource dispatch actions, incentive values, and the state at the next moment together form a complete state transition sample. As training iterates, a set of state transition samples covering different fault scenarios and dispatch modes is continuously accumulated.

[0069] Then, a preset importance scoring algorithm is called to calculate the importance score for each state transition sample. The scoring process comprehensively considers four core dimensions: the incentive value corresponding to the state transition sample, the frequency of occurrence of this type of state transition, the scope of influence covered by the state change, and the degree of difference in decision results. After setting corresponding weighting coefficients for each dimension, the final importance score result is obtained. The importance score calculation process is expressed as follows: ; In the formula, The importance score of the state transition sample; The excitation value corresponding to the state transition sample; The state transition frequency; The scope of the state's influence; For decision-making differences; , , and For the weighting coefficients, satisfying ; Then, based on the importance score hierarchy, state transition samples are stored in the corresponding shared areas. Samples with a wide impact and high decision-making reference value are stored in the global shared area, samples related to neighborhood collaboration are stored in the assistance shared area, and single-node routine decision samples are stored in the local shared area. After completing the hierarchical storage of samples, the sampling probability of all state transition samples in each shared area is calculated. The sampling probability is determined by combining the timeliness of the sample and the number of updates of the decision network, and is expressed as: ; In the formula, The sampling probability of the sample; For the timeliness of the sample; Update the decay coefficient for decision-making; The decision is updated several times. Finally, training samples of the corresponding size are extracted from the local shared area, the cooperative shared area, and the global shared area according to the sampling probability, forming training sample sets corresponding to a single agent, a group of neighboring agents, and the entire group of agents, respectively.

[0070] Subsequently, the decision network parameters of independent agents are updated hierarchically using training samples at each level. Samples extracted from the local shared region are used to update the decision network parameters of a single agent; samples extracted from the assisting shared region are used to update the decision network parameters of neighboring agent groups; and samples extracted from the global shared region are used to update the decision network parameters of all agents. This hierarchical parameter update mechanism achieves simultaneous optimization of local decision-making capabilities and global collaborative capabilities, resulting in independent agents with updated decision networks. Then, combining the historical decision data of each updated independent agent with the preset network topology of each dispatch node, the element values ​​of the weight matrix are calculated to construct the agent's weight matrix. The objective function of the near-end decision optimization algorithm is decomposed into three optimization branches at different time scales: immediate optimization, periodic optimization, and long-term optimization. The weight coefficients of each optimization term are calculated using the weight matrix and power system state information. This allows for multi-scale secondary optimization adjustments to the initially updated decision network, enabling it to simultaneously adapt to multiple dispatch requirements, including instantaneous fault response, phased handling advancement, and long-term operational optimization. The multi-scale optimization objective function of the near-end decision optimization algorithm can be expressed as: ; In the formula, , and These are respectively: immediate optimization items, periodic optimization items, and long-term optimization items; , The weights are the weighting coefficients for each optimization term in the weight matrix, satisfying... + + =1.

[0071] It is worth mentioning that during the training process, a basic feature parameter sharing mechanism can be used to train the decision network for agents with the same type of allocation nodes, and finally several target independent agents with training convergence are obtained. After the model training is completed, the real-time status data of the power system is input into each target independent agent, and each agent outputs the target resource allocation decision of the corresponding node based on the trained decision network.

[0072] In the above scheme, a three-layer shared area system (local, collaborative, and global) is constructed to store state transition samples. Samples are stored hierarchically based on importance scores. Training samples are selected from each shared area by calculating sampling probabilities to update the decision network parameters. A weight matrix is ​​then constructed based on the topology and historical decision data. The decision network is then optimized at multiple scales using a near-end decision optimization algorithm to obtain the target independent agent and generate the target resource allocation strategy. Thus, the hierarchical shared area system enables the classification and reuse of decision-making experience across different scopes. The importance scoring mechanism prioritizes the retention of high-value training samples, balancing individual decision-making, neighborhood collaboration, and global optimization during hierarchical parameter updates. Furthermore, the construction of the weight matrix and the use of the near-end decision optimization algorithm to optimize the updated decision network accelerate the convergence speed of the multi-agent reinforcement learning model within the independent agent, improve the multi-objective optimization effect and scenario adaptability of the allocation strategy, and enhance the reliability of power grid operation.

