A power distribution network resource scheduling management method and system based on large model calculation

Abstract

The application discloses a power distribution network resource scheduling management method and system based on large model calculation, relates to the technical field of power grids, and comprises the following steps: a neural network base class is inherited through a class completion; the relationship in the power distribution network is processed; feature extraction and mapping are performed through three full connection layers; and a ReLU activation function is used to provide nonlinear expression capability; a hierarchical progressive processing mechanism is adopted to complete intelligent scheduling and decision of the power distribution network resources. The application significantly improves the operation reliability and safety of the power distribution network, greatly improves the intelligent level and management efficiency of the power distribution network resource scheduling, realizes dynamic topology self-adaptation and multi-objective optimization, reduces the scheduling response time by 50%, improves the economic benefit by 30%, improves the system reliability by 25%, improves the distributed energy consumption rate by 40%, has strong practical value and adaptability, can be popularized and applied, has low maintenance cost, has good scalability, has commercialization potential, and provides strong support for efficient operation of modern power systems.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of power grid, in particular to a power distribution network resource scheduling management method and system based on large model calculation. BACKGROUND

[0002] With the development of digital information technology, power grid enterprises continue to promote the construction, operation and production technology upgrading of intelligent measurement and control general terminal, take "quality improvement and efficiency increase" as the overall goal, carry out intelligent and integrated construction of intelligent measurement and control general terminal, give full play to the role of data production factors in the field of power transformation, and effectively support the upgrading of digital power grid, digital enterprise, digital service and digital industry. Among them, the intelligent measurement and control general terminal can include various power grid sensors.

[0003] At present, the existing power distribution network resource scheduling management mainly has the following problems: the traditional scheduling method is difficult to effectively deal with the high-dimensional and nonlinear power distribution network resource optimization problem; the existing deep learning method is not suitable for the dynamic topology change of power distribution network, and lacks the unified coordination and scheduling ability of new type of resources such as distributed energy and energy storage; the existing method is difficult to realize the balance of multi-objective trade-off and real-time response, and the explainability of scheduling decision is insufficient, which is difficult to obtain the trust of operation personnel. SUMMARY

[0004] In view of the problems existing in the prior art power distribution network resource scheduling management method and system based on large model calculation, the present application is proposed.

[0005] Therefore, the present application is aimed at the problem that the prior art is difficult to realize the balance of multi-objective trade-off and real-time response, and the explainability of scheduling decision is insufficient, which is difficult to obtain the trust of operation personnel, and the present application solves the problem by using a power distribution network resource scheduling management method and system based on large model calculation.

[0006] To solve the above technical problems, the present application provides the following technical scheme:

[0007] In the first aspect, the present application provides a power distribution network resource scheduling management method based on large model calculation, which includes: using an improved heterogeneous graph neural network structure, inheriting from a neural network base class through class completion, and processing the relationship in the power distribution network; using a language writing module to construct a three-layer intelligent agent collaborative decision-making mechanism, performing feature extraction and mapping through three fully connected layers, and using a ReLU activation function to provide nonlinear expression ability; for scheduling decision generation under normal operation state of the power distribution network, a hierarchical progressive processing mechanism is adopted to complete intelligent scheduling decision of the power distribution network resources.

[0008] As a preferred scheme of the power distribution network resource scheduling management method based on large model calculation of the application, wherein: the improved heterogeneous graph neural network structure includes the organic combination of node feature coding, edge feature coding and multi-head graph attention mechanism, the feature coding includes designing a three-layer deep encoder, preventing overfitting by setting a dropout rate of 0.1, and completing the unified representation of heterogeneous nodes by using different types of power distribution network device node features, the node feature encoder adopts a residual connection structure to retain the original feature information, the edge feature coding includes using a double-layer EdgeEncoder structure to process the line parameter features in the power distribution network, the encoder extracts key information such as line impedance, power flow and capacity, and converts it into a feature vector consistent with the node feature dimension, and the attention mechanism is also integrated in the edge feature encoder, so that the model can adaptively focus on different types of edge features.

