Power market multi-agent intelligent decision-making method based on graph neural network and block chain

Abstract

The invention discloses an electricity market multi-agent intelligent decision-making method based on a graph neural network and a block chain, and relates to the technical field of electricity market transaction and artificial intelligence crossing. The method comprises the following steps: defining an association relationship between market subjects based on the types of the market subjects stored in a block chain, and constructing an initial undirected weighted relation graph by taking the market subjects as nodes and the association relationship between the market subjects as edges; inputting the initial undirected weighted relation graph into a graph neural network model for node feature optimization, outputting optimized node relation features, constructing a multi-subject non-cooperative game model according to the optimized node relation features in combination with real-time market parameters and power grid security constraints, and performing characterization learning and equilibrium decision solving to obtain a multi-subject non-cooperative game model; a Nash equilibrium strategy meeting global constraints is obtained; and the Nash equilibrium strategy is subjected to evidence storage and automatic execution through the block chain smart contract, and is recorded and traced. Therefore, efficient and orderly operation of the distributed power market is realized.

Description

Technical Field

[0001] The embodiments of this invention relate to the interdisciplinary fields of electricity market trading and artificial intelligence, and in particular to a multi-stakeholder intelligent decision-making method for the electricity market based on graph neural networks and blockchain. Background Technology

[0002] With the large-scale integration of distributed power sources (photovoltaics, wind power, etc.), the electricity market is gradually exhibiting characteristics of multi-entity, decentralized, and high-frequency trading. Traditional centralized trading models are no longer suitable for the flexible needs of distributed power trading. Currently, multi-entity relationship modeling and game theory decision-making in the electricity market face several technical bottlenecks: First, traditional relationship modeling methods can only capture local correlations between entities, failing to integrate multi-dimensional features such as space, time, and preferences, and thus failing to reflect the impact of the global market environment on local relationships, resulting in inaccurate characterization of relationship features. Second, existing attention mechanisms mostly employ scalar attention allocation, failing to distinguish the semantic importance of features across different dimensions, leading to poor feature aggregation. Third, multi-entity game theory decision-making relies heavily on traditional optimization algorithms, failing to fully consider dynamic correlations between entities and global constraints such as grid security, easily resulting in equilibrium solutions deviating from actual trading needs. Fourth, high-dimensional game states contain a large amount of task-irrelevant interference information, leading to low efficiency in reinforcement learning representation learning and insufficient stability and accuracy in game theory decision-making. Simultaneously, the lack of a reliable control mechanism in the trading process, coupled with issues such as parameter tampering and difficulties in decision tracing, affects market fairness.

[0003] Although existing technologies include graph neural networks for feature extraction, reinforcement learning for decision optimization, and blockchain for data storage, an integrated solution encompassing relationship modeling, feature optimization, game decision-making, and blockchain management has not yet been formed. This solution cannot simultaneously address core issues such as global characterization of multi-agent relationships, semantic differentiation of feature dimensions, effective representation of game states, and reliable execution of transaction processes, thus hindering the efficient and orderly operation of the distributed power market. Summary of the Invention

[0004] This invention addresses the core issues that existing technical solutions cannot simultaneously solve, such as global characterization of multi-agent relationships, semantic differentiation of feature dimensions, effective representation of game states, and reliable execution of transaction processes. It provides a multi-agent intelligent decision-making method for the power market based on graph neural networks and blockchain, which enables the efficient and orderly operation of the distributed power market.

[0005] In a first aspect, embodiments of the present invention provide a multi-stakeholder intelligent decision-making method for the electricity market based on graph neural networks and blockchain, comprising:

[0006] Based on the types of market entities stored in the blockchain, the relationships between market entities are defined, and an initial undirected weighted relationship graph is constructed with each market entity as a node and the relationships between market entities as edges.

[0007] The initial undirected weighted relationship graph is input into a graph neural network model adapted to the multi-entity characteristics of the electricity market. The graph neural network model performs dimensional attention weighted aggregation and global-local feature fusion to output optimized node relationship features. The graph neural network model includes a dimensional attention weighted aggregation module and a global-local feature fusion module.

[0008] Based on the optimized node relationship characteristics, and combined with real-time market parameters and power grid security constraints, a multi-agent non-cooperative game model is constructed.

[0009] Based on the multi-agent non-cooperative game model, representation learning and equilibrium decision-making are performed to obtain the Nash equilibrium strategy that satisfies global constraints.

[0010] The Nash equilibrium strategy is stored and automatically executed through a blockchain smart contract, and the entire decision-making process is recorded and traced.

