Electric power knowledge graph driven agent dialogue system and method

Abstract

The application relates to the technical field of artificial intelligence, in particular to an intelligent agent dialogue system and method driven by a power knowledge graph, which comprises the following steps: a power grid knowledge graph construction module constructs a power knowledge graph covering all links of power generation, power transmission, power transformation, power distribution and power utilization; an intelligent agent construction module receives triad data, converts the triad data into the state of an intelligent agent and an external environment, forms a state space representation, represents the intention of the intelligent agent in the state space, and triggers actions; a dialogue data collection module collects power equipment state and equipment maintenance procedure data through online dialogue; a dialogue data analysis module identifies the dialogue intention of a user, converts dialogue data into a state space of a power grid knowledge graph intelligent agent; and a dispatching instruction analysis module generates power grid dispatching instruction analysis by using the state space and deep reinforcement learning. The power knowledge graph covering all links is constructed, and the structured representation and efficient management of power field knowledge are realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of artificial intelligence, specifically to an intelligent agent dialogue system and method driven by a power knowledge graph, and particularly to an intelligent dialogue system for realizing power grid dispatching instruction analysis using a knowledge graph and reinforcement learning technology. BACKGROUND

[0002] With the increasing complexity and intelligence of the power system, traditional power grid dispatching faces many challenges. Currently, power dispatching mainly relies on human experience for decision-making, which often has slow response speed and poor decision quality when facing complex and dynamic power grid environments. In particular, in large power grid systems, the number of devices is large and the connection relationship is complex, making it difficult for human decision-making to fully consider all factors, which can lead to suboptimal or even incorrect dispatching instructions.

[0003] In recent years, knowledge graph technology has been widely applied in various fields. Knowledge graphs construct structured knowledge representations through entities, relationships, and attributes, which can effectively organize and manage domain knowledge. However, existing power domain knowledge graphs are mainly used for static knowledge representation and simple queries, and cannot support complex dispatching decisions. At the same time, although reinforcement learning technology has shown great potential in decision-making and control, its application in power systems still faces challenges such as high state space dimension, complex decision-making environment, and strict safety constraints.

[0004] In addition, existing intelligent dialogue technology for power systems mostly stays at the level of simple question and answer, lacks deep semantic understanding, and cannot effectively convert natural language into precise power grid dispatching instructions. Traditional dialogue systems are not closely integrated with knowledge graphs and intelligent decision-making technology, making it difficult to achieve knowledge-driven intelligent decision-making.

[0005] Therefore, how to organically combine power knowledge graphs, reinforcement learning technology, and intelligent dialogue systems to build an intelligent agent dialogue system that can realize knowledge-driven decision-making has become an important research direction for the intelligent development of power systems. SUMMARY

[0006] The present application aims to overcome the problems in the prior art and provide an intelligent agent dialogue system and method driven by a power knowledge graph, which organically combines power knowledge graphs, reinforcement learning technology, and intelligent dialogue systems to realize intelligent conversion from natural language to power grid dispatching instructions, improving the efficiency and quality of power grid dispatching.

[0007] The present application proposes an intelligent agent dialogue system driven by a power knowledge graph, which includes:

[0008] A power grid knowledge graph construction module is configured to construct a power grid knowledge graph covering the whole process of power generation, transmission, transformation, distribution and utilization, which includes device parameters, circuit connection information and power grid operation parameters of power generation and dispatching;

[0009] An agent construction module is connected with the power grid knowledge graph construction module and configured to receive triadic data in the power grid knowledge graph, convert the triadic data into internal states and external environment of an agent, form a state space representation, and represent the intention of the agent in the state space and trigger actions.

[0010] A triadic-to-state-action space knowledge conversion engine is configured to accurately map the triadic data to a state-action space understandable by the agent;

[0011] A hierarchical adaptive state representation system is configured to decompose the power grid state into multiple levels of states and dynamically fuse them;

[0012] A multi-objective reward mechanism under the constraint of power grid physics is configured to integrate multiple objectives such as network loss optimization, power balance and device safety;

[0013] A topology-aware intelligent decision-making network is configured to capture the complex topological relationship between power grid devices and generate decisions based thereon;

[0014] An experience accumulation type continuous learning architecture is configured to continuously improve and evolve the capabilities of the agent;

[0015] A dialogue data collection module is connected with the agent construction module and configured to collect power device state and device maintenance procedure data through online dialogue;

[0016] A dialogue data analysis module is connected with the dialogue data collection module and configured to receive data collected by the dialogue data collection module, identify user dialogue intention through a trained intention recognition method, and convert dialogue data into a state space of the power grid knowledge graph agent;

[0017] A dispatching instruction analysis module is connected with the dialogue data analysis module and configured to receive voice recognition content input by a user, use the state space obtained by the dialogue data analysis module, and generate power grid dispatching instruction analysis using deep reinforcement learning.

[0018] Preferably, the triadic-to-state-action space knowledge conversion engine includes:

[0019] A triadic analysis module is configured to receive triadic data stream of the power grid knowledge graph, and decompose the triadic into three basic units of device state, device action and device connection relationship;

[0020] The graph data conversion processor is connected with the triple analysis module and is configured to convert the triples into a graph database storage format and maintain state information and attribute values of the device nodes.

[0021] The state space constructor is connected with the graph data conversion processor and is configured to extract internal operating parameters and state characteristics of the devices from the graph data, identify and quantify topological relationships and mutual influences between the devices, and integrate internal states and external environmental information into a unified state vector.

[0022] The action space definition module is connected with the state space constructor and is configured to identify effective action types in the power grid operation, discretize a continuous action space into a limited action set, and identify prerequisites and limiting factors for action execution.

[0023] Preferably, the hierarchical adaptive state representation system comprises:

[0024] The multi-layer state decomposition module is configured to decompose the power grid state into device layer states, topological layer states, system layer states, and time sequence layer states.

[0025] The state feature extractor is connected with the multi-layer state decomposition module and is configured to filter key factors affecting decision-making from the states of different layers, reduce the dimensionality of the state space, and unify the representation ranges of parameters of different dimensions; the context awareness fusioner is connected with the state feature extractor and is configured to identify the scenario type of the current power grid operation, obtain historical information and environmental factors related to the current decision-making, and integrate state information from different levels and sources.

[0026] The adaptive representation adjuster is connected with the context awareness fusioner and is configured to evaluate the effectiveness and accuracy of the current state representation, dynamically adjust the state dimensionality according to the complexity of decision-making, and detect and correct biases and abnormalities in the state representation.

[0027] Preferably, the multi-objective reward mechanism under the power grid physical constraints comprises:

[0028] The multi-objective evaluation index system is configured to calculate the power grid loss value under the current device state, evaluate the deviation of actual power from the set value of the dispatching instruction, monitor whether the device operating parameters are within a safe range, and analyze the stability of the power grid system under the current state.

[0029] The reward calculation engine is connected with the multi-objective evaluation index system and is configured to calculate original reward values corresponding to each index, unify reward values of different dimensions to the same scale, and integrate reward values of multiple objectives according to weight coefficients.

[0030] A dynamic weight adjustment mechanism connected with the reward calculation engine, configured to identify a scenario type of current power grid operation, determine a priority order of each target according to the scenario, and calculate a weight of each target based on the priority and system state;

[0031] A long-term reward optimization module connected with the dynamic weight adjustment mechanism, configured to evaluate a potential influence of current decision on future state, adjust a discount coefficient of future reward according to a decision time domain, and predict a long-term cumulative reward of a decision sequence.

