Electric power analysis method and system based on artificial intelligence

Abstract

The invention discloses an electric power analysis method and system based on artificial intelligence, and relates to the technical field of electric power system intelligent analysis. Establishing a hierarchical depth map neural network model, and setting a cross-layer information interaction channel to connect each sub-network; executing a bidirectional knowledge distillation algorithm, extracting a rule set from neural network output, and forming a parameter-rule bidirectional mapping matrix; deploying a federal reinforcement learning architecture, performing parameter aggregation on the local models on the distributed power nodes, and generating a power system state evaluation result and a risk coefficient matrix; and starting the self-repairing intelligent agent network, calculating a regulation and control parameter set of the multilevel power system, and outputting a regulation and control instruction. According to the invention, through the hierarchical depth map neural network and a cross-layer information interaction mechanism, collaborative optimization among all levels of the power system is realized, and the operation efficiency and stability of the system are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent analysis of power systems, and in particular to an artificial intelligence-based power analysis method and system. Background Art

[0002] As power systems expand in size and complexity, traditional power analysis methods face challenges such as insufficient data processing capabilities, difficulty in cross-level coordination, and slow system response. Existing technologies typically treat different levels (generation, transmission, transformation, and distribution) as independent systems, lacking effective overall collaborative optimization mechanisms and relying on manual intervention for fault response, making them unable to meet the operational demands of modern smart grids.

[0003] Furthermore, existing AI technologies suffer from the "black box" problem in power system applications, making it difficult to effectively integrate expert domain knowledge. Models lack interpretability and credibility, limiting their adoption in critical power system applications. Protecting power data privacy and enabling distributed node collaborative learning are also pressing challenges. Summary of the Invention

[0004] The present invention provides an artificial intelligence-based power analysis method and system to solve technical problems such as the difficulty of hierarchical collaborative optimization, insufficient model interpretability, and automatic response and safety verification of system failures in traditional power analysis methods when dealing with complex power systems.

[0005] In view of this, the first aspect of the present invention provides an artificial intelligence-based power analysis method, including: collecting multi-source heterogeneous data from the power system, executing an adaptive data fusion algorithm through edge computing equipment, and generating a power feature data set; establishing a hierarchical deep graph neural network model, the model includes four sub-network structures: power generation layer, transmission layer, transformation layer and distribution layer, and setting up cross-layer information interaction channels to connect each sub-network; executing a bidirectional knowledge distillation algorithm to embed expert domain knowledge parameters into the neural network model parameter space, and at the same time extracting a rule set from the neural network output to form a parameter-rule bidirectional mapping matrix; deploying a federated reinforcement learning architecture to perform parameter aggregation on the local model on the distributed power node to generate a power system status assessment result and a risk coefficient matrix; starting a self-repairing intelligent agent network, calculating a multi-level power system control parameter set based on the risk coefficient matrix, and outputting control instructions.

[0006] Optionally, executing an adaptive data fusion algorithm through edge computing devices includes: deploying a multi-level sensor network to collect multi-source heterogeneous data of the power system, the multi-level sensor network includes sensor units on the power generation side, transmission network, distribution network and user side; constructing a multi-time scale parallel data processing architecture to classify and process multi-source heterogeneous data of the power system according to time characteristics; mapping the physical structure of the power grid into a dynamic graph structure, calculating the importance of nodes, and applying differentiated feature extraction strategies to different types of nodes; performing alignment and fusion of multi-source heterogeneous data of the power system to generate a power feature data set with a spatiotemporal hierarchical structure.

[0007] Optionally, establishing a hierarchical deep graph neural network model includes: constructing sub-networks of each layer of the hierarchical deep graph neural network model; designing heterogeneous graph convolution operators, and adopting differentiated convolution kernel functions for different types of nodes and edge relationships; constructing multi-level cross-layer information interaction channels, including state upload channels, constraint downlink channels, and collaborative optimization channels; introducing a physical constraint embedding module, and embedding the physical laws and operating constraints of the power system as prior knowledge into the graph neural network.

[0008] Optionally, executing the bidirectional knowledge distillation algorithm includes: constructing an expert knowledge base of the power system and encoding the expert domain knowledge in a parameterized form into a structured knowledge graph; designing a knowledge-parameter mapping mechanism to map the key parameters in the structured knowledge graph to the corresponding modules of the hierarchical deep graph neural network model; executing a progressive knowledge distillation process to enable the hierarchical deep graph neural network model to learn different levels of the structured knowledge graph in stages; extracting a rule set from the hierarchical deep graph neural network model; constructing a parameter-rule bidirectional mapping matrix to record the correspondence and influence intensity between the parameters of the hierarchical deep graph neural network model and the rule set.

[0009] Optionally, deploying a federated reinforcement learning architecture includes: deploying edge agents at key nodes of the power system, each agent maintaining a local environment state space and action space model; constructing a hierarchical reward function system; executing a federated learning process with differential privacy protection, where each edge agent independently trains a policy network based on local data, and transmits encrypted gradient information to a coordination center for secure aggregation; designing an asynchronous federated update mechanism, dynamically adjusting the set of nodes participating in the global update and the update weights based on the computing power, communication quality, and data quality of the edge nodes; and generating multi-dimensional power system state assessment results and a risk coefficient matrix.

[0010] Optionally, starting the self-repairing agent network includes: building a distributed self-repairing agent network, in which each agent works collaboratively based on a shared risk coefficient matrix; executing a multi-agent negotiation game algorithm, in which each agent performs an optimization game based on local resources and global goals to generate a candidate control parameter set; building a digital twin environment for the power system; executing a Monte Carlo verification process in the digital twin environment, evaluating the performance indicators of the candidate control parameter set through random perturbations and extreme scenario tests, and determining the optimal control parameter set; based on the optimal control parameter set, generating a control instruction sequence including the power generation side, transmission side, substation side and distribution side.

[0011] Optionally, the distributed self-repairing agent network includes a monitoring agent, a diagnosis agent, a decision-making agent and an execution agent.