[0011] As a preferred implementation, an initial undirected weighted relationship graph is constructed using each market entity as a node and the relationships between market entities as edges, including:

[0012] Obtain the historical interaction frequency between various market entities recorded in the blockchain, and assign corresponding relationship type weights based on the relationships between the market entities;

[0013] The historical interaction frequency and the relation type weight are weighted and fused together to obtain the edge weights of the relation graph;

[0014] Using each market entity as a node and the relationships between them as edges, an initial undirected weighted relationship graph is constructed by combining the edge weights of the relationship graph.

[0015] In a preferred embodiment, the initial undirected weighted relationship graph is input into a graph neural network model adapted to the multi-entity characteristics of the electricity market. The graph neural network model performs dimensional attention weighted aggregation and global-local feature fusion to output optimized node relationship features, including:

[0016] Based on the node pairs in the initial undirected weighted relation graph, extract spatial dimension features, temporal dimension features, and preference dimension features;

[0017] The spatial dimension features, temporal dimension features, and preference dimension features are concatenated to generate an initial associated feature vector;

[0018] The dimension attention weighted aggregation module calculates the attention coefficients of the initial associated feature vector dimension by dimension and performs dimension attention coefficient weighted aggregation to generate local node features.

[0019] The global-local feature fusion module fuses the local feature vector with the global market features to output an optimized node relationship feature vector.

[0020] In a preferred embodiment, the attention coefficients of the initial associated feature vector are calculated dimension by dimension and the dimensional attention coefficients are weighted and aggregated according to the dimensional attention weighting aggregation module to generate local node features, including:

[0021] The initial associated feature vector and the attribute features of the corresponding nodes are linearly projected to convert them into hidden layer feature vectors adapted for attention calculation.

[0022] Based on the hidden layer feature vector, the feature semantic correlation degree between node pairs is calculated dimension by dimension to obtain the dimensional attention coefficient vector, and the dimensional attention coefficient vector is normalized.

[0023] The normalized dimensional attention coefficient vector is aggregated with the hidden layer feature vectors of the corresponding neighboring nodes through the Hadamard product to generate local node features.

[0024] In a preferred embodiment, the local feature vector and global market features are fused according to the global-local feature fusion module, and an optimized node relationship feature vector is output, including:

[0025] Calculate global market characteristics based on the hidden layer feature vectors of each market participant;

[0026] Linear projection and Sigmoid activation are performed on the hidden layer feature vector of each market entity to obtain global attention weights used to adjust the fusion ratio;

[0027] Based on the local feature vector, global market features, and global attention weight, an optimized node relationship feature vector is generated.

[0028] As a preferred implementation, based on the optimized node relationship characteristics and combined with real-time market parameters and power grid security constraints, a multi-agent non-cooperative game model is constructed, including:

[0029] The optimized node relationship features are used as the static relationship representation of each market entity;

[0030] The real-time market parameters and power grid security constraints are used as dynamic environment inputs.

[0031] Based on the static relationship representation and dynamic environment input, a multi-agent non-cooperative game model is constructed, with each market entity as the game participant, the declared electricity volume and declared electricity price as the strategy space, and subject to the power grid security constraints.

[0032] The electricity price quote is a multi-agent non-cooperative game model with a strategy space and subject to the power grid security constraints.

[0033] As a preferred implementation, representation learning and equilibrium decision-making are performed based on the multi-agent non-cooperative game model to obtain a Nash equilibrium strategy that satisfies global constraints, including:

[0034] Define the game state of each market participant at the decision-making moment, wherein the game state includes at least the optimized node relationship characteristics, real-time market environment data and power grid operation constraints;

[0035] Based on the deep dual-simulation metric framework, low-dimensional representation learning is performed on the game state. A compact latent representation space is constructed by measuring the behavioral similarity between states, so as to focus on the core information related to game decision-making.

[0036] Design differentiated revenue functions based on the type of market entity, and introduce global constraints to ensure the safe and stable operation of the system;

[0037] By improving the gradient ascent method to perform collaborative optimization of the payoff function, and iteratively updating the strategy parameters of each market participant type, a Nash equilibrium strategy that satisfies global constraints is obtained.

[0038] As a preferred implementation, the Nash equilibrium strategy is stored and automatically executed via a blockchain smart contract, and the entire decision-making process is recorded and traced, including:

[0039] The Nash equilibrium strategy is automatically executed through the deployed smart contract to calculate and output the equilibrium price;

[0040] The equilibrium price is verified by a preset consensus algorithm, and the verified equilibrium price is automatically matched according to a preset transaction matching logic.

[0041] If the transaction is successfully matched, a legally binding electronic contract will be generated.

[0042] The hash value, balanced quote, electronic contract content, and execution status are stored on the blockchain.