[0032] As preferred, the topology-aware intelligent decision network comprises:

[0033] A graph structure representation module configured to represent devices as nodes in a graph, encode relationships between devices as edges in the graph, build a complete power grid topology graph based on the nodes and edges, and extract a relevant local subgraph according to a decision requirement;

[0034] An attention mechanism processor connected with the graph structure representation module, configured to calculate an importance score of each node, evaluate an importance degree of different relationship types, and allocate a calculation resource according to the attention score;

[0035] An information aggregation and propagation module connected with the attention mechanism processor, configured to collect state information of neighbors of a node, process neighbor information according to an attention weight, integrate self information and neighbor information to update a node state, and realize multi-hop delivery of information in the graph;

[0036] A decision generation module connected with the information aggregation and propagation module, configured to integrate a node representation after information propagation, generate a global representation reflecting the entire graph structure, and generate an action policy based on the node representation and the graph-level representation.

[0037] As preferred, the experience accumulation type continuous learning architecture comprises:

[0038] An experience memory management module configured to collect experience samples of interaction between an agent and an environment, evaluate learning values of different experience samples, and maintain a limited-capacity experience buffer pool;

[0039] An incremental learning engine connected with the experience memory management module, configured to evaluate a performance and adaptability of a current model, arrange a timing and frequency of incremental update of the model, and optimize model parameters based on new experience;

[0040] A knowledge distillation and migration module connected with the incremental learning engine, configured to select a model with excellent performance as a teacher, construct a student model suitable for knowledge migration, and migrate knowledge of the teacher model to the student model;

[0041] The continuous evaluation and evolution module, connected to the knowledge distillation and transfer module, is used to continuously monitor the performance indicators of the intelligent agent, detect dynamic changes in the power grid environment, analyze the agent's adaptability to environmental changes, and formulate the agent's evolution strategy based on the evaluation results.

[0042] Preferably, the power grid knowledge graph construction module includes:

[0043] The power grid data processing unit is used to acquire power grid operation data, including power grid GIS information and historical data of generators and loads, and to perform deduplication on the power grid operation data. By using data filtering methods, it retains data on power generation and load power within a preset range.

[0044] The knowledge graph modeling unit is connected to the power grid data processing unit and is used to preprocess the power grid data, convert the data into a data format for knowledge graph recognition, use entity names as the basic units of the power grid knowledge graph, use relation names as the logical links between entities in the power grid knowledge graph, and use relation parameters as the strength of the relationship between entities.

[0045] The entity relationship verification unit, connected to the knowledge graph modeling unit, is used to establish an entity relationship model, provide power grid data as input to the entity relationship model and obtain entity relationships as output. It uses a GAT graph attention network to realize entity relationship recognition of the power grid knowledge graph, and verifies the relationships in the power grid knowledge graph to ensure the correctness of entity relationships in the power grid knowledge graph.

[0046] Preferably, the dialogue data analysis module includes:

[0047] The data preprocessing unit is used to preprocess the collected dialogue data, including data cleaning, data fusion, deduplication, imputation and data standardization;

[0048] The attention neural network unit, connected to the data preprocessing unit, is used to perform temporal modeling using bidirectional LSTM, embed contextual information into the attention mechanism to obtain the attention neural network model, and perform intent recognition on the collected dialogue data to obtain the user dialogue intent vector.

[0049] The state space transition unit, connected to the attention neural network unit, is used to transform dialogue data into the state space of the power grid knowledge graph agent. Based on the user's input instructions, the agent performs a series of actions, transitioning the agent from its current state space to the next state, and mapping the agent's transition from the current state to the next state to power grid dispatch instructions.

[0050] Preferably, the scheduling instruction parsing module includes:

[0051] The state estimation and analysis unit is used to obtain state estimation sections and analyze equipment operating modes, equipment operating states, and event information parameters.

[0052] The graph link retraining unit is connected to the state estimation analysis unit and is used to retrain the graph link relationships.

[0053] The candidate information generation unit, connected to the graph link retraining unit, is used to generate labels for candidate device operating mode, candidate device operating status and candidate event information parameters;

[0054] The real-time dialogue execution unit, connected to the candidate information generation unit, is used to conduct dialogue using a retrained power knowledge graph to obtain the real-time operating mode, real-time operating status, and real-time events of the equipment.

[0055] The result sorting output unit, connected to the real-time dialogue execution unit, is used to compare and sort the real-time operating mode, real-time operating status, and real-time events of the device with the predicted candidate device operating mode, candidate device operating status, and candidate event information parameters, and select the data with the higher ranking as the final response.

[0056] A power knowledge graph-driven intelligent agent dialogue method includes the following steps:

[0057] Construct a power knowledge graph covering the entire process of power generation, transmission, transformation, distribution and consumption. The power knowledge graph includes equipment parameters, circuit connection information and power grid operation parameters for power production and dispatch.

[0058] The process of transforming the triplet data in the power knowledge graph into the internal state and external environment of the agent to form a state space representation, and representing the agent's intentions and triggering actions in the state space, includes the following steps:

[0059] Through the knowledge conversion engine from triples to state-action space, triple data is accurately mapped to a state-action space that the agent can understand.

[0060] The hierarchical adaptive state representation system decomposes the power grid state into multi-level states and dynamically merges them.

[0061] By integrating multiple objectives such as network loss optimization, power balance, and equipment safety through a multi-objective reward mechanism under the physical constraints of the power grid;

[0062] By using a topology-aware intelligent decision-making network, complex topological relationships between power grid devices are captured and decisions are generated based on these relationships.

[0063] Through an experience-accumulating continuous learning architecture, the capabilities of intelligent agents can be continuously improved and they can evolve on their own.

[0064] Data on the status of power equipment and equipment maintenance procedures are collected through online dialogue.

[0065] The collected dialogue data is preprocessed, and the user's dialogue intent is identified by a trained intent recognition method. The dialogue data is then transformed into the state space of the power grid knowledge graph agent.

[0066] The system receives user-input speech recognition content and uses the state space obtained in the previous step to generate power grid dispatch instruction parsing using deep reinforcement learning.

[0067] The beneficial effects of this invention include:

[0068] 1. A power knowledge graph covering the entire process of power generation, transmission, transformation, distribution and consumption has been constructed, realizing the structured representation and efficient management of knowledge in the power field, and providing a reliable knowledge foundation for intelligent decision-making.

[0069] 2. An innovative knowledge conversion engine from triples to state-action space was designed, which bridges the gap between static knowledge and dynamic decision-making environment, and solves the technical bottleneck of combining knowledge graphs and reinforcement learning.

[0070] 3. A hierarchical adaptive state representation system is proposed. Through multi-level state decomposition and dynamic fusion, the dimensionality of the state space is effectively reduced, and the representation accuracy and computational efficiency are improved.

[0071] 4. A multi-objective reward mechanism based on power grid physical constraints was designed, which simultaneously considers multiple objectives such as network loss optimization, power balance and equipment safety, and realizes global optimization of power grid dispatch.

[0072] 5. A topology-aware intelligent decision-making network was developed, which can capture the complex topological relationships between power grid devices and make intelligent decisions based on the global network structure, thereby improving the accuracy and practicality of the decisions.

[0073] 6. An experience-accumulating continuous learning architecture was created, enabling the agent to continuously improve its capabilities and evolve itself, thus solving the problems of knowledge obsolescence and performance degradation in traditional systems over long periods of operation.