[0012] The second aspect of the present invention provides an artificial intelligence-based power analysis system, including: a multi-source data acquisition and fusion module, which is used to collect multi-source heterogeneous data of the power system, execute an adaptive data fusion algorithm through an edge computing device, and generate a power feature data set; a hierarchical deep graph neural network module, which is used to establish a hierarchical deep graph neural network model, the model includes four sub-network structures: power generation layer, transmission layer, transformation layer and distribution layer, and a cross-layer information interaction channel is set to connect each sub-network; a bidirectional knowledge distillation module, which is used to execute a bidirectional knowledge distillation algorithm, embed expert domain knowledge parameters into the neural network model parameter space, and extract a rule set from the neural network output to form a parameter-rule bidirectional mapping matrix; a federated reinforcement learning module, which is used to deploy a federated reinforcement learning architecture, perform parameter aggregation on local models on distributed power nodes, and generate power system status assessment results and risk coefficient matrices; a self-repairing intelligent agent network module, which is used to start the self-repairing intelligent agent network, calculate the multi-level power system control parameter set based on the risk coefficient matrix, and output control instructions.

[0013] The beneficial effects of the present invention are as follows: the present invention realizes the collaborative optimization among various levels of the power system through a hierarchical deep graph neural network and a cross-layer information interaction mechanism, thereby improving the system operation efficiency and stability; utilizes a bidirectional knowledge distillation algorithm to establish a parameter-rule mapping matrix, thereby solving the "black box" problem of the AI model and enhancing the model interpretability; based on federated reinforcement learning and differential privacy protection mechanisms, data localization processing and distributed collaborative learning are realized, thereby protecting data privacy and reducing the communication burden; relying on self-healing intelligent agent networks and digital twin environment verification, the system fault response speed and recovery capability are improved, while ensuring the security of the control strategy. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0015] Figure 1 The following is a framework flow chart of an artificial intelligence-based power analysis method.

[0016] Figure 2 Construct a flowchart for a hierarchical deep graph neural network model for an AI-based power analysis approach.

[0017] Figure 3 Flowchart for the implementation of a bidirectional knowledge distillation algorithm for an AI-based power analysis approach. DETAILED DESCRIPTION

[0018] In order to make the purpose, features, and advantages of the present invention more obvious and easy to understand, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the embodiments described below are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0019] Example 1, reference Figures 1 to 3 , which is the first embodiment of the present invention, provides an artificial intelligence-based power analysis method, the framework flow chart is as follows Figure 1 Shown, including: S1: Collect multi-source heterogeneous data from the power system, execute adaptive data fusion algorithms through edge computing devices, and generate power feature data sets.

[0020] In a preferred embodiment, the method specifically includes the following sub-steps: S1.1: Deploy a multi-layer sensor network to collect multi-source heterogeneous data from the power system.

[0021] In this embodiment, the multi-layered sensor network includes high-speed dynamic sensor units on the power generation side, wide-area phasor measurement devices on the transmission network, multi-dimensional situational awareness sensors on the distribution network, and smart meters on the user side. The sampling frequency and response time of the sensor units at each level are adaptively adjusted based on the power level to ensure accurate and timely data collection.

[0022] S1.2: Build a multi-time-scale parallel data processing architecture on edge computing devices to classify and process multi-source heterogeneous data of the power system according to their time characteristics.

[0023] Specifically, the system dynamically allocates computing resources based on the time-varying nature of data types to optimize the processing efficiency of edge computing devices.

[0024] S1.3: Map the physical structure of the power grid into a dynamic graph structure, use graph attention weights to calculate the importance of nodes in the power system, and apply differentiated feature extraction strategies for different types of nodes.

[0025] Specifically, the dynamic graph structure is represented by G=(V,E,A), where V is the node set (including power generation, transmission, transformation, and distribution), E is the edge set, and A is the attribute set. Based on this, a graph attention network is applied to calculate the importance index of each node. The attention coefficients between nodes are processed using the LeakyReLU activation function and softmax normalization. The node importance level is determined based on the node degree, and nodes are classified into three levels: core nodes, important nodes, and general nodes. Finally, differentiated feature extraction strategies are implemented for nodes of different importance levels: for core nodes, a 3-5-layer deep feature extraction network is used to extract full-spectrum features, maintaining a high feature dimension (usually 60%-80% of the original data dimension); for important nodes, a 2-3-layer medium-complexity feature extraction network is used to extract key time domain features and main frequency domain features, maintaining a medium feature dimension (usually 40%-60% of the original data dimension); for general nodes, a lightweight feature extraction algorithm is applied to extract statistical features and trend features, maintaining a low feature dimension (usually 20%-40% of the original data dimension), thereby improving computational efficiency while ensuring analysis quality.

[0026] It is important to note that during the feature extraction process, feature extraction parameters are dynamically adjusted based on the node type and data characteristics. For example, for power generation nodes, the focus is on extracting power fluctuations and frequency characteristics; for transmission nodes, the focus is on extracting load distribution and line impedance characteristics; and for distribution nodes, the focus is on extracting voltage deviation and harmonic content characteristics.

[0027] S1.4: Perform alignment and fusion of multi-source heterogeneous data of the power system to generate a power feature dataset with a spatiotemporal hierarchical structure.

[0028] In this embodiment, the multi-source heterogeneous data of the power system includes but is not limited to electrical parameter data, meteorological and environmental data, equipment status data, and historical operating data. The system establishes associations between these multi-source heterogeneous data through timestamp alignment and spatial coordinate mapping, and uses uncertainty-aware Bayesian reasoning to address data missing and anomalies.

[0029] Furthermore, the generated power feature dataset has a multi-level spatiotemporal structure: the temporal dimension encompasses millisecond-level transient features, second-level dynamic features, minute-level quasi-static features, and hour-level static features; the spatial dimension encompasses a multi-scale spatial structure at the device, site, regional, and system levels; and each feature point contains multidimensional attribute information, establishing relationships between attributes. The power feature dataset uses a hierarchical storage structure, with system-level global features at the top level, followed by regional, site, and device-level features. Each layer of features contains both the time series data and static attribute data for that level, forming a complete spatiotemporal hierarchical structure.

[0030] Optimally, this invention solves the technical challenges of multi-source heterogeneous data processing in power systems, such as the difficulty of fusion and the high real-time requirements, by implementing an adaptive data fusion algorithm on edge computing devices. The edge computing model shifts data processing forward, reducing the computational burden on the central system. A graph attention mechanism is introduced to assess the importance of network nodes, enabling rational resource allocation. A differentiated feature extraction strategy is designed to balance computational efficiency and information integrity. Finally, a power feature dataset with a spatiotemporal hierarchical structure is constructed, providing structured data support for subsequent deep learning analysis.