[0009] As a preferred scheme of the power distribution network resource scheduling management method based on large model calculation of the application, wherein: the processing of the relationship in the power distribution network includes encoding the input node feature x to obtain an initial node embedding vector h, and encoding the edge feature edge_attr to obtain an edge feature vector edge_h, considering the local and global topological relationship of the node through a four-layer graph attention network, iteratively updating the feature representation of the node, and introducing the modulation effect of the edge feature when calculating the attention weight between nodes, considering the similarity of node features, and taking the feature of the connecting edge as a modulation factor;

[0010] The modulation factor includes a double detection strategy, which accurately compares the node features and edge features, judges the differences between the nodes in the new and old topological graphs during the node feature detection process, detects whether there are newly added or deleted nodes, determines whether the node features have changed through the relative error of the feature vectors of the nodes that exist together, and flexibly adjusts the detection sensitivity through a configurable threshold parameter.

[0011] As a preferred scheme of the power distribution network resource scheduling management method based on large model calculation of the application, wherein: the attention mechanism includes using a lightweight neural network structure, inputting local state information and decision information of the upper layer coordination agent, performing feature fusion through a 128-dimensional hidden layer, setting the action space dimension to one quarter of the global action space, and introducing a residual connection mechanism.

[0012] The introduced residual connection mechanism includes generating a scheduling strategy according to the overall state of the system, combining the regional coordination agent with the global strategy and the regional state, extracting the regional characteristics through the_get_region_state function, generating the regional coordination instruction, executing the agent to fuse the local state information and the coordination instruction, completing the control action, and establishing a bidirectional information flow mechanism.

[0013] As a preferred scheme of the power distribution network resource scheduling management method based on large model calculation, wherein: the three-layer agent collaborative decision mechanism includes setting a global scheduling agent at the top layer, responsible for formulating the overall scheduling strategy, configuring multiple regional coordination agents at the middle layer to complete the coordination and optimization of regional resources, and deploying local execution agents at the bottom layer to execute control instructions.

[0014] Feature extraction and mapping are performed through three fully connected layers, the hidden layer has a dimension of 512, the ReLU activation function is used to provide nonlinear expression capability, and the output layer generates action space probability distribution through the Softmax function to complete probability modeling of the global scheduling strategy.

[0015] The probability modeling includes using the PPO algorithm to design the architecture, introducing an adaptive clipping ratio and a multiple loss function optimization mechanism; in the algorithm initialization stage, a network model, an optimizer and a clipping ratio framework are constructed, the network model includes using a deep neural network structure to approximate the policy function and the value function, the optimizer selects the Adam optimizer, and the clipping ratio is set to 0.2 to limit the policy update step.

[0016] The PPO algorithm design architecture includes calculating the ratio of action probabilities under the new and old policies, introducing a clipping mechanism to constrain the policy ratio through exponential operation of the logarithmic difference; when the ratio exceeds the range [0.8, 1.2], clipping is performed.

[0017] As a preferred scheme of the power distribution network resource scheduling management method based on large model calculation, wherein: the scheduling decision generation for the normal operation state of the power distribution network includes processing by an emergency handling mechanism in an emergency handling mode, and when an abnormal working condition is detected, switching to an emergency control mode in time.

[0018] The emergency control mode includes processing through a hybrid decision mechanism that fuses deep learning and the system to handle faults, this mode accurately locates system abnormalities through an intelligent fault locating module, completes millisecond-level fault feature extraction based on an improved convolutional neural network, and completes rapid positioning of fault points of the power distribution network resource scheduling through deep analysis of multi-dimensional data of transient waveforms and phase characteristics.

[0019] As a preferred scheme of the power distribution network resource scheduling management method based on large model calculation of the present application, wherein: the hierarchical progressive processing mechanism includes adopting a distributed collaborative control strategy, the control system sorts the scheduling instructions according to time priority, quickly locates the fault point of the power distribution network resource scheduling, establishes a real-time feedback mechanism, and completes the intelligent scheduling decision of the power distribution network resource.

[0020] In a second aspect, the embodiments of the present application provide a power distribution network resource scheduling management system based on large model calculation, which comprises: a processing module adopting an improved heterogeneous graph neural network structure, which inherits from a neural network base class through class completion to process the relationships in the power distribution network.

[0021] A construction module adopts a language writing module to construct a three-layer intelligent agent collaborative decision-making mechanism, performs feature extraction and mapping through three fully connected layers, and uses a ReLU activation function to provide nonlinear expression ability.

[0022] A scheduling module generates scheduling decisions for the normal operation state of the power distribution network, adopts a hierarchical progressive processing mechanism, and completes the intelligent scheduling decision of the power distribution network resource.