[0073] A preferred embodiment involves converting the target resource allocation strategy into several allocation tasks corresponding to independent intelligent agents based on a preset rule knowledge base, and executing the corresponding allocation tasks in each independent intelligent agent to allocate dynamic resources of the power system. This includes: parsing the target resource allocation strategy to obtain an initial allocation scheme; performing constraint verification on the initial allocation scheme based on the preset rule knowledge base and calculating the rule matching score of the initial allocation scheme; if the initial allocation scheme does not meet the preset constraint verification requirements or the rule matching score of the initial allocation scheme does not meet the preset matching requirements, then revising the initial allocation scheme until… The revised initial allocation scheme satisfies the preset constraint verification requirements, and the rule matching score of the revised initial allocation scheme satisfies the preset matching degree requirements, thus obtaining a revised allocation scheme. Based on the pre-acquired historical allocation execution data, the execution risk probability of the revised allocation scheme in different preset time windows is calculated, and an execution priority is generated according to the execution risk probability of the revised allocation scheme in different preset time windows. The revised allocation scheme is sorted according to the execution priority to obtain allocation tasks. The allocation tasks are distributed to each of the independent intelligent agents so that each of the independent intelligent agents executes the corresponding allocation task to dynamically allocate resources in the power system.

[0074] For example, the output target resource allocation strategy is first structured and parsed. Combining the fault location results and resource demand data of each node, the abstract allocation instruction output by the model is decomposed into an initial allocation scheme that includes allocation node pairs, multiple types of emergency resources, corresponding allocation quantities, and allocation time windows. For example, the target resource allocation strategy indicates "the regional allocation center allocates emergency resources to the faulty substation, to be completed within 2 hours." Combining the real-time status of the power system and resource demand data, the initial allocation scheme is parsed as "the allocation object is from the regional allocation center to the target substation; the allocation resources are 2 emergency generators." The system includes a vehicle, 20 repair personnel, and one communication relay device; the allocation time window is within 2 hours. In this embodiment, the preset rule knowledge base includes three core rule systems: power system physical constraint rules, industry business constraint rules, and fault handling priority rules. Physical constraint rules include hard boundary requirements such as power equipment carrying capacity, emergency resource transportation restrictions, and communication link transmission capacity. Industry business constraint rules include management requirements such as power industry fault handling standards and resource allocation approval processes. Fault handling priority rules include handling principles such as fault urgency ranking and node importance classification. Based on the preset rule knowledge base, the initial allocation plan undergoes multi-dimensional constraint verification, checking whether each plan conforms to physical operating boundaries, meets business process requirements, and aligns with fault handling priorities. Simultaneously, a rule matching score function is constructed to comprehensively calculate the rule matching score of the initial allocation plan from three dimensions: physical constraint satisfaction, fault handling target achievement, and emergency resource utilization efficiency. The calculation process is expressed as follows: ; In the formula, Score the rule matching degree; For physical constraint satisfaction; For the degree of achievement of business objectives; For resource utilization rate; , and Weighting coefficient for the rating.

[0075] If the initial allocation plan fails any constraint verification or the rule matching score fails to reach the preset matching threshold, an automatic plan correction process is initiated. The plan elements are adjusted based on the specific reasons for the verification failure. Possible correction methods include splitting a single centralized allocation into multiple batches with time-sharing allocations, fine-tuning the allocation time window to avoid resource conflicts, and adjusting resource configuration combinations to adapt to equipment capacity boundaries. For example, a single allocation within 2 hours can be split into two batches within 30 minutes and 1.5 hours to avoid resource transportation conflicts. While ensuring the core objective of fault handling, the plan must meet all rule requirements. That is, when the corrected initial allocation plan meets all constraint verifications or the rule matching score reaches the preset matching threshold, a compliant and feasible corrected allocation plan is obtained. After obtaining the corrected allocation plan, based on pre-acquired historical allocation execution data of the power system, key indicators such as allocation delay rate, resource conflict rate, and fault propagation rate from historical cases are extracted. The execution risk probability of the corrected allocation plan under different preset time windows is calculated. The calculation process is as follows: ; In the formula, To determine the probability of execution risk; To adjust the latency rate; Resource conflict rate; This represents the fault propagation rate. , and This is a risk weighting factor. The higher the risk of allocation delay, resource conflict, and failure propagation within the time window, the higher the probability of execution risk.