[0031] S2: Establish a hierarchical deep graph neural network model, which includes four sub-network structures: power generation layer, transmission layer, transformation layer and distribution layer, and set up cross-layer information interaction channels to connect each sub-network.

[0032] In a preferred embodiment, the hierarchical deep graph neural network model construction flow chart is as follows: Figure 2 As shown, it specifically includes the following sub-steps: S2.1: Based on the power system characteristic dataset, construct the basic graph structure of each sub-network.

[0033] This paper represents the entire power system as a multi-layer graph structure G = {G1, G2, G3, G4}, where G1 represents the generation layer, G2 the transmission layer, G3 the transformation layer, and G4 the distribution layer. The underlying graph structure of each subnetwork layer consists of a set of nodes and a set of edges.

[0034] In a preferred embodiment, the present invention designs a dedicated graph neural network structure for each layer based on the characteristics of different layers of the power system: the graph structure design of the power generation layer captures the dynamic interaction between the generator sets, and adopts a hybrid graph convolutional network, which represents the generator sets as nodes and the electrical connections and control dependencies between the units as edges. ,in is the number of power generation nodes, For the feature dimension, the update process of the hybrid graph convolutional network is as follows: ; in, is a nonlinear activation function, is the node feature matrix of the lth power generation layer, is the normalized adjacency matrix, is the learnable weight matrix of layer l, is a self-connection matrix (usually a unit matrix or a learnable diagonal matrix used to retain node information). This design can effectively capture the coupling effects and dynamic response characteristics between different generator sets.

[0035] In view of the cross-regional and large-scale power transmission characteristics of the transmission network, the transmission layer adopts a spatial-spectral domain graph neural network, which represents the transmission line as an edge and the bus as a node. , using the following update rules: ; in, is a nonlinear activation function, is the spectral domain convolution of the Chebyshev polynomial expansion, is the node feature matrix of the lth transmission layer, L T is the Laplace matrix, is the graph attention mechanism, is the fusion coefficient, A T is the adjacency matrix of the transmission layer. This design can simultaneously consider the global topological characteristics and local connectivity of the transmission network.

[0036] In view of the nonlinear characteristics of transformers and power electronic equipment in substations, the substation layer adopts a graph heterogeneity attention network, which represents different types of equipment (such as transformers, reactors, capacitors, etc.) as different types of nodes, and the electrical connections between devices are represented as edges. , the heterogeneous attention mechanism is as follows: ; ; in, is a nonlinear activation function, is the node feature matrix of the lth substation layer, and is a learnable weight matrix (used for transformation of node features and relationship features respectively), is the attention coefficient between node i and node j, LeakReLU is the activation function, and softmax is the normalization function. is the learnable attention vector, The relationship type between node i and node j The eigenvector of is the feature vector of node i and node j. This design can distinguish and handle complex interactions between different types of devices.

[0037] In order to capture the load fluctuation and distributed resource coordination characteristics in the distribution network, the distribution layer adopts a graph time series network, which represents the distribution network nodes as nodes in the graph, the electrical connections as edges, and integrates time series information. , the update rule is: ; in, is a nonlinear activation function, is the node feature matrix of the lth distribution layer (features of distribution equipment), A D is the adjacency matrix of the distribution layer, GCN is the graph convolutional network (capturing the spatial topology), is the node feature matrix of the past k time steps, The long short-term memory network (capturing time series characteristics) is a design that can simultaneously consider the spatial topology and time series characteristics of the distribution network.

[0038] Through the above steps, a hierarchical deep graph neural network model was constructed, which can effectively handle the characteristics and dynamic behaviors of different levels in the power system.

[0039] S2.2: Design a heterogeneous graph convolution operator and use differentiated convolution kernel functions for different types of nodes and edge relationships in the power system.

[0040] Specifically, for different types of nodes (generation nodes, transmission nodes, substation nodes, distribution nodes, load nodes and energy storage nodes), type-specific parameterized convolution kernels are designed. , whose convolution operation is defined as: ; in, is a nonlinear activation function, is the feature vector of node i in layer l, is the node type, The type of node i is The neighbor set of For type The learnable weight matrix of the nodes in layer l, This design enables different types of nodes to use targeted feature extractors.

[0041] Furthermore, for different types of edge relationships (physical connection relationships, electrical related relationships, and power flow relationships), relationship-specific message passing functions are designed. , whose message passing function is defined as: ; in, The type from node i to node j is News, is the edge type, edge-type specific message passing functions, is the feature vector of node i and node j in layer l, is the edge eigenvector.

[0042] Exemplarily, if it is a physical connection relationship, the edge type specific message passing function uses a distance-weighted message passing function; if it is an electrical correlation relationship, the edge type specific message passing function uses a message passing function based on an electrical sensitivity matrix; if it is a power flow relationship, the edge type specific message passing function uses a message passing function based on power flow direction and size.

[0043] Furthermore, to balance model complexity and expressiveness, a hierarchical parameter sharing strategy is adopted, where nodes and edges of similar types within the same hierarchy share some parameters. This design improves the model's generalization capabilities while maintaining heterogeneity.

[0044] S2.3: Construct multi-level cross-layer information interaction channels, including bottom-up state upload channels, top-down constraint downlink channels, and horizontal collaborative optimization channels.

[0045] It should be noted that the bottom-up state upload channel transmits state information from the lower-level network to the upper-level network, enabling state perception and propagation; the top-down constraint downlink channel transmits constraints and control information from the upper-level network to the lower-level network, enabling hierarchical control and optimization; and the horizontal collaborative optimization channel enables collaborative optimization between different sub-networks within the same layer. Furthermore, a gated update mechanism is designed for information transmission across different channels, dynamically adjusting the interaction strength based on the reliability and importance of the information. This design dynamically adjusts the influence of each channel according to different scenarios, improving the adaptability and robustness of the model.

[0046] S2.4: Design a hierarchical graph pooling and upsampling mechanism.

[0047] Specifically, hierarchical clustering based on electrical distance and electrical similarity is used to achieve dimensionality reduction and dimensionality increase of the graph structure, thereby reducing computational complexity while retaining key topological information.

[0048] S2.5: Introduce a physical constraint embedding module to embed the physical laws and operating constraints of the power system as prior knowledge into the loss function and regularization term of the graph neural network.

[0049] It should be noted that the physical laws and operating constraints of the power system are embedded in the graph neural network as prior knowledge to ensure that the model output conforms to the physical characteristics and safe operation requirements of the power system. The physical constraints include power flow constraints, thermal stability constraints, voltage stability constraints, and frequency stability constraints.