[0074] 7. The intelligent dialogue system enables the intelligent conversion from natural language to power grid dispatching commands, improving the convenience and efficiency of human-computer interaction.

[0075] 8. The overall system has strong scalability and adaptability, and can cope with power grid systems of different sizes and types, and adapt to the dynamic changes of the power system. Attached Figure Description

[0076] Figure 1 This is the overall architecture diagram of the intelligent agent dialogue system driven by the power knowledge graph of this invention;

[0077] Figure 2 This is a schematic diagram of the structure of the power grid knowledge graph construction module of the present invention;

[0078] Figure 3 This is a schematic diagram of the intelligent agent construction module of the present invention;

[0079] Figure 4 This is a flowchart of the knowledge conversion engine from triples to state-action space of this invention;

[0080] Figure 5 This is a schematic diagram of the working principle of the hierarchical adaptive state representation system of the present invention;

[0081] Figure 6 This is a schematic diagram of the multi-objective reward mechanism under the physical constraints of the power grid in this invention;

[0082] Figure 7 This is a structural diagram of the topology-aware intelligent decision-making network of the present invention;

[0083] Figure 8 This is a framework diagram of the experience-accumulating continuous learning architecture of this invention;

[0084] Figure 9 This is a flowchart of the dialogue data analysis module of the present invention;

[0085] Figure 10 This is a flowchart of the scheduling instruction parsing module of the present invention;

[0086] Figure 11 This is a flowchart of the intelligent agent dialogue method driven by the power knowledge graph of the present invention. Detailed Implementation

[0087] Please refer to the attached document. Figures 1-11 The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments, but the implementation of the present invention is not limited thereto.

[0088] like Figure 1 As shown, this invention provides an intelligent agent dialogue system driven by a power grid knowledge graph, including a power grid knowledge graph construction module 1, an intelligent agent construction module 2, a dialogue data collection module 3, a dialogue data analysis module 4, and a dispatch command parsing module 5. The power grid knowledge graph construction module 1 is connected to the intelligent agent construction module 2, the intelligent agent construction module 2 is connected to the dialogue data collection module 3, the dialogue data collection module 3 is connected to the dialogue data analysis module 4, and the dialogue data analysis module 4 is connected to the dispatch command parsing module 5, forming a complete technical chain.

[0089] Reference Figure 2The power grid knowledge graph construction module 1 is used to construct a power knowledge graph covering the entire process of power generation, transmission, transformation, distribution and consumption (i.e., the five main processes of power generation, transmission, transformation, distribution and consumption in the power system, representing the complete process of power from production to consumption). The power knowledge graph includes equipment parameters, circuit connection information and power grid operation parameters for power production and dispatch.

[0090] In one embodiment of the present invention, the power grid knowledge graph construction module 1 includes a power grid data processing unit 11, a knowledge graph modeling unit 12, and an entity relationship verification unit 13.

[0091] The power grid data processing unit 11 is used to acquire power grid operation data, including power grid GIS information and historical data of generators and loads. It performs deduplication on the power grid operation data, retaining data where generation power and load power are within a preset range using data filtering methods. Preferably, the preset range can be adjusted according to the power grid type and scale. For example, for a 330kV power grid, the generation power range can be set to 50MW-500MW, and the load power range can be set to 20MW-300MW. Such settings can effectively filter abnormal data and improve the reliability of subsequent processing.

[0092] The knowledge graph modeling unit 12 is connected to the power grid data processing unit 11 and is used to preprocess power grid data, including data cleaning, data fusion, deduplication, imputation, and data standardization, converting the data into a data format that can be recognized by the knowledge graph. In this unit, entity names are used as the basic units of the power grid knowledge graph, relation names are used as the logical links between entities in the power grid knowledge graph, and relation parameters are used as the strength of the relationships between entities. For example, the connection relationship between the transformer entity and the switch entity can be represented as (Transformer A, Connection, Switch B, {Connection Strength: 1.0}), where the connection strength represents the determinism of the relationship, and 1.0 represents complete determinism.

[0093] The entity relationship verification unit 13 is connected to the knowledge graph modeling unit 12 and is used to establish an entity relationship model, providing power grid data as input and obtaining entity relationships as output. In this invention, a GAT graph attention network is used to identify entity relationships in the power grid knowledge graph, verify the relationships in the power grid knowledge graph, and ensure the correctness of entity relationships in the power grid knowledge graph.

[0094] The core of the GAT graph attention network is to assign different weights to different neighbor nodes through an attention mechanism, thereby more accurately aggregating neighbor information. The specific implementation is as follows:

[0095] For node i, its representation vector is The attention coefficient of the node j connected to it is calculated using the attention mechanism:

[0096] ,

[0097] Where W is the weight matrix, For attention vectors, This represents the vector concatenation operation. Let i represent the set of neighboring nodes of node i, and LeakyReLU be the activation function.

[0098] Then, the node representation is updated based on the attention coefficient:

[0099] ,

[0100] in, This is a non-linear activation function, such as ReLU. To improve the stability and expressive power of the model, a multi-head attention mechanism is usually employed, i.e.:

[0101] ,

[0102] Where K represents the number of attention heads, This indicates a splicing operation. and These are the attention coefficients and weight matrix of the k-th attention head, respectively.

[0103] In the power knowledge graph, nodes represent power equipment, and edges represent the relationships between equipment. Through the GAT network, the system can automatically learn the importance of different relationships, improving the accuracy of entity relationship recognition. Preferably, in this system, the number of attention heads K=8, the activation function is ELU, the learning rate is 0.005, and the number of training epochs is 200. These parameter settings are based on extensive experimental verification and achieve a good performance balance in the power knowledge graph task.

[0104] Reference Figure 3 The intelligent agent construction module 2 is connected to the power grid knowledge graph construction module 1. It is used to receive triple data in the power knowledge graph, transform the triple data into the internal state and external environment of the intelligent agent, form a state space representation, and represent the intention of the intelligent agent in the state space and trigger actions.

[0105] In an embodiment of the present invention, the agent construction module 2 includes a knowledge conversion engine 21 from triples to state-action space, a hierarchical adaptive state representation system 22, a multi-objective reward mechanism under power grid physical constraints 23, a topology-aware intelligent decision-making network 24, and an experience-accumulating continuous learning architecture 25. The state space refers to the set of all states that the agent may be in, and the action space refers to the set of all actions that the agent may take. Together, they constitute the basis for the agent to make decisions.

[0106] Reference Figure 4 The triple-to-state-action space knowledge conversion engine 21 is used to accurately map triple data to a state-action space that the agent can understand. The engine includes a triple parsing module 211, a graph data conversion processor 212, a state space builder 213, and an action space definition module 214.

[0107] The triplet parsing module 211 receives triplet data streams from the power grid knowledge graph and decomposes the triplets into three basic units: device status, device action, and device connection relationship. For example, for the triplet (Transformer A, Voltage Level, 110kV), this module will identify Transformer A as the device, Voltage Level as the status attribute, and 110kV as the attribute value.

[0108] The graph data conversion processor 212 is connected to the triple parsing module 211 to convert triples into a graph database storage format, maintaining the state information and attribute values ​​of device nodes. In this invention, an attribute graph model is used to store power grid knowledge. Each node represents a device, and the node attributes contain the device's state information. Edges represent the relationships between devices. For example, a transformer node may have attributes such as ID, type, rated capacity, and voltage level, while the edge between the transformer and the switch may have attributes such as connection type and distance.