[0076] Then, based on the execution risk probability, the execution priority of each modified allocation plan is generated. The higher the execution risk probability of a modified allocation plan, the lower its execution priority is typically set. All modified allocation plans are then sorted according to the generated execution priorities, further refining the plans into standardized allocation tasks that clearly define the allocation targets, allocation quantities, allocation times, and execution entities. Each task clearly defines the resource allocation and transfer nodes, the specific configuration scale of various emergency resources, the precise execution time nodes, and the responsible execution entity, forming a task list that can be directly issued for execution. Subsequently, the allocation tasks corresponding to each allocation node are issued to the corresponding independent intelligent agents, i.e., the independent computing units of each allocation node, for execution by each allocation node. After receiving tasks, the node's local computing unit executes them sequentially according to priority. In this embodiment, during task execution, allocation tasks with low risk and high urgency are executed first. The low risk of an allocation task can be characterized by a low probability of execution risk for the modified allocation scheme corresponding to the allocation task. The urgency of an allocation task can be characterized by the urgency of the fault corresponding to the modified allocation scheme in the constraint verification step when it meets the fault handling priority rule. When the risks of allocation tasks are the same, allocation tasks with high urgency are executed first. When the urgency of allocation tasks is the same, allocation tasks with low risk are executed first. When the risks and urgency of allocation tasks are different, allocation tasks with low risk are executed first.

[0077] It is worth mentioning that during the allocation task execution, each node continuously monitors the local status perception results and resource allocation execution progress. When an abnormal perception status is detected, such as the expansion of the fault scope, the emergence of new characteristics of coupled faults, or resource allocation conflicts such as resource transportation obstruction or insufficient local resource supply, an emergency assistance request is automatically initiated to adjacent allocation nodes. The node receiving the assistance request generates response information including response time, allocation consumption, and resource inventory by combining its own historical cooperation frequency, current fault handling intensity, and available resource status. An emergency handling plan is reached through local negotiation between nodes. During the negotiation process, the negotiation weight is determined by combining node credibility and network hop count, and a consensus is gradually formed through weighted iteration. If negotiations stall, random disturbances are introduced to restart them, ensuring a feasible emergency response plan is quickly formed. Simultaneously, the heartbeat information and response latency of each participating dispatch node are continuously monitored. When a node's heartbeat is interrupted or its response latency exceeds a preset threshold, the corresponding node is deemed to be malfunctioning. Unfinished negotiation tasks from the malfunctioning node are extracted, and the task reception suitability of each node is calculated using existing methods based on the resource sufficiency, time response rate, and distance suitability of the remaining normal nodes. Under the premise of meeting resource constraints, unfinished tasks are reassigned to nodes with higher suitability. If necessary, tasks are split and shared by multiple nodes to ensure uninterrupted emergency response processes, thereby completing the dynamic resource allocation of the power system.

[0078] In the above scheme, the target resource allocation strategy is parsed into an initial allocation plan. Constraint verification and automatic correction are performed using a pre-set rule knowledge base. Execution risks are assessed based on historical execution data, and execution priorities are generated. These are then converted into allocation tasks and distributed to each intelligent agent for execution. Thus, by verifying power industry rules, the allocation plan ensures compliance with physical constraints and business management specifications. Risk prioritization improves the reliability of allocation task execution, thereby enhancing the reliability of power grid operation.

[0079] Based on the above method embodiments, corresponding apparatus embodiments are provided; see [link to apparatus embodiments]. Figure 2 , Figure 2 This is a schematic diagram of the module structure of a dynamic resource allocation system for a power system according to a certain embodiment of the present invention. Figure 2 As shown, this embodiment of the invention also provides a dynamic resource allocation system for a power system, including a network construction module 201, a feature extraction module 202, a model training module 203, a strategy generation module 204, and a resource allocation module 205; wherein: the network construction module 201 is used to construct a full-dimensional state perception network for the power system based on pre-collected multi-source sensing data from each allocation node; the feature extraction module 202 is used to extract fault status features from the multi-source sensing data based on the full-dimensional state perception network of the power system to obtain the real-time state of the power system and the corresponding resource demand data; the model training module 203 is used to... A multi-agent reinforcement learning model is constructed based on the multi-source sensing data and the resource demand data. The policy generation module 204 is used to configure corresponding independent agents in each of the allocation nodes based on the multi-agent reinforcement learning model, and generate a target resource allocation strategy according to the real-time status of the power system and each of the independent agents through a preset near-end decision optimization algorithm. The resource allocation module 205 is used to convert the target resource allocation strategy into several allocation tasks corresponding to each independent agent based on a preset rule knowledge base, and execute the corresponding allocation tasks in each of the independent agents to allocate dynamic resources of the power system.