[0050] S3: Execute the bidirectional knowledge distillation algorithm to embed the expert domain knowledge parameters into the neural network model parameter space, and extract the rule set from the neural network output to form a parameter-rule bidirectional mapping matrix.

[0051] In a preferred embodiment, the bidirectional knowledge distillation algorithm execution flow chart is as follows: Figure 3 As shown, it specifically includes the following sub-steps: S3.1: Construct an expert knowledge base for the power system and encode the expert domain knowledge into a structured knowledge graph in a parameterized form.

[0052] First, a power system expert knowledge base was constructed, covering four key areas: power system stability analysis rules, fault diagnosis experience, safety constraints, and control decision-making methods. The knowledge base construction process was as follows: Expert knowledge resources from power system experts were collected, including summaries of expert experience, industry standards and specifications, and typical case analysis reports. The collected knowledge was categorized and organized according to attributes such as topic, scope of application, and credibility, and its accuracy and completeness were confirmed through expert review. The organized knowledge was then converted into a structured knowledge graph, where nodes represent conceptual entities in the power system and edges represent relationships between nodes. Weight parameters were assigned to nodes and edges in the knowledge graph to reflect their importance and credibility in the power system.

[0053] In this step, knowledge is parameterized and encoded using triples (entity 1, relationship, entity 2, weight), creating a knowledge graph with 1,000-5,000 nodes and 5,000-20,000 edges. This structured representation allows expert knowledge to be efficiently processed and utilized by computer programs.

[0054] S3.2: Design a knowledge-parameter mapping mechanism to map key parameters in the structured knowledge graph to the corresponding modules of the hierarchical deep graph neural network model.

[0055] First, based on the characteristics of different types of knowledge, their corresponding relationships with each layer of the neural network are determined. Specifically, stability analysis knowledge is mapped to convolutional layer parameters, fault diagnosis knowledge is mapped to feature extraction layer parameters, safety constraints are mapped to loss function parameters, and regulatory decision knowledge is mapped to output layer parameters. Second, an attention guidance mechanism is constructed. This mechanism guides the network to focus on relevant features by converting the importance information of key nodes in the knowledge graph into attention weights in the neural network. Third, parameter initialization and alignment are completed, and expert knowledge parameters are used to initialize the weights of the corresponding modules of the neural network. Finally, a parameter constraint mechanism is designed to limit the value range of the neural network parameters according to the deterministic rules in the expert knowledge. This mapping mechanism enables expert knowledge to be integrated into the different layers of the neural network in an appropriate form while preserving the learning ability of the neural network, realizing the fusion of knowledge and data-driven.

[0056] S3.3: Perform a progressive knowledge distillation process to enable the hierarchical deep graph neural network model to learn different levels of the structured knowledge graph in stages.

[0057] Specifically, the progressive knowledge distillation process consists of four stages: concept identification, relationship understanding, rule application, and decision generation. In the concept identification stage, the neural network is trained to identify basic power system concept entities using node information from the knowledge graph. In the relationship understanding stage, the neural network is trained to understand the causal relationships between power system concepts using edge information from the knowledge graph. In the rule application stage, the neural network is trained to apply power system analysis rules and diagnostic methods using the rule set from the knowledge graph. In the decision generation stage, the neural network is trained to generate control decisions that are consistent with expert experience using decision patterns from the knowledge graph.

[0058] It is important to note that each stage employs a teacher-student model framework, transferring knowledge by minimizing the loss function between the outputs of the student and teacher models. This progressive knowledge distillation process ensures that the neural network systematically learns expert knowledge, from fundamental concepts to advanced decision-making, improving its performance in power system analysis.

[0059] S3.4: Extracting rule sets from hierarchical deep graph neural network models.

[0060] First, local interpretability methods are applied to analyze the model's sensitivity to input features. Feature attribution techniques are used to calculate the sensitivity of the model's prediction results to each input feature and identify key influencing factors. Second, global interpretability methods are applied to extract decision trees and rule sets. Through model distillation or rule extraction algorithms, the complex decision-making process of the neural network is simplified into understandable decision trees or rule sets. Third, rule validation and optimization are performed, and the extracted rules are compared and analyzed with the original expert knowledge to verify their rationality and innovation and confirm their effectiveness. Finally, a rule set knowledge base is constructed and the verified rule sets are integrated into the knowledge base. Attributes such as their source, applicable conditions, and credibility are labeled to facilitate subsequent knowledge updates and applications.

[0061] S3.5: Construct a parameter-rule bidirectional mapping matrix to record the correspondence and influence intensity between the parameters of the hierarchical deep graph neural network model and the rule set.

[0062] Specifically, the correspondence between neural network parameters and rules is analyzed, and the influence of network parameters on the formation of specific rules is determined through sensitivity analysis and back propagation tracing; a parameter-rule bidirectional mapping matrix M is constructed, where the matrix elements M ij The system represents the strength of the correspondence between neural network parameter group i and rule j, a value determined by calculating the influence of the parameters on the rule output. A dynamic update mechanism based on the mapping matrix is established to locate and adjust affected network parameters when expert knowledge base rules are updated. When neural network parameters change, the system identifies and updates potentially changing rules. This creates a closed-loop optimization mechanism for knowledge, data, and rules. Through a bidirectional parameter-rule mapping matrix, the system achieves collaborative optimization among expert knowledge, training data, and extracted rules. This bidirectional mapping mechanism enables the system to continuously update knowledge and adaptively adjust the model, overcoming the shortcomings of traditional neural networks that are difficult to interpret and update, avoiding the static nature of the expert knowledge base, and achieving an organic fusion of knowledge and data.

[0063] Optimally, compared with traditional methods, the bidirectional knowledge distillation algorithm of the present invention significantly improves the accuracy of power system stability analysis, significantly accelerates fault diagnosis, and excels in small-sample learning tasks, maintaining high performance even with significantly reduced data volumes. Furthermore, through interpretability analysis, the present invention extracts multiple new power system analysis rules from the trained model. Expert verification confirms that most of these rules are consistent with professional knowledge, and some provide new analytical perspectives, enriching the expert knowledge base.

[0064] S4: Deploy a federated reinforcement learning architecture to aggregate parameters of local models on distributed power nodes and generate power system state assessment results and risk coefficient matrices.

[0065] Specifically, the method includes the following steps: S4.1: Deploy edge agents at key nodes of the power system, each of which maintains a local environment state space and action space model.