[0109] The state space builder 213 is connected to the graph data conversion processor 212 and is used to extract the internal operating parameters and state characteristics of the devices from the graph data, identify and quantify the topological relationships and mutual influences between devices, and integrate the internal state and external environment information into a unified state vector. Preferably, the state vector can be represented as:

[0110] ,

[0111] in, This represents the internal state vector, which contains the state parameters of all devices. This represents the external environment vector, which includes the topological relationships and mutual influences between devices.

[0112] The internal state vector can be further represented as:

[0113] ,

[0114] in, This represents the state vector of the i-th device, containing all relevant parameters of that device. For example, for a transformer, its state vector might contain parameters such as "voltage", "current", "temperature", and "load rate".

[0115] The external environment vector can be represented as a combination of the adjacency matrix and the feature matrix:

[0116] ,

[0117] Where A is the adjacency matrix, representing the connection relationships between devices; F is the characteristic matrix, representing the attributes of the connection relationships. For example, for a power grid containing n devices, A is... The matrix, This indicates that a connection exists between device i and device j; otherwise... F may contain information such as connection type and impedance.

[0118] The action space definition module 214 is connected to the state space builder 213 and is used to identify the valid action types in power grid operation, discretize the continuous action space into a finite set of actions, and identify the preconditions and limiting factors for action execution. In power systems, typical actions include switching operations (closing / opening), transformer tap adjustment, capacitor / reactor switching, etc.

[0119] Preferably, the action space can be represented as:

[0120] ,

[0121] in, This represents the i-th action, which may include information such as action type, target device, and operating parameters. For example, adjusting the tap position of transformer T1 to level 5 can be represented as (adjust_tap, T1, 5).

[0122] To reduce the complexity of the action space, this invention employs an action mask mechanism to filter out infeasible actions based on the current state and operational constraints. An action mask can be represented as:

[0123] ,

[0124] in, , Indicates the action in state s It is feasible. This indicates that it is not feasible.

[0125] Reference Figure 5 The hierarchical adaptive state representation system 22 is used to decompose the power grid state into multi-level states and dynamically fuse them. The system includes a multi-level state decomposition module 221, a state feature extractor 222, a context-aware fusion unit 223, and an adaptive representation adjuster 224.

[0126] The multi-layer state decomposition module 221 is used to decompose the power grid state into equipment-level state, topology-level state, system-level state, and time-series state. The equipment-level state focuses on the operating parameters and state characteristics of individual devices; the topology-level state focuses on the connection relationships and interaction patterns between devices; the system-level state focuses on the overall operation status and macroscopic indicators of the power grid; and the time-series state focuses on the dynamic characteristics of the state changing over time.

[0127] The state feature extractor 222 is connected to the multi-level state decomposition module 221, and is used to screen key factors affecting decision-making from each level of state, reduce the dimensionality of the state space, and unify the representation range of parameters with different dimensions. Preferably, principal component analysis (PCA) is used to reduce the state dimensionality while retaining the main information. The calculation process of PCA is as follows:

[0128] 1. Calculate the covariance matrix of the data:

[0129] ,

[0130] in, For sample vectors, The sample mean. This represents the number of samples.

[0131] 2. Calculate the eigenvalues ​​and eigenvectors of the covariance matrix:

[0132] ,

[0133] in, For eigenvalues, This is the corresponding feature vector.

[0134] 3. Select the eigenvectors corresponding to the k largest eigenvalues ​​to construct the projection matrix P:

[0135] ,

[0136] 4. Project the original data into k-dimensional space:

[0137] ,

[0138] In this system, the k value is preferably set so that the amount of information retained reaches more than 95%, that is:

[0139] ,

[0140] Context-aware fusion unit 223 is connected to state feature extractor 222 to identify the current power grid operation scenario type, acquire historical information and environmental factors related to the current decision, and integrate state information from different levels and sources. Scene recognition employs a clustering algorithm to match the current state with predefined scenario templates. Preferably, an improved K-means clustering algorithm is used, taking into account power system characteristics, particularly the periodicity of the load curve.

[0141] State fusion employs an attention mechanism, automatically adjusting the weights of state information at different levels based on the current scene. The fusion process can be represented as:

[0142] ,

[0143] in, This represents the state representation of the i-th layer. This represents the corresponding attention weight, where L represents the layer number. The attention weights are calculated as follows:

[0144] ,

[0145] in, It is an assessment of the status. The function related to the current scene c can be a simple inner product operation. .

[0146] The adaptive representation adjuster 224 is connected to the context-aware fusion unit 223 to evaluate the effectiveness and accuracy of the current state representation, dynamically adjust the state dimension according to the decision complexity, and detect and correct deviations and anomalies in the state representation. Preferably, the state dimension is dynamically adjusted according to a balance between decision accuracy and computational efficiency, using low-dimensional representation to improve efficiency in simple scenarios and high-dimensional representation to improve accuracy in complex scenarios.

[0147] Reference Figure 6 The multi-objective reward mechanism 23 under the physical constraints of the power grid is used to integrate multiple objectives such as network loss optimization, power balance, and equipment safety. This mechanism includes a multi-objective evaluation index system 231, a reward calculation engine 232, a dynamic weight adjustment mechanism 233, and a long-term reward optimization module 234.

[0148] The multi-objective evaluation index system 231 is used to calculate the grid loss value under the current equipment status, assess the deviation between the actual power and the dispatch command setting value, monitor whether the equipment operating parameters are within the safe range, and analyze the stability of the grid system under the current status. Specific evaluation indicators include:

[0149] 1. Network loss index: Calculates the active power loss in the power grid, expressed as:

[0150] ,

[0151] in, Number of busbars Let i be the electrical conductance between nodes i and j. and These represent the voltage magnitude and phase angle at node i, respectively.

[0152] 2. Power Balance Index: The deviation between the actual power and the dispatch command setting is calculated and expressed as:

[0153] ,

[0154] in, For the number of generators, This represents the actual output of the i-th generator. Set values ​​for the corresponding scheduling instructions.

[0155] 3. Equipment safety indicators: Monitoring whether the operating parameters of the equipment are within the safe range, which can be expressed as a function of the degree of exceeding the limit:

[0156] ,

[0157] in, For the number of devices, These are the operating parameters (such as current, power, etc.) of the i-th device. This corresponds to the safety upper limit.

[0158] 4. System stability indicators: The dynamic stability margin of a system can be evaluated by the critical fault clearing time (CCT).

[0159] ,

[0160] in, To account for the number of faults, Let be the critical fault clearing time for the i-th fault.

[0161] The reward calculation engine 232 is connected to the multi-objective evaluation index system 231 to calculate the original reward value corresponding to each index separately, unify reward values ​​of different dimensions to the same scale, and integrate the reward values ​​of multiple objectives according to weight coefficients. Reward normalization adopts the Min-Max method.

[0162] ,

[0163] in, The original reward value. and These are the preset minimum and maximum values, respectively. The formula for calculating the comprehensive reward is as follows.

[0164] The formula for calculating the comprehensive reward is:

[0165] ,

[0166] Where M is the target quantity. The weight of the i-th objective satisfies .

[0167] The dynamic weight adjustment mechanism 233 is connected to the reward calculation engine 232 to identify the current power grid operation scenario type, determine the priority order of each objective based on the scenario, and calculate the weight of each objective based on the priority and system state. Preferably, the weight adjustment strategy is as follows:

[0168] 1. During normal operation: Network loss optimization has a high weight (0.5), while power balance (0.3) and equipment safety weight (0.2) are moderate;

[0169] 2. Peak load phase: Power balance has the highest weight (0.6), network loss optimization has a lower weight (0.2), and equipment safety has a weight (0.2).