[0023] In a third aspect, the embodiments of the present application provide a computer device comprising a memory and a processor, the memory storing a computer program, wherein the processor implements any step of the power distribution network resource scheduling management method based on large model calculation described above when executing the computer program.

[0024] In a fourth aspect, the embodiments of the present application provide a computer readable storage medium having a computer program stored thereon, wherein the computer program is executed by a processor to implement any step of the power distribution network resource scheduling management method based on large model calculation described above.

[0025] The present application has the following beneficial effects: the present application adopts an improved heterogeneous graph neural network structure, a three-layer intelligent agent collaborative decision-making mechanism and a hierarchical progressive processing mechanism to realize fine and intelligent management of power distribution network resources, not only improves the accuracy of resource representation through node feature encoding and edge feature encoding, but also enhances the dynamic optimization ability and decision efficiency of the model through a multi-head graph attention mechanism and a bidirectional information flow mechanism, and the emergency processing mode and fault point rapid positioning technology ensures the rapid response and recovery of the system under abnormal working conditions, significantly improves the operation reliability and safety of the power distribution network, and overall, the present application greatly improves the intelligent level and management efficiency of the power distribution network resource scheduling, and provides strong support for the efficient operation of modern power systems. BRIEF DESCRIPTION OF DRAWINGS

[0026] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings needed to be used in the embodiments will be briefly introduced as follows. Obviously, the drawings in the following description only represent some embodiments of the present application, and other drawings can be obtained by those skilled in the art without any creative effort.

[0027] Figure 1 A specific flowchart of a power distribution network resource scheduling management method and system based on large model calculation provided by an embodiment of the present application.

[0028] Figure 2 A knowledge migration mechanism schematic diagram of a power distribution network resource scheduling management method and system based on large model calculation provided by an embodiment of the present application.

[0029] Figure 3 An improved PPO algorithm schematic diagram of a power distribution network resource scheduling management method and system based on large model calculation provided by an embodiment of the present application. DETAILED DESCRIPTION

[0030] In order to make the above objectives, features and advantages of the present application more apparent and comprehensible, the specific embodiments of the present application will be described in detail below with reference to the drawings. Obviously, the described embodiments are only some of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without any creative effort should fall within the protection scope of the present application.

[0031] In the following description, many specific details are set forth in order to provide a thorough understanding of the present application. However, the present application can be practiced without the specific details, other than those described herein, and it is understood that the present application is not limited to the embodiments described herein and can be practiced with or without the other embodiments. Embodiments of the present application can be implemented in hardware, software, firmware, or a combination thereof. Embodiments of the present application can be implemented in one or more computer programs or program elements.

[0032] Secondly, the "one embodiment" or "embodiment" referred to herein means that the specific features, structures or characteristics can be included in at least one implementation of the present application. The "in one embodiment" appearing in different places in the specification does not mean the same embodiment, nor is it an independent or alternative embodiment that excludes other embodiments.

[0033] The present application is described in detail in conjunction with the schematic diagram, and in the detailed description of the embodiments of the present application, the sectional view of the device structure will be partially enlarged without the general proportion for the convenience of description, and the schematic diagram is only an example, which should not limit the protection scope of the present application herein. In addition, the three-dimensional spatial dimensions of length, width and depth should be included in the actual manufacture.

[0034] Meanwhile, in the description of the present application, it should be noted that the terms "upper, lower, inner and outer" and the like indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, and are only for the convenience of describing the present application and simplifying the description, and do not indicate or imply that the devices or elements referred to must have a particular orientation, be constructed and operated in a particular orientation, and therefore cannot be understood as a limitation on the present application. In addition, the terms "first, second or third" are only for descriptive purposes and cannot be understood as indicating or implying relative importance.

[0035] Unless otherwise explicitly specified and limited, the terms "mounting, connecting, connecting" in the present application should be understood broadly, for example: it can be fixed connection, detachable connection or integral connection; it can also be mechanical connection, electrical connection or direct connection, it can also be indirectly connected through intermediate medium, or it can be the communication inside two elements. For those skilled in the art, the specific meaning of the above terms in the present application can be understood according to the specific circumstances.

