CN-122021998-A - Power market multi-subject intelligent decision method based on graphic neural network and blockchain

CN122021998ACN 122021998 ACN122021998 ACN 122021998ACN-122021998-A

Abstract

The invention discloses an electric power market multi-subject intelligent decision method based on a graph neural network and a blockchain, and relates to the technical field of electric power market transaction and artificial intelligence intersection. The method comprises the steps of defining association relations among market subjects based on types of market subjects stored in a blockchain, constructing an initial undirected weighted relation graph by taking each market subject as a node and the association relations among the market subjects as edges, inputting the initial undirected weighted relation graph into a graph neural network model to perform node characteristic optimization, outputting the optimized node relation characteristics, constructing a multi-subject non-cooperative game model according to the optimized node relation characteristics and combining real-time market parameters with power grid safety constraints, performing characterization learning and balanced decision solving to obtain a Nash balanced strategy meeting global constraints, and performing storage and automatic execution, recording and tracing on the Nash balanced strategy through the blockchain intelligent contracts. Thereby realizing efficient and orderly operation of the distributed power market.

Inventors

PAN XI
ZHOU CHAOXI
ZHU YUNAN
XU HAIYANG
PEI ZIXIA
MA JIKE
XU WEI
LIU YUNPENG
SHAN CHAO
LIU LIU

Assignees

国网江苏省电力有限公司营销服务中心
江苏方天电力技术有限公司

Dates

Publication Date: 20260512
Application Date: 20251222

Claims (10)