[0170] 3. Equipment over-limit stage: Equipment safety weight is significantly increased (0.7), power balance weight is (0.2), and network loss optimization weight is reduced (0.1);

[0171] 4. System instability phase: System stability weight dominates (0.8), while the weights of other objectives decrease (0.067 each).

[0172] To ensure smooth weight changes and avoid decision oscillations, a smooth transition function is used:

[0173] ,

[0174] in, This is a smoothing factor; a larger value indicates a smoother change. Preferably, it is set to... This ensures that weight changes have both fast response and sufficient smoothness.

[0175] The long-term reward optimization module 234 is connected to the dynamic weight adjustment mechanism 233. It is used to evaluate the potential impact of the current decision on future states, adjust the discount coefficient of future rewards according to the decision time domain, and predict the long-term cumulative reward of the decision sequence. The long-term cumulative reward is calculated using a discounted accumulation method.

[0176] ,

[0177] in, As a discount factor, For at any time The instant reward received.

[0178] In power systems, different decisions have significantly different time scales. Therefore, an adaptive discount factor is used to dynamically adjust the discount factor according to the decision type. value:

[0179] Short-term operational decisions (such as switching operations): ;

[0180] Medium-term scheduling decisions (such as intraday economic scheduling): ;

[0181] Long-term planning decisions (such as seasonal adjustments): .

[0182] Reference Figure 7 The topology-aware intelligent decision-making network 24 is used to capture the complex topological relationships between power grid devices and generate decisions based on them. The network includes a graph structure representation module 241, an attention mechanism processor 242, an information aggregation and dissemination module 243, and a decision generation module 244.

[0183] The graph structure representation module 241 is used to represent devices as nodes in the graph, encode the relationships between devices as edges in the graph, construct a complete power grid topology graph based on nodes and edges, and extract relevant local subgraphs according to decision requirements. In this invention, the power grid topology graph G=(V, E) consists of a node set V and an edge set E, where nodes represent power equipment and edges represent physical connections between devices.

[0184] Node features are represented as follows:

[0185] ,

[0186] in, Let d represent the i-th feature of node v, where d is the feature dimension. For example, for a transformer node, features might include voltage level, rated capacity, impedance, etc.

[0187] Edge features are represented as:

[0188] ,

[0189] in, Let the i-th feature of edge e be represented. This refers to the characteristic dimension. For example, for a line edge, characteristics might include impedance, rated capacity, length, etc.

[0190] The attention mechanism processor 242 is connected to the graph structure representation module 241 and is used to calculate the importance score of each node, evaluate the importance of different relationship types, and allocate computing resources according to the attention scores. This invention employs a Graph Attention Network (GAT) to assign dynamic weights to different nodes and edges. Node-level attention calculation is as follows:

[0191] ,

[0192] in, Let W be the feature representation of node i, W be the feature transformation matrix, and a be the attention vector. Let i represent the set of neighbors of node i.

[0193] The edge attention is calculated as follows:

[0194] ,

[0195] in, For the edge Feature representation, W and For learnable parameters, This is the activation function.

[0196] The information aggregation and propagation module 243 is connected to the attention mechanism processor 242. It collects the state information of node neighbors, processes neighbor information according to attention weights, integrates its own information and neighbor information to update the node state, and realizes multi-hop information transmission in the graph. The node state update formula is:

[0197] ,

[0198] in, This represents the state of node i in the l-th layer. For the corresponding attention coefficient, Let be the weight matrix of the l-th layer.

[0199] To capture multi-hop relationships, this invention employs a multi-layer graph attention network, preferably with three layers, enabling each node to perceive information from all nodes within a distance of three hops. This is crucial for power grid topology analysis.

[0200] The decision generation module 244 is connected to the information aggregation and dissemination module 243. It integrates the node representations after information dissemination to generate a global representation reflecting the entire graph structure, and generates action strategies based on the node and graph-level representations. The graph-level representation is obtained by weighted averaging of all node representations.

[0201] ,

[0202] in, Let i be the representation of node i in the last layer. The importance weight of node i can be calculated using a graph-level attention mechanism.

[0203] Based on graph-level representation and key node representation, action policies are generated using a multilayer perceptron (MLP):

[0204] ,

[0205] in, This represents key nodes (such as devices that need to be operated). It is a multilayer perceptron.

[0206] Reference Figure 8 The experience-accumulating continuous learning architecture 25 is used to enable the agent to continuously improve its capabilities and self-evolve. This architecture includes an experience memory management module 251, an incremental learning engine 252, a knowledge distillation and transfer module 253, and a continuous evaluation and evolution module 254.

[0207] The experience memory management module 251 is used to collect experience samples of the agent's interaction with the environment, evaluate the learning value of different experience samples, and maintain a limited-capacity experience buffer pool. The format of the experience samples is as follows: This means that in state s, action a is performed, reward r is obtained, and the user transitions to a new state s'.

[0208] To effectively manage experience samples, this invention employs Priority Experience Playback (PER) technology, which determines the priority of samples based on their TD error:

[0209] ,

[0210] in, Indicates TD error, It is a small positive number to ensure that all samples have a chance to be sampled.

[0211] The sampling probability is directly proportional to the priority:

[0212] ,

[0213] in, Control the degree of priority usage. Indicates uniform sampling. This indicates sampling is performed entirely according to priority. Preferably, it is set to... A balance is struck between priority sampling and random exploration.

[0214] The incremental learning engine 252 is connected to the experience memory management module 251 to evaluate the performance and adaptability of the current model, schedule the timing and frequency of incremental updates to the model, and optimize model parameters based on new experience. This invention uses a deep Q-network (DQN) as the basic algorithm and combines it with a dual network architecture to improve learning stability.

[0215] The formula for updating the Q value is:

[0216] ,

[0217] in, For learning rate, This is the discount factor.

[0218] To avoid instability caused by excessive updates, a target network fixing technique is adopted, that is, using a separate target network. Calculate the target value:

[0219] ,

[0220] The target network parameters are updated every C steps. Preferably, C=1000 is set to achieve a balance between learning stability and timely updates.

[0221] The knowledge distillation and transfer module 253 is connected to the incremental learning engine 252. It is used to select high-performing models as teachers, construct student models suitable for knowledge transfer, and transfer the knowledge of the teacher models to the student models. The core of knowledge distillation is to enable the student models to learn the output distribution of the teacher models, rather than just the final decision.

[0222] The distillation loss function is:

[0223] ,

[0224] in, Cross-entropy loss is given by KL, and KL divergence is given by KL. and The outputs of logits for the teacher model and the student model are respectively. For the softmax function, For temperature parameters, This is a balance factor. Preferably, it is set to... , This allows student models to learn the knowledge distribution of teacher models and adapt to new tasks.

[0225] The continuous evaluation and evolution module 254 is connected to the knowledge distillation and transfer module 253 to continuously monitor the performance indicators of the intelligent agent, detect dynamic changes in the power grid environment, analyze the agent's adaptability to environmental changes, and formulate the agent's evolution strategy based on the evaluation results.

[0226] Performance metrics include average reward, success rate, and stability. Environmental change detection employs statistical hypothesis testing methods, such as the KL divergence test, to detect the difference between the current state distribution and historical distributions. Based on performance evaluation and the degree of environmental change, the system automatically adjusts its learning strategy, such as adjusting the exploration rate, update frequency, and model complexity.