[0036] Embodiment 1

[0037] Reference Figures 1-3 For the first embodiment of the present application, the embodiment provides a power distribution network resource scheduling management method based on large model calculation, comprising:

[0038] S1: An improved heterogeneous graph neural network structure is adopted, which includes the organic combination of node feature encoding, edge feature encoding and multi-head graph attention mechanism. The feature encoding includes designing a three-layer deep encoder, and a dropout rate of 0.1 is set to prevent overfitting. The node feature of different types of power distribution network equipment is encoded to complete the unified representation of heterogeneous nodes. The node feature encoder adopts a residual connection structure to retain the original feature information. The edge feature encoding includes adopting a double-layer EdgeEncoder structure to process the line parameter features in the power distribution network. The encoder extracts the key information of line impedance, power flow and capacity, and converts it into a feature vector consistent with the node feature dimension. The attention mechanism is also integrated in the edge feature encoder, so that the model can adaptively pay attention to different types of edge features.

[0039] In terms of node feature encoding, the present application designs a three-layer deep NodeEncoder encoder, which sets a dropout rate of 0.1 to prevent overfitting. This encoder can map the operating parameters of different types of power distribution network equipment, such as transformers, circuit breakers, busbars, etc., to a unified hidden_dim-dimensional feature space, achieving unified representation of heterogeneous nodes. The node feature encoder adopts a residual connection structure to ensure the stability of deep network training while retaining the original feature information.

[0040] The edge feature code adopts a double-layer EdgeEncoder structure, which is specially used for processing line parameter features in the power distribution network. The encoder can effectively extract key information such as line impedance, power flow and capacity, and convert it into a representation vector consistent with the dimension of the node feature. The attention mechanism is integrated into the edge feature encoder, so that the model can adaptively focus on different types of edge features, improving the relevance of feature extraction.

[0041] At the level of graph attention propagation, the application innovatively designs a four-layer stacked multi-head graph attention network. Each MultiHeadGATLayer contains 8 attention heads, which improves the generalization ability of the model through batch normalization and dropout mechanism. The multi-head attention mechanism enables the model to learn the relationship between nodes from multiple subspaces simultaneously, significantly improving the comprehensiveness and robustness of feature extraction.

[0042] During the forward propagation process, the model first encodes the input node feature x to obtain the initial node embedding vector h, then encodes the edge feature edge_attr to obtain the edge representation vector edge_h, and finally updates the feature representation of the node through the four-layer graph attention network by considering the local and global topological relationships of the node. This hierarchical feature extraction method ensures that the model can fully capture the multi-scale topological information in the power distribution network.

[0043] An important innovation of this module is the introduction of edge feature modulation in attention calculation. Specifically, when calculating the attention weight between nodes, not only the similarity of node features is considered, but also the feature of the connecting edge is taken as a modulation factor, making the attention distribution more consistent with the actual physical meaning. This design significantly improves the model's ability to perceive the power flow characteristics of the power distribution network.

[0044] Finally, the improved heterogeneous graph neural network can output node representation vectors with rich semantic information, providing a reliable feature basis for subsequent scheduling decisions. Experiments show that this module improves the feature extraction effect by 35% and the dynamic scene adaptability by 40% compared to traditional methods, demonstrating significant technical advantages.

[0045] S1.1: Processing the relationship in the power distribution network includes encoding the input node feature x to obtain the initial node embedding vector h, encoding the edge feature edge_attr to obtain the edge representation vector edge_h, and updating the feature representation of the node through the four-layer graph attention network by considering the local and global topological relationships of the node. The introduction of edge feature modulation in attention calculation considers the similarity of node features and takes the feature of the connecting edge as a modulation factor, making the attention distribution more consistent with the actual physical meaning. The algorithm of the dynamic topological reconstruction mechanism of the attention mechanism is:

[0046]

[0047]

[0048] Further, the knowledge transfer mechanism of the present application has significant advantages in practical application. Through the mechanism, the inference speed of the lightweight model trained is more than 3 times higher than that of the original teacher model, and the accuracy is only reduced by less than 5 percentage points. In addition, the mechanism has good generalization and can adapt to power distribution systems of different scales and topologies, providing a flexible solution for actual deployment.

[0049] In terms of model updating, the knowledge transfer mechanism of the present application supports incremental learning. When new operation modes or fault types appear in the system, targeted training can be performed on the teacher model first, and then the new knowledge is transferred to the deployed lightweight model through knowledge distillation, so as to realize dynamic improvement of model capability and avoid the computational overhead caused by complete retraining.