[0080] This invention proposes a dynamic resource allocation system for power systems. It establishes a comprehensive power system state perception network to accurately perceive various types of fault characteristics and predict resource demand data. Then, it constructs a multi-agent distributed decision-making architecture, configuring corresponding independent agents in each allocation node. A pre-defined near-end decision optimization algorithm is used to generate target resource allocation strategies, transforming resource allocation decision-making from centralized to distributed, reducing computational complexity and shortening allocation response time. Finally, a rule transformation mechanism converts the target resource allocation strategy into allocation tasks for each independent agent, completing the dynamic resource allocation of the power system and ensuring the feasibility of the allocation tasks. Thus, the entire process from perception and decision-making to execution optimizes the ability to respond to emergencies. This enables each allocation node to flexibly generate resource allocation decisions and execute corresponding resource allocation tasks based on comprehensive perception results when facing sudden emergencies, improving the reliability of power grid operation.

[0081] It is understood that the above-described device embodiments correspond to the method embodiments of the present invention, and can implement the dynamic resource allocation method for a power system provided by any of the above-described method embodiments of the present invention.

[0082] It should be noted that the device embodiments described above are merely illustrative, and some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Furthermore, in the accompanying drawings of the device embodiments provided by this invention, the connection relationships between modules indicate that they have communication connections, which can specifically be implemented as one or more communication buses or signal lines. Those skilled in the art can understand and implement this without any creative effort.

[0083] Based on the above-described embodiment of a dynamic resource allocation method for a power system, another embodiment of the present invention provides a terminal device, which includes a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor. When the processor executes the computer program, it implements a dynamic resource allocation method for a power system according to any embodiment of the present invention.

[0084] For example, in this embodiment, the computer program can be divided into one or more modules, which are stored in the memory and executed by the processor to complete the present invention. The one or more modules may be a series of computer program instruction segments capable of performing a specific function, which describe the execution process of the computer program in the terminal device.

[0085] The terminal device may be a desktop computer, laptop, handheld computer, or cloud server, etc. The terminal device may include, but is not limited to, a processor and a memory.

[0086] The processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor. The processor is the control center of the terminal device, connecting all parts of the terminal device via various interfaces and lines.

[0087] Based on the above-described method embodiments, another embodiment of the present invention provides a computer-readable storage medium including a stored computer program, wherein, when the computer program is executed, it controls the device where the computer-readable storage medium is located to execute the dynamic resource allocation method for a power system as described in any of the above-described method embodiments of the present invention.

[0088] The modules / units integrated in the device / terminal equipment, if implemented as software functional units and sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the above embodiments of the present invention can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include: any entity or device capable of carrying the computer program code, a recording medium, a USB flash drive, a portable hard drive, a magnetic disk, an optical disk, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electrical carrier signal, a telecommunication signal, and a software distribution medium, etc.

[0089] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. A method for dynamic resource allocation in a power system, characterized in that, include: A full-dimensional state perception network for the power system is constructed based on multi-source sensing data from pre-collected dispatch nodes. Based on the power system's full-dimensional state perception network, fault status features are extracted from the multi-source perception data to obtain the real-time state of the power system and the corresponding resource demand data. A multi-agent reinforcement learning model is constructed based on the multi-source perception data and the resource demand data. Based on the multi-agent reinforcement learning model, corresponding independent agents are configured in each of the allocation nodes, and a target resource allocation strategy is generated by a preset near-end decision optimization algorithm according to the real-time status of the power system and each of the independent agents. Based on a preset rule knowledge base, the target resource allocation strategy is converted into several allocation tasks corresponding to each independent intelligent agent, and the corresponding allocation tasks are executed in each independent intelligent agent to allocate dynamic resources of the power system.