[0066] S4.2: Build a hierarchical reward function system.

[0067] The hierarchical reward function system includes node-level local reward functions, regional-level collaborative reward functions, and system-level global reward functions. Local rewards focus on equipment safety and efficiency, collaborative rewards focus on regional balance and stability, and global rewards focus on system reliability and economy.

[0068] S4.3: Perform a federated learning process with differential privacy protection. Each edge agent independently trains a policy network based on local data and transmits the encrypted gradient information to the coordination center for secure aggregation.

[0069] Specifically, the edge agent first independently trains the policy network based on local data, uses the stochastic gradient descent method to update the network parameters, and uses experience replay and target network fixing techniques to improve training stability and efficiency during training; then, the differential privacy mechanism is adopted to add Gaussian noise to the uploaded parameters to achieve data privacy protection; then, the gradient information is encrypted through homomorphic encryption technology and transmitted to the coordination center, so that the center can perform parameter aggregation in an encrypted state; finally, the coordination center executes a secure aggregation protocol based on secret sharing to calculate the global gradient without obtaining the original gradient, effectively resisting potential malicious behavior interference.

[0070] S4.4: Design an asynchronous federated update mechanism to dynamically adjust the set of nodes participating in the global update and the update weight based on the computing power, communication quality, and data quality of the edge nodes.

[0071] Specifically, the set of nodes participating in the global update is dynamically selected based on the computing power, communication quality, and data quality of the edge nodes, where a multi-factor scoring mechanism is used for node selection. An asynchronous update strategy is implemented, allowing nodes to participate in the global model update at different rates to avoid the efficiency loss caused by waiting for slow nodes. The weight of each node in the global update is dynamically adjusted according to the node's characteristics and contribution. Through version control and gradient compensation mechanisms, the model consistency under asynchronous update conditions is maintained.

[0072] S4.5: Generate multi-dimensional power system status assessment results and risk factor matrix.

[0073] Specifically, the state assessment includes stable state probability, overload risk level, and abnormal trend prediction. The stable state probability is calculated by integrating the prediction results of multiple federated learning models. A weighted voting mechanism is used to integrate the outputs of each model, with weights dynamically adjusted based on the model's historical accuracy. The system categorizes the stable state probability into five levels: extremely stable (>0.95), high stable (0.85-0.95), moderately stable (0.7-0.85), low stable (0.5-0.7), and unstable (<0.5), providing operators with a clear stability assessment. The overload risk level assessment for critical equipment is based on a comprehensive assessment of current load factors, environmental factors, and historical fault data. It is divided into four levels: safe, warning, alert, and dangerous. Each level has a specific load factor threshold and response strategy. Abnormal trend prediction uses time series analysis and pattern recognition techniques to predict the changing trends of system parameters. An LSTM network model is used to achieve multi-timescale predictions for the next 15 minutes, 30 minutes, and 60 minutes. The prediction results include the direction of state change, the rate of change, and the time of possible turning points.

[0074] Furthermore, the risk coefficient matrix is constructed into a three-dimensional structure based on risk type, impact scope, and outbreak time, and the uncertainty level is quantified through information entropy calculation. The risk type dimension includes four major categories: equipment failure risk, power quality risk, supply and demand balance risk, and network topology risk. Each risk category is further subdivided into multiple subcategories, and each subcategory is assigned a risk probability coefficient between 0 and 1. The impact scope dimension describes the potential impact of a risk event and is divided into four levels: equipment level, site level, regional level, and system level. Each level has a clear definition and quantification criteria for the impact scope. The outbreak time dimension represents the time frame in which a risk event may occur. It is divided into immediate risk (0-1 hour), short-term risk (1-24 hours), medium-term risk (1-7 days), and long-term risk (>7 days). Each time frame is assigned a time urgency coefficient between 0 and 1. Each element in the risk coefficient matrix is calculated as the weighted product of the risk type coefficient, impact scope coefficient, and time urgency coefficient. The weights are determined based on power system operation experience. The risk coefficient is divided into five risk rating levels and associated with corresponding response strategies.

[0075] S5: Start the self-repairing intelligent agent network, calculate the multi-level power system control parameter set based on the risk coefficient matrix, and output the control instructions.

[0076] Specifically, the method includes the following steps: S5.1: Build a distributed network of self-repairing agents, where each agent works collaboratively based on a shared risk matrix.

[0077] Among them, the distributed self-repairing intelligent agent network includes monitoring agents, diagnostic agents, decision-making agents and execution agents. These four types of agents are deployed at different levels of nodes in the power system.

[0078] Specifically, the intelligent agents deployed at the nodes on the power generation side simultaneously perform monitoring and execution functions. They are responsible for monitoring the operating status of power generation equipment, collecting key parameters such as the output power, speed, and temperature of the generator sets, and maintaining the local environmental state space Sg (including information such as operating status, spare capacity, and fuel storage) and the action space Ag (including operations such as power adjustment, starting and stopping units, and mode switching). The power generation-side intelligent agent receives risk assessment information from the risk coefficient matrix, paying special attention to high-risk coefficient items such as the risk of power generation equipment failure and the risk of supply and demand balance, and gives priority to adjusting power generation parameters related to high-risk factors. Based on the Actor-Critic architecture, the intelligent agent uses the risk coefficient as an additional input to the state space and dynamically adjusts the output distribution of the policy network to make the control behavior under high-risk conditions more conservative and robust.

[0079] Furthermore, the intelligent agents deployed at substation nodes primarily perform diagnostic and decision-making functions, continuously analyzing power quality and network topology risks in the risk coefficient matrix. They maintain the state space St (including parameters such as voltage, current, and power) and the action space At (including transformer tap adjustment, reactive power compensation control, and line switching). The substation agent aligns the outbreak time dimension in the risk matrix with its own decision-making cycle, implementing emergency response strategies for immediate risks (0-1 hour) and developing preventive control plans for short-term risks (1-24 hours). The agent manages risk response tasks at different timescales through a priority queue to ensure optimal resource allocation.

[0080] Furthermore, agents deployed at distribution station nodes assume monitoring and execution functions, monitoring distribution network parameters and maintaining the local environmental state space Sd (including information such as line load, voltage deviation, and device status) and the action space Ad (including operations such as distributed power generation scheduling, load shifting, and topology adjustment). The distribution station agents pay particular attention to the impact scope dimension of the risk coefficient matrix and activate collaborative defense mechanisms for high-risk factors that could cause regional and system-level impacts. When the risk coefficient exceeds a preset threshold, the relevant distribution station agents form a temporary collaborative alliance and collaboratively develop a defense strategy through a distributed consensus algorithm.