[0050] The specific algorithm of the attention mechanism knowledge transfer mechanism is:

[0051]

[0052] S2: A three-layer intelligent agent collaborative decision-making mechanism is constructed using a language writing module. Feature extraction and mapping are performed through three fully connected layers, and a ReLU activation function is used to provide nonlinear expression capability.

[0053] Among them, the three-layer intelligent agent collaborative decision-making mechanism includes setting a global scheduling intelligent agent at the top layer, responsible for formulating the overall scheduling strategy, configuring multiple regional coordination intelligent agents at the middle layer, completing the coordination and optimization of regional resources, and deploying local execution intelligent agents at the bottom layer to execute control instructions.

[0054] Feature extraction and mapping are performed through three fully connected layers, the hidden layer dimension is 512, a ReLU activation function is used to provide nonlinear expression capability, and the output layer generates an action space probability distribution through a Softmax function to complete the probability modeling of the global scheduling strategy.

[0055] The improved PPO algorithm is used for processing, and the specific algorithm is:

[0056]

[0057]

[0058] Further, the multi-agent network structure adopts a hierarchical architecture design, and a high-efficiency three-layer agent collaborative decision mechanism is constructed. A global scheduling agent is arranged at the top layer, responsible for formulating the overall scheduling strategy; a plurality of regional coordination agents are configured at the middle layer, realizing the coordination and optimization of resources in the region; and a local execution agent is arranged at the bottom layer, executing specific control instructions.

[0059] The global scheduling agent adopts a deep neural network structure, the input layer dimension is the product of the state space dimension and the number of agents, feature extraction and mapping are performed through three fully connected layers, the hidden layer dimension is 512, and the ReLU activation function is used to provide nonlinear expression capability. The output layer generates an action space probability distribution through the Softmax function, realizing probability modeling of the global scheduling strategy.

[0060] The regional coordination agent designs an adaptive network structure according to the characteristics of different regions. The input dimension of each regional coordination agent is the same as the state dimension of a single agent, and feature transformation is performed through two hidden layers, and the hidden layer dimension is 256. Considering the fine requirement of regional scheduling, the action space dimension is set to half of the global action space, realizing accurate control in the region. At the same time, the attention mechanism is introduced, so that the coordination agent can adaptively focus on important state information in the region.

[0061] The local execution agent, as the most basic control unit, adopts a lightweight neural network structure. The input includes local state information and decision information of the upper coordination agent, and feature fusion is performed through a 128-dimensional hidden layer. Considering the specificity requirement at the execution level, the action space dimension is set to one quarter of the global action space, ensuring accurate execution of the control instruction. At the same time, the residual connection mechanism is introduced, effectively preventing the gradient vanishing problem.

[0062] In the action selection process, the invention designs a top-down decision transmission mechanism. First, the global scheduling agent generates a macro scheduling strategy according to the overall system state. Second, the regional coordination agent combines the global strategy and regional state, extracts regional features through the_get_region_state function, and generates regional coordination instructions. Finally, the local execution agent fuses local state information and coordination instructions to generate specific control actions. This hierarchical decision mechanism not only ensures the realization of the global optimization goal, but also takes into account the flexibility of local control.

[0063] In order to improve the reliability of decision-making, the invention establishes a bidirectional information flow mechanism between agents. The downward decision flow ensures the effective transmission of the strategy, and the upward state feedback provides the basis for the upper-level decision-making. Through this bidirectional information flow, the system can quickly adapt to environmental changes and realize dynamic optimization and adjustment. At the same time, the asynchronous update mechanism is adopted by each layer of agent, avoiding the decision bottleneck and improving the system response speed.

[0064] In practical applications, the multi-agent network structure of the present application shows good scalability and robustness. By adjusting the number of agents and network parameters, it can be flexibly adapted to different sizes of power distribution network systems. At the same time, the multi-level decision mechanism and redundant design improve the fault tolerance of the system, ensuring that the basic scheduling function can be maintained in the case of partial agent failure.

[0065] For the generation of scheduling decisions under normal operation state of the power distribution network, a hierarchical progressive processing mechanism is adopted to complete the intelligent scheduling decision of the power distribution network resources.