2. The method for dynamic resource allocation in a power system as described in claim 1, characterized in that, The power system's full-dimensional state perception network, constructed based on pre-collected multi-source sensing data from various dispatch nodes, includes: The multi-source sensing data of each dispatch node is preprocessed, and the preprocessed multi-source sensing data is divided into real-time monitoring sequences and historical correlation sequences. A real-time sensing layer is built based on the aforementioned real-time monitoring sequence; A trend perception layer is built based on the aforementioned historical correlation sequences; Based on a preset two-way information flow mechanism, a two-way information interaction channel is set up between the real-time perception layer and the trend perception layer to obtain a full-dimensional state perception network for the power system.

3. The method for dynamic resource allocation in a power system as described in claim 2, characterized in that, Based on the aforementioned power system full-dimensional state perception network, fault state features are extracted from the multi-source perception data to obtain the real-time state of the power system and corresponding resource demand data, including: Based on the power system's full-dimensional state perception network, the fault state features of the real-time monitoring sequence are extracted through the real-time perception layer to obtain instantaneous state features. These instantaneous state features are then transmitted to the trend perception layer through the two-way information interaction channel to update the parameters of the trend perception layer, thereby obtaining an optimized trend perception layer. Based on the power system's full-dimensional state perception network, fault status features are extracted from the historical correlation sequence through the trend perception layer to obtain fault evolution features, and the fault evolution features are embedded into the input of the real-time perception layer to obtain an optimized real-time perception layer. Based on the optimized trend perception layer and the optimized real-time perception layer, the power system full-dimensional state perception network is optimized and trained by a preset multi-scale loss function to obtain the optimized power system full-dimensional state perception network. The optimized real-time perception layer output and the optimized trend perception layer output are obtained through the optimized power system full-dimensional state perception network. The optimized real-time perception layer output and the optimized trend perception layer output are integrated according to a preset integration algorithm to obtain the initial fault perception result and the corresponding initial resource demand data. The initial fault perception result and the corresponding initial resource demand data are then calibrated according to preset power system network topology constraints to obtain the real-time status of the power system and the corresponding resource demand data.

4. The method for dynamic resource allocation in a power system as described in claim 1, characterized in that, A multi-agent reinforcement learning model is constructed based on the multi-source sensing data and the resource demand data, including: Based on the multi-source sensing data and the resource demand data, a state vector is constructed for each of the allocation nodes. Based on the preset topology between each of the allocated nodes, an adjacency matrix is ​​constructed; An action vector is constructed based on the multi-source sensing data corresponding to each of the allocation nodes; Based on the multi-source sensing data and the resource demand data, a global activation function value and several local activation function values ​​are constructed, and a multi-objective activation function is obtained based on the pre-constructed target weight coefficient, the global activation function value, and the several local activation function values. Based on the state vector, adjacency matrix, action vector, and multi-objective activation function of each of the aforementioned allocation nodes, a multi-agent reinforcement learning model is constructed for each of the aforementioned allocation nodes.

5. A method for dynamic resource allocation in a power system as described in claim 4, characterized in that, Based on the multi-source sensing data and the resource demand data, a global activation function value and several local activation function values ​​are constructed. Based on the pre-constructed target weight coefficients, the global activation function value, and the several local activation function values, a multi-objective activation function is obtained, including: The fault levels are divided based on the multi-source sensing data to determine the emergency resource allocation benchmark corresponding to each fault level. The standardized resource supply and demand deviation is calculated based on the emergency resource allocation benchmark corresponding to each fault level and the resource demand data. The resource balance incentive value is calculated based on the standardized resource supply and demand deviation and the preset time decay factor. Based on the multi-source sensing data, the immediate handling efficiency, stage handling efficiency, and overall handling efficiency of each dispatch node are calculated, and the fault handling efficiency incentive value is calculated based on the immediate handling efficiency, the stage handling efficiency, and the overall handling efficiency. The consumption cost is calculated based on the multi-source sensing data, and the consumption incentive value is calculated based on the consumption cost. Calculate the resource allocation difference ratio between adjacent allocation nodes and construct a difference matrix. Calculate the distance delay weight coefficient of each adjacent allocation node based on the multi-source sensing data. Calculate the global assistance incentive value based on the difference matrix and the distance delay weight coefficient. The resource balance incentive value, the fault handling efficiency incentive value, and the allocation consumption incentive value are used as several local incentive function values, and the global assistance incentive value is used as the global incentive function value. Initial weight coefficients are set for several local activation function values ​​and the global activation function values, and weight constraints are set for the initial weight coefficients to obtain target weight coefficients; Based on the target weight coefficients, the global activation function value and several local activation function values ​​are weighted and summed to obtain a multi-objective activation function.