[0227] The dialogue data collection module 3 is connected to the intelligent agent construction module 2 and is used to collect data on the status of power equipment and equipment maintenance procedures through online dialogue.

[0228] This module allows users to describe equipment status and maintenance needs using natural language through a natural language interface. For example, users can input information such as "Transformer No. 3 oil temperature has risen to 75℃" or "Line No. 2 needs maintenance, scheduled for tomorrow from 10:00 to 12:00." The system will convert these natural language descriptions into structured data to update the knowledge graph and agent state space.

[0229] To improve the accuracy of data collection, this module also supports the integration of multiple input methods, such as text input, voice input, and image upload, adapting to different operating environments and user habits. Simultaneously, the system performs preliminary verification of the collected data, checking its completeness and reasonableness, and marking or requesting user confirmation for abnormal data.

[0230] Reference Figure 9 The dialogue data analysis module 4 is connected to the dialogue data collection module 3. It is used to receive the data collected by the dialogue data collection module, identify the user's dialogue intent through the trained intent recognition method, and transform the dialogue data into the state space of the power grid knowledge graph intelligent agent.

[0231] In an embodiment of the present invention, the dialogue data analysis module 4 includes a data preprocessing unit 41, an attention neural network unit 42, and a state space transition unit 43.

[0232] The data preprocessing unit 41 is used to preprocess the collected dialogue data, including data cleaning, data fusion, deduplication, imputation, and data standardization. Data cleaning includes removing irrelevant content, correcting spelling errors, and standardizing terminology; data fusion involves unifying the processing of data from different sources and formats; deduplication avoids bias caused by duplicate data; imputation handles missing values; and data standardization ensures the comparability of data with different dimensions.

[0233] The attention neural network unit 42 is connected to the data preprocessing unit 41 and is used to perform temporal modeling using bidirectional LSTM, embed contextual information into the attention mechanism to obtain the attention neural network model, and perform intent recognition on the collected dialogue data to obtain the user dialogue intent vector.

[0234] The structure of a bidirectional LSTM is as follows:

[0235] ,

[0236] ,

[0237] ,

[0238] in, and These are the hidden states of the forward and backward LSTMs, respectively. It is the input at time t. This is the fused hidden state. The attention mechanism is used to weight the hidden states at different time steps.

[0239] Attention mechanisms are used to weight the hidden states at different time steps:

[0240] ,

[0241] ,

[0242] ,

[0243] Where v, W, and b are learnable parameters. is the attention weight at time t, and c is the context vector.

[0244] Finally, the intent classification is obtained through a fully connected layer and a softmax function:

[0245] ,

[0246] in, and y represents the output layer parameters, and y represents the intention probability distribution.

[0247] The state space transition unit 43 is connected to the attention neural network unit 42 and is used to transform dialogue data into the state space of the power grid knowledge graph agent. According to the user's input instructions, the agent performs a series of actions to transition the agent from its current state space to the next state, and maps the agent's transition from the current state to the next state to the power grid dispatch instructions.

[0248] This process includes:

[0249] 1. Extract key entities and parameters from the dialogue based on the identified intent;

[0250] 2. Map the extracted information to entities and relationships in the knowledge graph;

[0251] 3. Update the corresponding state variables in the state space;

[0252] 4. Generate possible action sequences for subsequent decision-making and execution.

[0253] For example, when the system detects the intention to adjust the tap of transformer No. 1 to level 3, it will extract transformer No. 1 as the target device, tap adjustment as the operation type, level 3 as the target parameter, then update the relevant state of transformer No. 1 in the state space and generate the corresponding scheduling instructions.

[0254] Reference Figure 10 The dispatch instruction parsing module 5 is connected to the dialogue data analysis module 4. It is used to receive the voice recognition content input by the user and generate power grid dispatch instruction parsing using the state space obtained by the dialogue data analysis module and deep reinforcement learning.

[0255] In an embodiment of the present invention, the scheduling instruction parsing module 5 includes a state estimation and analysis unit 51, a graph link retraining unit 52, a candidate information generation unit 53, a real-time dialogue execution unit 54, and a result sorting and output unit 55.

[0256] The state estimation analysis unit 51 is used to acquire state estimation profiles and analyze equipment operating modes, operating states, and event information parameters. State estimation employs a weighted least squares method, calculating system state variables (such as node voltage amplitude and phase angle) based on remote measurement data. The objective function for state estimation is:

[0257] ,

[0258] in, For measured values, State variables The function, This represents the measurement error variance.

[0259] The graph link retraining unit 52 is connected to the state estimation and analysis unit 51 and is used to retrain the graph link relationships. The retraining process uses the latest state estimation results to update the entity relationships in the knowledge graph, especially those dynamic relationships that change over time. For example, updating the line power flow distribution based on load changes, and updating the network topology based on switch state changes, etc.

[0260] The candidate information generation unit 53 is connected to the graph linking and retraining unit 52, and is used to generate labels for candidate equipment operating modes, candidate equipment operating states, and candidate event information parameters. Based on the current state and historical data, this unit predicts possible equipment state changes and event occurrences, generating labels and probability estimates for each possible scenario. For example, for load forecasting, labels such as high load, medium load, and low load may be generated, and corresponding probabilities may be assigned.

[0261] The real-time dialogue execution unit 54 is connected to the candidate information generation unit 53 and is used to conduct dialogue using a retrained power knowledge graph to obtain the real-time operating mode, real-time operating status, and real-time events of the equipment. This unit updates the status information in real time during the dialogue and dynamically adjusts the knowledge graph and state space based on user input and system responses.

[0262] The result sorting output unit 55 is connected to the real-time dialogue execution unit 54. It compares and sorts the real-time operating mode, real-time operating status, and real-time events of the device with the predicted candidate device operating mode, candidate device operating status, and candidate event information parameters, selecting the data with the highest ranking as the final response. The sorting process considers factors such as the reliability of real-time data, the accuracy of prediction, and the safety of operation, determining the final scheduling instruction through a comprehensive score. Preferably, only instructions with scores exceeding a threshold (e.g., 0.8) are recommended for execution, ensuring the safety and reliability of the scheduling decision.

[0263] Reference Figure 11 The present invention also provides a power knowledge graph-driven intelligent agent dialogue method, comprising the following steps:

[0264] 1. Construct a power knowledge graph covering the entire process of power generation, transmission, transformation, distribution and consumption. The power knowledge graph includes equipment parameters, circuit connection information and power grid operation parameters for power production and dispatch.

[0265] 2. Transforming the triplet data in the power knowledge graph into the internal state and external environment of the agent to form a state space representation, and representing the agent's intention and triggering actions in the state space, this step includes:

[0266] Through the knowledge conversion engine from triples to state-action space, triple data is accurately mapped to a state-action space that the agent can understand.

[0267] The hierarchical adaptive state representation system decomposes the power grid state into multi-level states and dynamically merges them.

[0268] By integrating multiple objectives such as network loss optimization, power balance, and equipment safety through a multi-objective reward mechanism under the physical constraints of the power grid;

[0269] By using a topology-aware intelligent decision-making network, complex topological relationships between power grid devices are captured and decisions are generated based on these relationships.

[0270] Through an experience-accumulating continuous learning architecture, the capabilities of intelligent agents can be continuously improved and they can evolve on their own.