[0081] Furthermore, the intelligent agents deployed at user-side nodes primarily perform decision-making and execution functions, monitoring user loads and distributed resources, and maintaining the local environmental state space Su (containing information such as electricity demand, controllable loads, and user preferences) and the action space Au (including functions such as demand response, energy management, and load forecasting). The user-side intelligent agents receive risk information related to the supply-demand balance from the risk coefficient matrix and dynamically adjust the aggressiveness of the demand response strategy based on the risk level, expanding the range of controllable loads in high-risk situations and prioritizing user comfort in low-risk situations.

[0082] All agents share updated risk matrix information via a secure distributed communication network, employing a blockchain-based data sharing mechanism to ensure information consistency and integrity. The agent network employs a multi-layered collaboration mechanism, encompassing both horizontal collaboration between agents on the same layer and vertical collaboration between agents across layers, creating an adaptive and flexible regulatory network structure. The entire network adheres to the principle of "local decision-making first, global coordination as a guarantee," maximizing decision-making efficiency while ensuring system security.

[0083] S5.2: Execute the multi-agent negotiation game algorithm.

[0084] Specifically, for the potential risks identified in the risk coefficient matrix, each intelligent agent conducts an optimization game based on local resources and global goals, generates multiple sets of candidate control parameter sets, and determines the final control plan through a step-by-step concession method.

[0085] S5.3: Build a digital twin environment for the power system.

[0086] Specifically, the environment integrates electrical models, communication models and environmental models, and can simulate the physical processes and information interactions of each link of power generation, transmission, transformation and distribution, and supports millisecond-level simulation and rapid scene switching.

[0087] S5.4: Perform the Monte Carlo validation process in the digital twin environment to evaluate the performance indicators of the candidate control parameter sets through random perturbations and extreme scenario tests to determine the optimal control parameter set.

[0088] Among them, performance indicators include success rate, stability margin and economic indicators.

[0089] S5.5: Based on the optimal control parameter set, generate a control instruction sequence including the power generation side, transmission side, transformation side and distribution side.

[0090] Specifically, the control instruction sequence includes output adjustment instructions on the power generation side, network topology reconstruction instructions on the transmission side, voltage control instructions on the substation side, and load control instructions on the distribution side. Each instruction is sorted according to time urgency and execution priority, and status checkpoints are set to ensure the controllability of the execution process.

[0091] Furthermore, this embodiment also provides an artificial intelligence-based power analysis system, including: a multi-source data acquisition and fusion module, which is used to collect multi-source heterogeneous data from the power system, execute an adaptive data fusion algorithm through edge computing equipment, and generate a power feature data set; a hierarchical deep graph neural network module, which is used to establish a hierarchical deep graph neural network model, which includes four sub-network structures: power generation layer, transmission layer, transformation layer, and distribution layer, and sets cross-layer information interaction channels to connect each sub-network; a bidirectional knowledge distillation module, which is used to execute a bidirectional knowledge distillation algorithm, embed expert domain knowledge parameters into the neural network model parameter space, and extract a rule set from the neural network output to form a parameter-rule bidirectional mapping matrix; a federated reinforcement learning module, which is used to deploy a federated reinforcement learning architecture, perform parameter aggregation on local models on distributed power nodes, and generate power system status assessment results and risk coefficient matrices; a self-repairing intelligent agent network module, which is used to start the self-repairing intelligent agent network, calculate a multi-level power system control parameter set based on the risk coefficient matrix, and output control instructions.

[0092] In summary, the present invention realizes the collaborative optimization among various levels of the power system through a hierarchical deep graph neural network and a cross-layer information interaction mechanism, thereby improving the system operation efficiency and stability; uses a bidirectional knowledge distillation algorithm to establish a parameter-rule mapping matrix, solves the "black box" problem of the AI model, and enhances the model interpretability; based on federated reinforcement learning and differential privacy protection mechanism, it realizes data localization processing and distributed collaborative learning, protects data privacy and reduces the communication burden; relies on self-healing intelligent agent network and digital twin environment verification, improves the system fault response speed and recovery capability, and ensures the security of the control strategy.

[0093] Example 2, reference Figures 1 to 3 , which is the second embodiment of the present invention, provides an artificial intelligence-based power analysis method. In order to verify the beneficial effects of the present invention, scientific demonstration is carried out through economic benefit calculation and simulation experiments.

[0094] To validate the effectiveness of AI-based power analysis methods in real-world power systems, this study selected a provincial power grid in East China as a test scenario. This grid, covering approximately 100,000 square kilometers, includes 15 power plants, 32 500kV substations, 213 220kV substations, and approximately 1,200 distribution stations at 110kV and below. This diverse network boasts a diverse range of power sources, including hydropower, thermal power, wind power, and photovoltaics, making it an ideal environment for verifying the applicability of this invention.

[0095] First, experiments were conducted in the data collection and fusion phase. A multi-layered sensor network was deployed: high-speed dynamic sensor units with a sampling frequency of 200Hz were installed on the generator side to capture fluctuations in the electrical parameters of the generator sets; 50 wide-area phasor measurement units (PMUs) were deployed at key nodes in the transmission network, with a synchronous sampling rate of 120Hz; and 300 multi-dimensional situational awareness sensors were deployed in the distribution network with a sampling frequency of 1Hz. Approximately 200,000 smart meters were connected to the user side, with data upload intervals of 15 minutes. Edge computing devices, using industrial computers equipped with Intel i7 processors, 32GB of RAM, and a 1TB SSD, communicated with the sensor network via the MQTT protocol. During data processing, graph attention analysis was used to identify 35 core nodes, 127 important nodes, and 843 general nodes. Using a differentiated feature extraction strategy, core nodes retained 75% of the feature information of the original data dimensions, important nodes retained 52%, and general nodes retained 35%, which is consistent with theoretical design values. During the data fusion process, time synchronization accuracy reached 1 millisecond, and spatial coordinate mapping accuracy exceeded 99.5%. According to tests, it only takes about 4.3 hours to process 1TB of raw data at this stage, which is nearly three times more efficient than the traditional centralized processing method (about 12.7 hours). The amount of data after fusion is reduced to about 42% of the original data, while retaining key spatiotemporal features.