6. A method for dynamic resource allocation in a power system as described in claim 4, characterized in that, Based on the multi-agent reinforcement learning model, corresponding independent agents are configured in each of the allocation nodes, and a target resource allocation strategy is generated according to the real-time state of the power system and each of the independent agents through a preset near-end decision optimization algorithm, including: Based on the multi-agent reinforcement learning model corresponding to each of the allocation nodes, a corresponding independent agent is configured for each of the allocation nodes, and a local sharing area, a cooperative sharing area and a global sharing area are constructed in each independent agent. Each independent agent is trained using a pre-defined proximal decision optimization algorithm to generate several state transition samples; The importance scores of several state transition samples are calculated based on a preset importance scoring algorithm. Based on the importance scores of several state transition samples, the several state transition samples are stored in the corresponding local shared area, assistance shared area or global shared area. Calculate the sampling probability of several state transition samples, and select several training samples from the local shared area, the assisted shared area and the global shared area according to the sampling probability; Based on several training samples, the decision network parameters of each independent agent are updated to obtain the independent agents with updated decision networks. A weight matrix is ​​constructed based on the historical decision data of the independent agents after the decision network is updated and the preset topology; The decision networks of independent agents after the decision networks of each decision network are updated are optimized according to the weight matrix and the preset proximal decision optimization algorithm to obtain several target independent agents; Based on the real-time status of the power system, a target resource allocation strategy is generated through several target independent intelligent agents.

7. A method for dynamic resource allocation in a power system as described in claim 1, characterized in that, Based on a preset rule knowledge base, the target resource allocation strategy is converted into several allocation tasks corresponding to each independent intelligent agent, and the corresponding allocation tasks are executed in each independent intelligent agent to allocate dynamic resources of the power system, including: The target resource allocation strategy is analyzed to obtain an initial allocation plan; The initial allocation scheme is constrained and verified based on the preset rule knowledge base, and the rule matching score of the initial allocation scheme is calculated. If the initial allocation scheme does not meet the preset constraint verification requirements or the rule matching score of the initial allocation scheme does not meet the preset matching requirements, the initial allocation scheme is modified until the modified initial allocation scheme meets the preset constraint verification requirements and the rule matching score of the modified initial allocation scheme meets the preset matching requirements, thus obtaining a modified allocation scheme. Calculate the execution risk probability of the modified allocation scheme in different preset time windows based on the pre-acquired historical allocation execution data, and generate execution priorities based on the execution risk probability of the modified allocation scheme in different preset time windows; The revised allocation schemes are sorted based on the execution priority to obtain allocation tasks; The allocation task is distributed to each of the independent intelligent agents so that each independent intelligent agent executes the corresponding allocation task to dynamically allocate resources in the power system.

8. A dynamic resource allocation system for a power system, characterized in that, It includes a network construction module, a feature extraction module, a model training module, a policy generation module, and a resource allocation module; among which: The network construction module is used to construct a full-dimensional state perception network for the power system based on the multi-source sensing data of each pre-collected dispatch node. The feature extraction module is used to extract fault status features from the multi-source sensing data based on the power system's full-dimensional state perception network, so as to obtain the real-time state of the power system and the corresponding resource demand data. The model training module is used to construct a multi-agent reinforcement learning model based on the multi-source perception data and the resource demand data. The strategy generation module is used to configure corresponding independent agents in each of the allocation nodes based on the multi-agent reinforcement learning model, and generate a target resource allocation strategy according to the real-time status of the power system and each of the independent agents through a preset near-end decision optimization algorithm. The resource allocation module is used to convert the target resource allocation strategy into several allocation tasks corresponding to each independent intelligent agent based on a preset rule knowledge base, and to execute the corresponding allocation tasks in each independent intelligent agent in order to allocate the dynamic resources of the power system.

9. A dynamic resource allocation terminal for a power system, characterized in that, The terminal includes a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein the processor, when executing the computer program, implements a dynamic resource allocation method for a power system as described in any one of claims 1 to 7.

10. A storage medium for a dynamic resource allocation method in a power system, characterized in that, The storage medium includes a stored computer program, wherein, when the computer program is executed, it controls the device where the storage medium is located to perform a dynamic resource allocation method for a power system as described in any one of claims 1 to 7.

Images