[0271] 3. Collect data on the status of power equipment and equipment maintenance procedures through online dialogue;

[0272] 4. Preprocess the collected dialogue data, identify the user's dialogue intent using a trained intent recognition method, and transform the dialogue data into the state space of the power grid knowledge graph agent.

[0273] 5. Receive the user's voice recognition input, and use the state space obtained in the previous step to generate a parser for power grid dispatching instructions using deep reinforcement learning.

[0274] Each step of this method corresponds one-to-one with the function of each module of the above system, and the specific implementation is also completely consistent, so it will not be repeated here.

[0275] To more clearly illustrate the working process of this invention, a complete application example is given below.

[0276] Suppose a power dispatcher needs to handle a power grid dispatching problem during a peak load period. The specific scenario is: the load in an urban area is growing rapidly, and the load on a certain transmission line is approaching its limit. It is necessary to adjust the output of generator sets and the tap positions of transformers to balance the load and reduce network losses.

[0277] 1. The power grid knowledge graph construction module collects and processes power grid GIS information and historical operation data to construct a knowledge graph containing equipment such as generators, transformers, lines, and loads. For example, it forms the following triples:

[0278] (Generator G1, rated capacity, 300MW);

[0279] (Generator G1, connection, bus B1);

[0280] (Transformer T1, tap position, 2);

[0281] (Line L1, connection, {bus B1, bus B2});

[0282] (Line L1, impedance, 0.05 + j0.2);

[0283] (Load D1, Power, 200MW);

[0284] 2. The agent construction module transforms these triples into a state-space representation. The internal state includes parameters of each device, such as generator output, transformer tap position, and line power flow; the external environment includes the topological relationships and mutual influences between devices.

[0285] 3. The dialogue data collection module receives voice input from the dispatcher: The load rate of line 3 has reached 85%. Please suggest how to adjust the generator output and transformer parameters.

[0286] 4. The dialogue data analysis module uses an attention neural network to identify the user's intent as a request for scheduling suggestions. The key entity is line 3, and the key parameter is the load rate of 85%. This information is then transformed into constraints in the state space.

[0287] 5. The scheduling instruction parsing module initiates the decision-making process based on the current state space and user requirements:

[0288] Obtain the state estimation results to confirm the current power grid operating status;

[0289] Retrain the graph link relationships and update the power flow distribution model;

[0290] Generate candidate adjustment schemes, such as increasing the output of G2, decreasing the output of G1, and adjusting the T1 tap changer;

[0291] The effectiveness of each option is evaluated using a topology-aware intelligent decision-making network.

[0292] The proposals are scored and ranked according to a multi-objective reward mechanism (considering network loss, power balance, and equipment safety);

[0293] Choose the optimal solution as the scheduling suggestion.

[0294] 6. The system provides the dispatcher with the following suggestions: reduce the output of generator No. 1 by 30MW and increase the output of generator No. 2 by 35MW. At the same time, adjust the tap of the main transformer No. 1 from level 2 to level 3. This will reduce the load rate of line No. 3 to 75% and reduce network losses by about 5%.

[0295] 7. After the scheduler confirms the execution suggestion, the system stores the interaction experience in the experience buffer pool for subsequent reinforcement learning and model optimization.

[0296] 8. After the execution results are fed back to the system, the experience-accumulating continuous learning architecture evaluates the difference between the actual effect and the expectation, updates the model parameters, and improves the accuracy of future decisions.

[0297] This example demonstrates that the present invention can effectively combine power knowledge graphs, reinforcement learning, and intelligent dialogue technologies to achieve intelligent conversion from natural language to power grid dispatching instructions, thereby improving the efficiency and quality of power grid dispatching.

[0298] This invention provides a power knowledge graph-driven intelligent agent dialogue system and method, which organically combines power knowledge graphs, reinforcement learning technology, and intelligent dialogue systems to achieve intelligent conversion from natural language to power grid dispatching commands. The system constructs a knowledge graph covering all aspects of the power industry and innovatively designs a knowledge conversion engine from triples to state-action space, a hierarchical adaptive state representation system, a multi-objective reward mechanism, a topology-aware intelligent decision-making network, and an experience-accumulating continuous learning architecture. This effectively solves the technical bottlenecks of combining knowledge graphs and reinforcement learning, improving the efficiency and quality of power grid dispatching.

[0299] This invention is applicable to power systems of various sizes and types, possessing strong scalability and adaptability, and can provide powerful support for the intelligent development of power systems. In the future, system functions can be further expanded, such as adding fault diagnosis, predictive maintenance, and emergency response capabilities, further enhancing the intelligence level of power systems.

[0300] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A power knowledge graph-driven intelligent agent dialogue system, characterized in that, include: The power knowledge graph construction module is used to construct a power knowledge graph covering the entire process of power generation, transmission, transformation, distribution and consumption. The power knowledge graph includes equipment parameters, circuit connection information and power grid operation parameters for power production and dispatch. The intelligent agent construction module is connected to the power knowledge graph construction module. It is used to receive triple data in the power knowledge graph, transform the triple data into the internal state and external environment of the intelligent agent, form a state space representation, and represent the intention of the intelligent agent and trigger actions in the state space. The intelligent agent construction module includes: A knowledge conversion engine from triples to state-action space is used to accurately map the triple data to a state-action space that the agent can understand. A hierarchical adaptive state representation system is used to decompose the power grid state into multi-level states and dynamically fuse them. A multi-objective reward mechanism under the physical constraints of the power grid is used to integrate multiple objectives such as network loss optimization, power balance and equipment safety; Topology-aware intelligent decision-making networks are used to capture complex topological relationships between power grid devices and generate decisions based on these relationships. An experience-accumulating continuous learning architecture is used to enable agents to continuously improve their capabilities and evolve themselves. The dialogue data collection module, connected to the intelligent agent construction module, is used to collect data on the status of power equipment and equipment maintenance procedures through online dialogue. The dialogue data analysis module is connected to the dialogue data collection module and is used to receive the data collected by the dialogue data collection module, identify the user's dialogue intent through a trained intent recognition method, and transform the dialogue data into the state space of the power knowledge graph intelligent agent. The dispatch instruction parsing module, connected to the dialogue data analysis module, is used to receive the voice recognition content input by the user, and generate power grid dispatch instruction parsing using the state space obtained by the dialogue data analysis module and deep reinforcement learning. The knowledge conversion engine from triples to state-action space includes: The triplet parsing module is used to receive triplet data streams from the power knowledge graph and decompose triplets into three basic units: device status, device action, and device connection relationship. The graph data conversion processor is connected to the triple parsing module and is used to convert triples into a graph database storage format to maintain the status information and attribute values ​​of device nodes. A state space builder, connected to the graph data conversion processor, is used to extract the internal operating parameters and state characteristics of the devices from the graph data, identify and quantify the topological relationships and mutual influences between devices, and integrate the internal state and external environment information into a unified state vector. The action space definition module, connected to the state space builder, is used to identify the effective action types in power grid operation, discretize the continuous action space into a finite set of actions, and identify the preconditions and limiting factors for action execution. The hierarchical adaptive state representation system includes: The multi-layer state decomposition module is used to decompose the power grid state into equipment layer state, topology layer state, system layer state and timing layer state; A state feature extractor, connected to the multi-layer state decomposition module, is used to filter key factors affecting decision-making from each layer of state, reduce the dimensionality of the state space, and unify the representation range of parameters with different dimensions; A context-aware fusion unit, connected to the state feature extractor, is used to identify the current power grid operation scenario type, obtain historical information and environmental factors related to the current decision, and integrate state information from different levels and sources. An adaptive representation adjuster, connected to the context-aware fusion unit, is used to evaluate the effectiveness and accuracy of the current state representation, dynamically adjust the state dimension according to the decision complexity, and detect and correct deviations and anomalies in the state representation.