1. The utility model provides a power market multi-subject intelligent decision method based on a graph neural network and a blockchain, which is characterized by comprising the following steps: defining an association relation among market subjects based on the types of the market subjects stored in the blockchain, and constructing an initial undirected weighted relation graph by taking the market subjects as nodes and the association relation among the market subjects as edges; Inputting the initial undirected weighted relation graph into a graph neural network model adapting to multi-main-body characteristics of the power market, carrying out dimension attention weighted aggregation and global-local characteristic fusion through the graph neural network model, and outputting optimized node relation characteristics, wherein the graph neural network model comprises a dimension attention weighted aggregation module and a global-local characteristic fusion module; According to the optimized node relation characteristics, combining real-time market parameters and power grid security constraints, and constructing a multi-main-body non-cooperative game model; Performing characterization learning and balanced decision solving according to the multi-main-body non-cooperative game model to obtain a Nash balanced strategy meeting global constraint; and carrying out certification and automatic execution on the Nash equilibrium strategy through a blockchain intelligent contract, and recording and tracing the whole decision making process.
2. The method of claim 1, wherein constructing an initial undirected weighted graph with the market subjects as nodes and the relationships between the market subjects as edges comprises: acquiring historical interaction frequency among market subjects recorded in a block chain, and assigning corresponding relation type weights according to the association relation among the market subjects; Weighting and fusing the historical interaction frequency and the relation type weight to obtain a relation graph edge weight; And constructing an initial undirected weighted relation graph by taking the market subjects as nodes and the association relations among the market subjects as edges and combining the relation graph edge weights.
3. The method of claim 1, wherein inputting the initial undirected weighted relationship graph into a graph neural network model that adapts multi-principal features of the power market, performing a dimension attention weighted aggregation and global-local feature fusion by the graph neural network model, and outputting optimized node relationship features, comprising: Extracting space dimension features, time dimension features and preference dimension features according to node pairs in the initial undirected weighted relation graph; splicing the space dimension feature, the time dimension feature and the preference dimension feature to generate an initial association feature vector; Calculating attention coefficients dimension by dimension according to the initial associated feature vectors by dimension according to the dimension attention weighted aggregation module, and carrying out dimension attention coefficient weighted aggregation to generate node local features; And according to the global-local feature fusion module, fusing the local feature vector and the global market feature, and outputting the optimized node relation feature vector.
4. A method according to claim 3, wherein computing attention coefficients dimension by dimension and performing dimension attention coefficient weighted aggregation on the initial associated feature vector according to the dimension attention weighted aggregation module to generate node local features comprises: Respectively carrying out linear projection on the initial association feature vector and the attribute features of the corresponding nodes, and converting the initial association feature vector and the attribute features of the corresponding nodes into hidden layer feature vectors for calculating the adaptive injection force; Based on the hidden layer feature vector, calculating feature semantic association degree between node pairs dimension by dimension to obtain a dimension attention coefficient vector, and carrying out normalization processing on the dimension attention coefficient vector; and aggregating the dimension attention coefficient vector after normalization processing and the hidden layer feature vector of the corresponding neighbor node through Hadamard products to generate node local features.
5. The method of claim 4, wherein fusing the local feature vector with global market features according to the global-local feature fusion module outputs an optimized node relationship feature vector, comprising: Calculating global market features based on hidden layer feature vectors of all market subjects; Performing linear projection and Sigmoid activation on the hidden layer feature vector of each market subject to obtain a global attention weight for adjusting the fusion proportion; and generating an optimized node relation feature vector according to the local feature vector, the global market feature and the global attention weight.
6. The method of claim 1, wherein constructing a multi-subject non-cooperative game model based on the optimized node relationship features in combination with real-time market parameters and grid security constraints, comprises: Taking the optimized node relation characteristic as a static relation representation of each market subject; taking the real-time market parameters and the power grid security constraint as dynamic environment input; Based on the static relation representation and dynamic environment input, a multi-main-body non-cooperative game model which takes each market main body as a game participant, takes the declared electric quantity and the declared electricity price as policy spaces and is constrained by the safety of the power grid is constructed.
7. The method of claim 6, wherein performing token learning and equalization decision solution according to the multi-subject non-cooperative game model results in a nash equalization strategy that satisfies global constraints, comprising: defining a game state of each market subject at a decision time, wherein the game state at least comprises the optimized node relation characteristics, real-time market environment data and power grid operation constraint conditions; Based on a deep double-simulation measurement framework, carrying out low-dimensional characterization learning on the game states, and constructing a compact potential characterization space through measuring the behavior similarity among the states so as to focus on core information related to game decisions; according to the market subject type, designing a differentiated yield function, and introducing global constraint conditions for ensuring safe and stable operation of the system; And carrying out collaborative optimization on the benefit function by improving a gradient rising method, and iteratively updating strategy parameters of each market subject type to obtain a Nash equilibrium strategy meeting global constraint.
8. The method of claim 7, wherein the nash equalization strategy is documented and automatically executed by a blockchain smart contract and the decision-making overall process is recorded and traced, comprising: Automatically executing the Nash equilibrium strategy through the deployed intelligent contract, and calculating and outputting an equilibrium quotation; verifying the balanced quotation through a preset consensus algorithm, and automatically matching the verified balanced quotation according to a preset deal matching logic; If the transaction matching is successful, generating an electronic contract with legal effectiveness; and the balanced quotation, the electronic contract content and the execution state are stored and verified on the blockchain.
9. The method according to claim 1, characterized in that the method further comprises: recording strategy iteration tracks of each market subject in a game decision process and storing the strategy iteration tracks on a blockchain; based on a preset abnormal behavior judgment rule, monitoring and analyzing the strategy iteration track on the blockchain; when the monitoring analysis result meets the abnormal condition, automatically triggering punishment clauses defined in intelligent contracts deployed on the blockchain; The abnormal condition comprises that the change amplitude of strategy parameters of any market subject in adjacent decision rounds exceeds a preset mutation threshold value.
10. An electric power market multi-subject intelligent decision system based on a graph neural network and a blockchain is characterized by comprising: The initial relation diagram construction module is used for defining the association relation among all market subjects based on all market subject types stored in the blockchain, and constructing an initial undirected weighted relation diagram by taking all market subjects as nodes and the association relation among all market subjects as edges; The node relation feature optimization module is used for inputting the initial undirected weighted relation graph into a graph neural network model adapting to multi-main-body features of the power market, carrying out dimension attention weighted aggregation and global-local feature fusion through the graph neural network model, and outputting optimized node relation features, wherein the graph neural network model comprises a dimension attention weighted aggregation module and a global-local feature fusion module; the game model construction module is used for constructing a multi-main non-cooperative game model according to the optimized node relation characteristics and by combining real-time market parameters and power grid security constraints; the Nash equilibrium strategy solving module is used for carrying out characterization learning and equilibrium decision solving according to the multi-main-body non-cooperative game model to obtain a Nash equilibrium strategy meeting global constraint; And the policy automation execution module is used for executing the license storage and automation execution of the Nash equilibrium policy through the blockchain intelligent contract, and recording and tracing the whole decision making process.