[0096] Next, a hierarchical deep graph neural network model was constructed. Specific graph neural network architectures were designed for each layer of the power system: the generation layer employed a hybrid graph convolutional network with an input feature dimension of 72 and 128 hidden layer neurons; the transmission layer employed a spatial-spectral graph neural network with a Chebyshev polynomial expansion order of 3 and 64 spectral filters; the substation layer employed a graph heterogeneous attention network with five defined relationship types and eight attention heads for each relationship; and the distribution layer employed a graph time series network with a time window length of 24 and 128 LSTM hidden units. The four subnetworks are connected via cross-layer information exchange channels. Both the upstream and downstream channels utilize a gated update mechanism, with the gating units using sigmoid functions to control the information flow ratio. The physical constraint embedding module introduces four key constraints: a loss weight of 0.5 for power flow constraints, 0.3 for thermal stability constraints, 0.4 for voltage stability constraints, and 0.3 for frequency stability constraints. The model has approximately 2.35 million parameters and was trained using the Adam optimizer, with an initial learning rate of 0.001, a cosine annealing learning rate strategy, a batch size of 64, and 200 epochs. During training, the total loss function decreased from an initial 4.27 to 0.31. The verification accuracy of each subnetwork layer reached 93.2% for the generation layer, 92.7% for the transmission layer, 90.5% for the transformation layer, and 89.8% for the distribution layer, demonstrating that the model successfully captures the key characteristics of each level of the power system.

[0097] Subsequently, experiments were conducted using a bidirectional knowledge distillation algorithm. First, a power system expert knowledge graph consisting of 3,750 nodes and 17,230 edges was constructed, covering four key areas: stability analysis rules, fault diagnosis experience, safety constraints, and control decision-making methods. In the knowledge-parameter mapping stage, stability analysis knowledge was mapped to the convolutional layer parameters of the graph neural network, influencing approximately 35% of the model parameters; fault diagnosis knowledge was mapped to the feature extraction layer parameters, influencing approximately 28% of the model parameters; safety constraints were mapped to the loss function parameters; and control decision-making knowledge was mapped to the output layer parameters. The progressive knowledge distillation process was performed in four stages. The distillation loss at each stage decreased from 1.85, 1.42, and 1.03 to 0.25, 0.19, 0.15, and 0.11, respectively, demonstrating that knowledge was gradually absorbed by the network. A local interpretability analysis method was used to extract 147 rules from the trained model. After evaluation by five power system experts, 132 rules were found to be consistent with the expert knowledge (89.8% agreement), and 15 rules were found to provide new analytical insights. These new rules focus on the impact of load fluctuations and distributed generation coordination on system stability, as well as the quantitative relationship between environmental factors and equipment failure rates, filling gaps in traditional knowledge. The constructed parameter-rule bidirectional mapping matrix measures 2350×147, recording the correspondence between network parameters and rules. The average mapping strength is 0.43, and the mapping coverage reaches 92.5%. Compared with traditional power analysis methods, this algorithm improves the accuracy of power system stability analysis by 15.7 percentage points and accelerates fault diagnosis by 73.2%. It also maintains an accuracy of 85.3% in a small-sample learning task using only 10% of the training data.

[0098] During the deployment phase of the federated reinforcement learning architecture, edge agents were deployed at 15 power generation nodes, 32 substations, 48 distribution nodes, and 120 key user-side nodes in the power system. Each agent maintained a local environment state space consisting of 60-120 feature dimensions and an action space consisting of 15-35 discrete actions. In the hierarchical reward function design, the local reward weight was 0.4, the collaborative reward weight was 0.3, and the global reward weight was 0.3. For differential privacy protection, the standard deviation of the added Gaussian noise was set to 0.05. During the asynchronous federated update, an average of 65% of the nodes were selected to participate in the global model update each round. Node selection was based on a comprehensive score of three factors: computing power, communication quality, and data quality. After 3000 rounds of training, the system generated a multi-dimensional power system state assessment result and a risk coefficient matrix, including stable state probability, overload risk level, and abnormal trend prediction. This matrix was constructed as a three-dimensional structure with 64 elements, based on four risk types, four levels of impact range, and four outbreak time. The matrix elements ranged from [0, 1], representing the magnitude of the risk coefficient. By comparing historical risk event data, the system achieved an accuracy rate of 91.3% for identifying high-risk events (risk coefficient > 0.8), 86.7% for medium-risk events (0.5 ≤ risk coefficient ≤ 0.8), and 82.5% for low-risk events (risk coefficient < 0.5). The system can provide an average of 18.5 minutes of early warning of potential risk events, a 153% improvement compared to the average warning time of 7.3 minutes for traditional SCADA systems.

[0099] Finally, a self-healing agent network verification experiment was launched. Based on the generated risk coefficient matrix, the self-healing agent network calculated a multi-level power system control parameter set. During the multi-agent negotiation game, the system generated eight candidate control parameter sets, and the final solution was determined through a step-by-step concession method. The constructed power system digital twin environment integrated electrical, communication, and environmental models, supporting millisecond-level simulation. The environment simulated the physical processes and information interactions of the power system, including 3,000 nodes and 4,500 electrical components. 10,000 Monte Carlo validation runs were performed within the digital twin environment, covering scenarios such as N-1 and N-2 safety checks and extreme weather conditions. Each simulation run lasted 24 hours. The verification results showed that the optimal control parameter set, while ensuring system safety, increased the renewable energy consumption rate by 9.4 percentage points, reduced the line loss rate by 2.1 percentage points, and improved the system stability margin by 15.6%. The resulting control instruction sequence included 27 output adjustment instructions on the generation side, 13 network topology reconfiguration instructions on the transmission side, 43 voltage control instructions on the substation side, and 52 load control instructions on the distribution side. These instructions were implemented in three batches based on execution priority. An evaluation of the actual deployment results showed that system response time was reduced from an average of 5.7 minutes to 1.2 minutes, fault recovery time was shortened by an average of 47.3%, and annual power outages were reduced by 32.5%, resulting in an increase in the system's annual economic benefits.

[0100] To comprehensively evaluate the technical advantages of the present invention, a comparative analysis was conducted with existing technical methods, as shown in Table 1. This table compares the differences between the present invention and traditional SCADA systems and conventional machine learning methods in terms of performance indicators, functional characteristics, and technical advantages.