2. The power knowledge graph-driven intelligent agent dialogue system according to claim 1, characterized in that, The multi-objective reward mechanism under the physical constraints of the power grid includes: A multi-objective evaluation index system is used to calculate the power grid loss value under the current equipment status, assess the deviation between the actual power and the dispatch command setting value, monitor whether the equipment operating parameters are within the safe range, and analyze the stability of the power grid system under the current status. The reward calculation engine is connected to the multi-objective evaluation index system and is used to calculate the original reward value corresponding to each index separately, unify the reward values ​​of different dimensions to the same scale, and integrate the reward values ​​of multiple objectives according to the weight coefficient. A dynamic weight adjustment mechanism, connected to the reward calculation engine, is used to identify the current power grid operation scenario type, determine the priority order of each objective according to the scenario, and calculate the weight of each objective based on the priority and system state. The long-term reward optimization module, connected to the dynamic weight adjustment mechanism, is used to assess the potential impact of the current decision on the future state, adjust the discount coefficient of the future reward according to the decision time domain, and predict the long-term cumulative reward of the decision sequence.

3. The power knowledge graph-driven intelligent agent dialogue system according to claim 1, characterized in that, The topology-aware intelligent decision-making network includes: The graph structure representation module is used to represent devices as nodes in a graph, encode the relationships between devices as edges in a graph, construct a complete power grid topology graph based on nodes and edges, and extract relevant local subgraphs according to decision requirements. An attention mechanism processor, connected to the graph structure representation module, is used to calculate the importance score of each node, evaluate the importance of different relationship types, and allocate computing resources according to the attention score; The information aggregation and dissemination module, connected to the attention mechanism processor, is used to collect the state information of the node's neighbors, process the neighbor information according to the attention weight, integrate its own information and neighbor information to update the node's state, and realize the multi-hop transmission of information in the graph. The decision generation module, connected to the information aggregation and propagation module, is used to integrate the node representations after information propagation, generate a global representation reflecting the entire graph structure, and generate action strategies based on the node representations and graph-level representations.

4. The power knowledge graph-driven intelligent agent dialogue system according to claim 1, characterized in that, The experience-accumulating continuous learning architecture includes: The experience memory management module is used to collect experience samples of the agent's interaction with the environment, evaluate the learning value of different experience samples, and maintain a limited-capacity experience buffer pool. An incremental learning engine, connected to the experience memory management module, is used to evaluate the performance and adaptability of the current model, schedule the timing and frequency of incremental updates to the model, and optimize model parameters based on new experiences. The knowledge distillation and transfer module, connected to the incremental learning engine, is used to select high-performance models as teachers, construct student models suitable for knowledge transfer, and transfer the knowledge of the teacher model to the student model. The continuous evaluation and evolution module, connected to the knowledge distillation and transfer module, is used to continuously monitor the performance indicators of the intelligent agent, detect dynamic changes in the power grid environment, analyze the agent's adaptability to environmental changes, and formulate the agent's evolution strategy based on the evaluation results.

5. The power knowledge graph-driven intelligent agent dialogue system according to claim 1, characterized in that, The power knowledge graph construction module includes: The power grid data processing unit is used to acquire power grid operation data, including power grid GIS information and historical data of generators and loads, and to perform deduplication on the power grid operation data. By using data filtering methods, it retains data on power generation and load power within a preset range. The knowledge graph modeling unit is connected to the power grid data processing unit and is used to preprocess the power grid data, convert the data into a data format for knowledge graph recognition, use entity names as the basic unit of the power knowledge graph, use relation names as the logical link between entities in the power knowledge graph, and use relation parameters as the strength of the relationship between entities. The entity relationship verification unit, connected to the knowledge graph modeling unit, is used to establish an entity relationship model, provide power grid data as input to the entity relationship model and obtain entity relationships as output. It uses a GAT graph attention network to realize entity relationship recognition of the power knowledge graph, and verifies the relationships in the power knowledge graph to ensure the correctness of entity relationships in the power knowledge graph.

6. The power knowledge graph-driven intelligent agent dialogue system according to claim 1, characterized in that, The dialogue data analysis module includes: The data preprocessing unit is used to preprocess the collected dialogue data, including data cleaning, data fusion, deduplication, imputation and data standardization; The attention neural network unit, connected to the data preprocessing unit, is used to perform temporal modeling using bidirectional LSTM, embed contextual information into the attention mechanism to obtain the attention neural network model, and perform intent recognition on the collected dialogue data to obtain the user dialogue intent vector. The state space transition unit, connected to the attention neural network unit, is used to transform dialogue data into the state space of the power knowledge graph agent. Based on the user's input instructions, the agent performs a series of actions, transitioning the agent from its current state space to the next state. The transition of the agent from the current state to the next state is mapped to power grid dispatch instructions.

7. The power knowledge graph-driven intelligent agent dialogue system according to claim 1, characterized in that, The scheduling instruction parsing module includes: The state estimation and analysis unit is used to obtain state estimation profiles and analyze equipment operating modes, equipment operating states, and event information parameters. The graph link retraining unit is connected to the state estimation analysis unit and is used to retrain the graph link relationships. The candidate information generation unit, connected to the graph link retraining unit, is used to generate labels for candidate device operating mode, candidate device operating status and candidate event information parameters; The real-time dialogue execution unit, connected to the candidate information generation unit, is used to conduct dialogue using a retrained power knowledge graph to obtain the real-time operating mode, real-time operating status, and real-time events of the equipment. The result sorting output unit, connected to the real-time dialogue execution unit, is used to compare and sort the real-time operating mode, real-time operating status, and real-time events of the device with the predicted candidate device operating mode, candidate device operating status, and candidate event information parameters, and select the data with the higher ranking as the final response.

8. A power knowledge graph-driven agent dialogue method, employing the power knowledge graph-driven agent dialogue system as described in any one of claims 1-7, characterized in that, Includes the following steps: Construct a power knowledge graph covering the entire process of power generation, transmission, transformation, distribution and consumption. The power knowledge graph includes equipment parameters, circuit connection information and power grid operation parameters for power production and dispatch. The process of transforming the triplet data in the power knowledge graph into the internal state and external environment of the agent to form a state space representation, and representing the agent's intentions and triggering actions in the state space, includes the following steps: Through the knowledge conversion engine from triples to state-action space, triple data is accurately mapped to a state-action space that the agent can understand. The hierarchical adaptive state representation system decomposes the power grid state into multi-level states and dynamically merges them. By integrating multiple objectives such as grid loss optimization, power balance, and equipment safety through a multi-objective reward mechanism under the physical constraints of the power grid; By using a topology-aware intelligent decision-making network, complex topological relationships between power grid devices are captured and decisions are generated based on these relationships. Through an experience-accumulating continuous learning architecture, the capabilities of intelligent agents can be continuously improved and they can evolve on their own. Data on the status of power equipment and equipment maintenance procedures are collected through online dialogue. The collected dialogue data is preprocessed, and the user's dialogue intent is identified by a trained intent recognition method. The dialogue data is then transformed into the state space of the power knowledge graph agent. The system receives user-input speech recognition content and uses the state space obtained in the previous step to generate power grid dispatch instruction parsing using deep reinforcement learning.

Images