[0066] Among them, the generation of scheduling decisions under normal operation state of the power distribution network includes the algorithm processing of emergency handling mode, and the algorithm includes:

[0067]

[0068] Preferably, the hierarchical progressive processing mechanism includes adopting a distributed collaborative control strategy, the control system sorts the scheduling instructions according to time priority, and realizes accurate control of various devices through an improved feedforward-feedback control algorithm, and at the same time establishes a real-time feedback mechanism.

[0069] In one preferred embodiment, a power distribution network resource scheduling management system based on large model calculation, the system includes a processing module, which adopts an improved heterogeneous graph neural network structure, and through the class completion inherits from the neural network base class, processes the relationship in the power distribution network; a construction module, which adopts a language writing module to construct a three-layer intelligent agent collaborative decision mechanism, extracts and maps features through three fully connected layers, and uses a ReLU activation function to provide nonlinear expression ability; a scheduling module, which generates scheduling decisions under normal operation state of the power distribution network, adopts a hierarchical progressive processing mechanism, and completes the intelligent scheduling decision of the power distribution network resources.

[0070] The present embodiment also provides a computer device suitable for the case of the multi-source power grid information fusion method based on the Internet of Things, comprising a memory and a processor; the memory is used to store computer executable instructions, and the processor is used to execute the computer executable instructions to realize the multi-source power grid information fusion method based on the Internet of Things as proposed in the above-mentioned embodiments.

[0071] The computer device can be a terminal, and the computer device includes a processor, a memory, a communication interface, a display screen and an input device connected through a system bus. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The communication interface of the computer device is used for wired or wireless communication with external terminals. The wireless mode can be achieved through WIFI, operator network, NFC (near field communication) or other technologies. The display screen of the computer device can be a liquid crystal display screen or an electronic ink display screen. The input device of the computer device can be a touch layer overlaid on the display screen, or a key, trackball or touchpad arranged on the shell of the computer device. It can also be an external keyboard, touchpad or mouse, etc.

[0072] To sum up, by adopting the improved heterogeneous graph neural network structure, the three-layer agent collaborative decision mechanism and the hierarchical progressive processing mechanism, the fine and intelligent management of the power distribution network resources is realized. Not only the accuracy of resource representation is improved through node feature encoding and edge feature encoding, but also the dynamic optimization capability and decision efficiency of the model are enhanced through the multi-head graph attention mechanism and the bidirectional information flow mechanism. The emergency processing mode and the fault point rapid positioning technology ensure the rapid response and recovery of the system under abnormal working conditions, significantly improve the operation reliability and safety of the power distribution network. Overall, the intelligent level and management efficiency of the power distribution network resource scheduling are greatly improved, which provides strong support for the efficient operation of the modern power system.

[0073] Embodiment 2

[0074] Reference Figures 1-3 For the second embodiment of the present application, the embodiment provides a power distribution network resource scheduling management method based on large model calculation. In order to verify the beneficial effects of the present application, a simulation experiment is carried out for scientific demonstration.

[0075] In the experimental process of the present application, a power distribution network resource scheduling management system based on an improved heterogeneous graph neural network structure is first constructed. The system is written in Python language and inherits from the neural network base class, which is used to process the complex relationships in the power distribution network. The experimental object is a medium-sized power distribution network containing 100 nodes and 150 edges. The node types include substations, distribution lines, user loads, etc., and the edge features include line impedance, power flow and capacity, etc. Key information.

[0076] The first step of the experiment is to encode the node and edge features. The node feature encoder uses a three-layer deep encoder structure with a dropout rate of 0.1 to prevent overfitting and preserves the original feature information through a residual connection structure. The edge feature encoder uses a double-layer EdgeEncoder structure to extract line parameter features and convert them into a representation vector consistent with the dimension of the node features.

[0077] The second step is to consider the local and global topological relationships of nodes through a four-layer graph attention network (GAT) to iteratively update the feature representation of the nodes. When calculating the attention weight between nodes, the modulation effect of edge features is introduced to make the attention allocation more consistent with the actual physical meaning. In the experiment, the initial embedding vector dimension of each node is 64, and the edge representation vector dimension is also 64. After four iterations of GAT, the final node embedding vector dimension remains 64.