[0101] Table 1 Performance comparison between the present invention and the prior art Performance indicators Traditional SCADA system Conventional machine learning methods The present invention Stability analysis accuracy 78.3% 83.5% 93.2% Fault diagnosis speed Baseline value Increased by 32.1% Increased by 73.2% Risk warning lead time 7.3min 11.5min 18.5min System response time 5.7min 3.8min 1.2min Line loss rate reduced Baseline value 0.8% 2.1% Improved system stability margin Baseline value 5.3% 15.6% Small sample learning ability not applicable 65.2% 85.3% Expert knowledge integration capabilities middle Low high Comparative results demonstrate that the proposed method significantly improves key metrics such as analysis accuracy, warning time, and system reliability, fully validating the innovative and practical value of AI-based power analysis methods. Furthermore, the proposed method implements a bidirectional knowledge distillation mechanism that combines expert knowledge with data-driven approaches, resolving the difficulty of traditional methods in balancing professional experience and adaptive learning, and providing a new technical path for intelligent analysis and control of power systems.

[0102] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. An artificial intelligence-based power analysis method, characterized in that: include: Collect multi-source heterogeneous data from the power system, execute adaptive data fusion algorithms through edge computing devices, and generate power feature data sets; Establish a hierarchical deep graph neural network model, which includes four sub-network structures: power generation layer, transmission layer, transformation layer and distribution layer, and set up cross-layer information interaction channels to connect the sub-networks; Execute a bidirectional knowledge distillation algorithm to embed expert domain knowledge parameters into the neural network model parameter space, and extract the rule set from the neural network output to form a parameter-rule bidirectional mapping matrix; Deploy a federated reinforcement learning architecture to aggregate parameters of local models on distributed power nodes and generate power system state assessment results and risk coefficient matrices; The self-repairing intelligent agent network is started, a multi-level power system control parameter set is calculated based on the risk coefficient matrix, and a control instruction is output.

2. The power analysis method based on artificial intelligence according to claim 1, characterized in that: The execution of the adaptive data fusion algorithm by the edge computing device includes: Deploy a multi-level sensor network to collect multi-source heterogeneous data from the power system, including sensor units on the generation side, transmission network, distribution network and user side; Constructing a multi-time-scale parallel data processing architecture to classify and process the multi-source heterogeneous data of the power system according to their time characteristics; Map the physical structure of the power grid into a dynamic graph structure, calculate the importance of nodes, and apply differentiated feature extraction strategies for different types of nodes; Perform alignment and fusion of multi-source heterogeneous data in the power system to generate a power feature dataset with a spatiotemporal hierarchical structure.

3. The power analysis method based on artificial intelligence according to claim 1, characterized in that: The establishment of a hierarchical deep graph neural network model includes: Construct sub-networks of each layer of the hierarchical deep graph neural network model; Design heterogeneous graph convolution operators and use differentiated convolution kernel functions for different types of nodes and edge relationships; Construct multi-level cross-layer information interaction channels, including status upload channels, constraint download channels, and collaborative optimization channels; A physical constraint embedding module is introduced to embed the physical laws and operating constraints of the power system into the graph neural network as prior knowledge.

4. The power analysis method based on artificial intelligence according to claim 1, characterized in that: The execution of the bidirectional knowledge distillation algorithm includes: Build an expert knowledge base for power systems and encode expert domain knowledge into a structured knowledge graph in a parameterized form; Designing a knowledge-parameter mapping mechanism to map key parameters in the structured knowledge graph to corresponding modules of the hierarchical deep graph neural network model; Performing a progressive knowledge distillation process to enable the hierarchical deep graph neural network model to learn different levels of the structured knowledge graph in stages; extracting a set of rules from the hierarchical deep graph neural network model; A parameter-rule bidirectional mapping matrix is constructed to record the correspondence and influence intensity between the parameters of the hierarchical deep graph neural network model and the rule set.

5. The power analysis method based on artificial intelligence according to claim 1, characterized in that: The deployment of the federated reinforcement learning architecture includes: Deploy edge agents at key nodes of the power system, each of which maintains a local environment state space and action space model; Build a hierarchical reward function system; Perform a federated learning process with differential privacy protection, where each edge agent independently trains a policy network based on local data and transmits encrypted gradient information to the coordination center for secure aggregation; Design an asynchronous federated update mechanism to dynamically adjust the set of nodes participating in the global update and the update weight based on the computing power, communication quality, and data quality of edge nodes; Generate multi-dimensional power system status assessment results and risk factor matrix.

6. The power analysis method based on artificial intelligence according to claim 1, characterized in that: The starting of the self-repairing agent network includes: Build a distributed network of self-repairing agents, each of which works together based on a shared risk coefficient matrix; Execute the multi-agent negotiation game algorithm, where each agent performs an optimization game based on local resources and global goals to generate a set of candidate control parameters; Build a digital twin environment for the power system; Performing a Monte Carlo validation process in the digital twin environment to evaluate the performance indicators of candidate control parameter sets through random perturbations and extreme scenario tests to determine the optimal control parameter set; Based on the optimal control parameter set, a control instruction sequence including the power generation side, the power transmission side, the power transformation side and the power distribution side is generated.

7. The power analysis method based on artificial intelligence according to claim 6, characterized in that: The distributed self-repairing agent network includes a monitoring agent, a diagnosis agent, a decision-making agent and an execution agent.

8. An artificial intelligence-based power analysis system, characterized in that: include: Multi-source data acquisition and fusion module, used to collect multi-source heterogeneous data from the power system, execute adaptive data fusion algorithms through edge computing devices, and generate power feature data sets; A hierarchical deep graph neural network module is used to establish a hierarchical deep graph neural network model. The model includes four sub-network structures: power generation layer, transmission layer, transformation layer, and distribution layer, and a cross-layer information interaction channel is set up to connect the sub-networks; The bidirectional knowledge distillation module is used to execute the bidirectional knowledge distillation algorithm, embed the expert domain knowledge parameters into the neural network model parameter space, and extract the rule set from the neural network output to form a parameter-rule bidirectional mapping matrix; The federated reinforcement learning module is used to deploy a federated reinforcement learning architecture, aggregate parameters of local models on distributed power nodes, and generate power system state assessment results and risk coefficient matrices; The self-repairing intelligent agent network module is used to start the self-repairing intelligent agent network, calculate the multi-level power system control parameter set based on the risk coefficient matrix, and output the control instructions.

Images