Description

Power market multi-subject intelligent decision method based on graphic neural network and blockchain Technical Field The embodiment of the invention relates to the technical field of electric power market transaction and artificial intelligence intersection, in particular to an electric power market multi-subject intelligent decision method based on a graphic neural network and a blockchain. Background With the large-scale access of distributed power sources (photovoltaic, wind power and the like), the electric power market gradually presents multi-main, decentralized and high-frequency transaction characteristics, and the traditional centralized transaction mode has difficulty in adapting to the flexible requirements of distributed electric power transaction. Firstly, the traditional relationship modeling method can only capture local association among the main bodies, cannot fuse multidimensional features such as space, time, preference and the like, and is difficult to reflect the influence of global market environment on local relationship, so that the relationship features are not accurately depicted; secondly, the existing attention mechanism mostly adopts scalar attention allocation, semantic importance of features with different dimensions cannot be distinguished, and feature aggregation effect is poor; thirdly, the multi-main-body game decision is based on a traditional optimization algorithm, global constraints such as dynamic association among main bodies and power grid safety are not fully considered, the problem that balanced solution deviates from actual transaction requirements easily occurs, fourthly, the high-dimensional game state contains a large amount of task irrelevant interference information, so that reinforcement learning characterization learning efficiency is low, stability and accuracy of the game decision are insufficient, meanwhile, a trusted management and control mechanism is lacked in a transaction process, and the problems of parameter tampering, decision tracing difficulty and the like affect market fairness. Although the prior art has independent application of the graph neural network for feature extraction, reinforcement learning for decision optimization and blockchain for data storage, an integrated solution of relationship modeling, feature optimization, game decision-blockchain management and control is not formed, so that the core problems of global characterization of multi-subject relationships, semantic distinction of feature dimensions, effective characterization of game states, trusted execution of transaction processes and the like cannot be solved at the same time, and efficient and orderly operation of a distributed power market is restricted. Disclosure of Invention Aiming at the core problems that the prior art scheme cannot simultaneously solve the overall description of multi-main-body relationships, the semantic distinction of characteristic dimensions, the effective representation of game states, the credible execution of transaction processes and the like, the invention provides the multi-main-body intelligent decision method for the electric power market based on the graph neural network and the blockchain, and the efficient and orderly operation of the distributed electric power market is realized. In a first aspect, an embodiment of the present invention provides a method for multi-main-body intelligent decision-making in an electric power market based on a graph neural network and a blockchain, including: defining an association relation among market subjects based on the types of the market subjects stored in the blockchain, and constructing an initial undirected weighted relation graph by taking the market subjects as nodes and the association relation among the market subjects as edges; Inputting the initial undirected weighted relation graph into a graph neural network model adapting to multi-main-body characteristics of the power market, carrying out dimension attention weighted aggregation and global-local characteristic fusion through the graph neural network model, and outputting optimized node relation characteristics, wherein the graph neural network model comprises a dimension attention weighted aggregation module and a global-local characteristic fusion module; According to the optimized node relation characteristics, combining real-time market parameters and power grid security constraints, and constructing a multi-main-body non-cooperative game model; Performing characterization learning and balanced decision solving according to the multi-main-body non-cooperative game model to obtain a Nash balanced strategy meeting global constraint; and carrying out certification and automatic execution on the Nash equilibrium strategy through a blockchain intelligent contract, and recording and tracing the whole decision making process. As a preferred embodiment, constructing an initial undirected weighted relation graph by taking the market s