[0078] The third step is to build a three-layer intelligent agent collaborative decision-making mechanism. The top-level global scheduling agent is responsible for formulating the overall scheduling strategy, the middle-level regional coordination agent completes the coordination and optimization of regional resources, and the bottom-level local execution agent executes the control instructions. The agent extracts and maps features through three fully connected layers, with a hidden layer dimension of 512 and a ReLU activation function providing non-linear expression capability. The output layer generates an action space probability distribution through a Softmax function. In the experiment, the global scheduling agent generates a scheduling strategy every 10 seconds, the regional coordination agent updates the optimization scheme every 5 seconds, and the local execution agent executes a control instruction every 1 second.

[0079] The fourth step is to generate scheduling decisions under normal operation of the power distribution network. The experiment uses a hierarchical progressive processing mechanism, and through a distributed collaborative control strategy, the control system sorts the scheduling instructions according to time priority, implements precise control of various devices through an improved feedforward-feedback control algorithm, and establishes a real-time feedback mechanism. The experimental results show that the system can efficiently schedule resources under normal operation, with an average scheduling response time of less than 2 seconds and a scheduling accuracy of more than 95%. The experimental data of the invention are shown in Table 1 below:

[0080] Table 1 Experimental data table of the invention

[0081] Data type Numerical value Number of nodes 100 Number of edges 150 Node embedding vector dimension 64 Edge representation vector dimension 64 Dropout rate 0.1 Number of graph attention network layers 4 Hidden layer dimension 512 Global dispatching agent dispatching frequency Every 10 seconds Regional coordination agent update frequency Every 5 seconds Local execution agent execution frequency Every 1 second Average dispatching response time <2 seconds Dispatching accuracy >95％ Emergency handling mode start-up time Within 1 second Preliminary emergency dispatching completion time Within 3 seconds Power grid recovery stability time Within 5 seconds

[0082] Table 1 clearly shows the key data points in the experiment, which helps to understand the experimental process and results. The comparison between the invention and the prior art is shown in Table 2 below:

[0083] Table 2 Comparison table of the invention and the prior art

[0084]

[0085] Table 2 clearly shows the advantages of the technical solution of the present application in multiple aspects, highlighting its significant improvement in node and edge feature encoding, attention mechanism, agent collaborative decision-making, scheduling response time, scheduling accuracy, emergency handling capability, model complexity, calculation accuracy, economy and scalability, etc.

[0086] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present application and not to limit it. Although the present application has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present application can be modified or equivalently replaced without departing from the spirit and scope of the technical solutions of the present application, which should be covered in the scope of the claims of the present application.

Claims

1. A distribution network resource scheduling and management method based on large-scale model calculation, characterized by: include, Adopting an improved heterogeneous graph neural network structure, inheriting from the neural network base class through classes, to handle the relationship in the distribution network; The improved heterogeneous graph neural network structure includes an organic combination of node feature encoding, edge feature encoding, and a multi-head graph attention mechanism. The node feature encoding includes designing a three-layer deep encoder, setting a dropout rate of 0.1 to prevent overfitting, and utilizing the node features of different types of distribution network equipment to complete the unified representation of heterogeneous nodes. The node feature encoder adopts a residual connection structure to retain the original feature information. The edge feature encoding includes a two-layer EdgeEncoder structure to process the line parameter features in the distribution network. The edge feature encoder extracts key information such as line impedance, power flow, and capacity, and converts it into a representation vector consistent with the node feature dimension. The edge feature encoder also integrates an attention mechanism so that the model can adaptively focus on different types of edge features. A three-layer intelligent collaborative decision-making mechanism was constructed using a language writing module. Feature extraction and mapping were performed through three fully connected layers, and the ReLU activation function was used to provide nonlinear expression capabilities. The three-layer intelligent agent collaborative decision-making mechanism is constructed, including setting up a global scheduling intelligent agent at the top level to be responsible for formulating the overall scheduling strategy, configuring multiple regional coordination intelligent agents in the middle level to complete the coordinated optimization of resources in the region, and deploying local execution intelligent agents at the bottom level to execute control instructions; Feature extraction and mapping are performed through three fully connected layers. The hidden layer dimension is 512, and the ReLU activation function is used to provide nonlinear expression capabilities. The output layer uses the Softmax function to generate the action space probability distribution, completing the probabilistic modeling of the global scheduling strategy. The probabilistic modeling includes designing an architecture using the PPO algorithm, introducing an adaptive clipping ratio and a multiple loss function optimization mechanism; In the algorithm initialization phase, a network model, optimizer, and clipping ratio framework were constructed. The network model includes a deep neural network structure for approximating the policy function and value function. The Adam optimizer is used as the optimizer, and the clipping ratio is set to 0.2 to limit the policy update step size. The PPO algorithm design architecture includes calculating the ratio of action probabilities under the new and old strategies, and introducing a clipping mechanism to constrain the strategy ratio by taking the exponential operation of the logarithmic difference; when the ratio exceeds the range of [0.8, 1.2], clipping is performed; For the dispatch decision generation under the normal operation of the distribution network, a hierarchical and progressive processing mechanism is adopted to complete the intelligent dispatch decision of distribution network resources.

2. The distribution network resource dispatching and management method based on large model calculation according to claim 1, characterized in that: The processing of relationships in the distribution network includes encoding the input node feature x to obtain the initial node embedding vector h, encoding the edge feature edge_attr to obtain the edge representation vector edge_h, and iteratively updating the node feature representation by comprehensively considering the local and global topological relationships of the nodes through a four-layer graph attention network. The modulation effect of edge features is introduced. When calculating the attention weight between nodes, the similarity of node features is considered and the features of the connecting edges are used as modulation factors. The modulation factor includes the system adopting a dual detection strategy to accurately compare node features and edge features. During the node feature detection process, the differences between nodes in the new and old topology maps are judged, and whether there are new or deleted nodes are detected. For co-existing nodes, the relative error of the feature vector is calculated to determine whether the node features have changed, and the detection sensitivity is flexibly adjusted through configurable threshold parameters.

3. The distribution network resource scheduling and management method based on large model calculation according to claim 2 is characterized in that: The local execution agent includes a lightweight neural network structure, which inputs local state information and decision information of the regional coordination agent, performs feature fusion through a 128-dimensional hidden layer, sets the action space dimension to one-quarter of the global action space, and introduces a residual connection mechanism; The global scheduling agent generates a scheduling strategy based on the overall state of the system, combines the regional coordination agent with the global strategy and regional state, extracts regional features through the _get_region_state function, generates regional coordination instructions, and the local execution agent integrates local state information and coordination instructions to complete the control action. A two-way information flow mechanism is established between the agents. The two-way information flow mechanism includes a downward decision flow to ensure the effective transmission of the strategy, and the upward state feedback provides a basis for upper-level decision-making. Through the two-way information flow, the system completes dynamic optimization and adjustment.

4. The distribution network resource scheduling and management method based on large model calculation according to claim 3 is characterized in that: The generation of dispatching decisions for the normal operation of the distribution network includes adopting an emergency processing mechanism of an emergency processing mode for processing, and when an abnormal operating condition is detected, timely switching to an emergency control mode; The emergency control mode includes fault handling through a hybrid decision-making mechanism that integrates deep learning and the system. This mode accurately locates system anomalies through an intelligent fault location module, extracts millisecond-level fault features based on an improved convolutional neural network, and quickly locates fault points for distribution network resource scheduling through in-depth analysis of transient waveforms and multi-dimensional phase feature data.

5. The distribution network resource dispatching and management method based on large model calculation according to claim 4 is characterized in that: The hierarchical and progressive processing mechanism includes adopting a distributed collaborative control strategy, the control system sorts the scheduling instructions according to time priority, quickly locates the fault point of distribution network resource scheduling, establishes a real-time feedback mechanism, and completes the intelligent scheduling decision of distribution network resources.

6. A distribution network resource dispatching and management system based on large model calculation, used to implement the distribution network resource dispatching and management method based on large model calculation according to any one of claims 1 to 5, characterized in that: include, The processing module uses an improved heterogeneous graph neural network structure and inherits from the neural network base class through classes to process the relationships in the distribution network; The building module uses a language-based writing module to build a three-layer agent collaborative decision-making mechanism, performs feature extraction and mapping through three fully connected layers, and uses the ReLU activation function to provide nonlinear expression capabilities; The dispatching module generates dispatching decisions under the normal operation of the distribution network and adopts a hierarchical and progressive processing mechanism to complete the intelligent dispatching decisions of distribution network resources.

7. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the distribution network resource scheduling management method based on large model calculation according to any one of claims 1 to 5 are implemented.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the distribution network resource scheduling management method based on large model calculation according to any one of claims 1 to 5 are implemented.

Images