[0043] In a preferred embodiment, the method further includes:

[0044] The strategy iteration trajectory of each market participant in the game decision-making process is recorded and stored on the blockchain;

[0045] Based on preset abnormal behavior judgment rules, the iterative trajectory of the strategy stored on the blockchain is monitored and analyzed;

[0046] When the monitoring and analysis results meet the abnormal conditions, the penalty clauses defined in the smart contract deployed on the blockchain are automatically triggered.

[0047] The abnormal conditions include: the change in strategy parameters of any market participant in adjacent decision-making rounds exceeds a preset mutation threshold.

[0048] Secondly, embodiments of the present invention also provide a multi-agent intelligent decision-making system for the electricity market based on graph neural networks and blockchain, comprising:

[0049] The initial relationship graph construction module is used to define the relationship between market entities based on the types of market entities stored in the blockchain, and to construct an initial undirected weighted relationship graph with each market entity as a node and the relationship between each market entity as an edge.

[0050] The node relationship feature optimization module is used to input the initial undirected weighted relationship graph into a graph neural network model adapted to the multi-subject features of the power market, and to perform dimensional attention weighted aggregation and global-local feature fusion through the graph neural network model to output optimized node relationship features. The graph neural network model includes a dimensional attention weighted aggregation module and a global-local feature fusion module.

[0051] The game model construction module is used to construct a multi-agent non-cooperative game model based on the optimized node relationship characteristics and combined with real-time market parameters and power grid security constraints.

[0052] The Nash equilibrium strategy solving module is used to perform representation learning and equilibrium decision solving based on the multi-agent non-cooperative game model to obtain the Nash equilibrium strategy that satisfies global constraints.

[0053] The strategy automation execution module is used to execute the Nash equilibrium strategy automatically by storing and verifying it through a blockchain smart contract, and to record and trace the entire decision-making process.

[0054] Thirdly, embodiments of the present invention also provide an electronic device, the electronic device comprising:

[0055] One or more processors;

[0056] Storage device for storing one or more programs;

[0057] When the one or more programs are executed by the one or more processors, the one or more processors implement the multi-agent intelligent decision-making method for the electricity market based on graph neural networks and blockchain as described in any embodiment of the present invention.

[0058] Fourthly, embodiments of the present invention also provide a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the multi-agent intelligent decision-making method for the electricity market based on graph neural networks and blockchain as described in any embodiment of the present invention.

[0059] Compared with the prior art, the present invention achieves the following beneficial effects:

[0060] (1) The graph neural network model proposed in this invention includes a dimension attention weighted aggregation module and a global-local feature fusion module. By fusing spatial, temporal and preference features and combining the global-local feature fusion mechanism, it captures the local correlation preferences between subjects and reflects the influence of the global market environment. It solves the problem that local relationships cannot reflect the global influence. The accuracy of the relationship feature characterization is improved by more than 30% compared with traditional methods.

[0061] (2) The Dimensional Attention Weighted Aggregation Mechanism (DGAT) adopted in this invention accurately distinguishes the semantic importance of features in different dimensions by allocating weights in each dimension. Compared with the traditional scalar attention mechanism, the recognition of features after aggregation is significantly improved, providing more effective input basis for subsequent game decision-making.

[0062] (3) The representation learning based on the DBC framework of this invention can filter out irrelevant interference in high-dimensional states and retain the core information of the game. Combined with the non-cooperative game + global constraint framework, the equilibrium decision scheme solved satisfies the demand for maximizing the benefits of each subject and strictly abides by the power grid safety constraints. The feasibility and stability of the equilibrium solution are improved by 40% compared with the traditional algorithm.

[0063] (4) This invention uses blockchain smart contracts to realize the on-chain storage of game parameters, automatic execution of decisions, and traceability and supervision of results, ensuring that game parameters are tamper-proof and the decision-making process is transparent and auditable, effectively preventing malicious manipulation of parameters, transaction repudiation and other behaviors, and improving the fairness and regulatory efficiency of electricity market transactions.

[0064] (5) The technical solution of the present invention can be compatible with the characteristics of multi-entity, high-frequency and decentralized transactions in the distributed power market, and can be flexibly adapted to the transaction needs of different types of distributed power sources and different users. It has a wide range of application scenarios and promotion value. Attached Figure Description

[0065] Other features, objects, and advantages of the invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings. The drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings:

[0066] Figure 1This is a flowchart of a multi-stakeholder intelligent decision-making method for the electricity market based on graph neural networks and blockchain, provided in an embodiment of the present invention.

[0067] Figure 2 This is a framework diagram of the multi-agent intelligent decision-making method for the electricity market based on graph neural networks and blockchain provided in this embodiment of the invention;

[0068] Figure 3 This is a distributed transaction deployment architecture diagram of the multi-agent intelligent decision-making method for the power market based on graph neural networks and blockchain provided in this embodiment of the invention;

[0069] Figure 4 This is the block information query interface provided in the embodiments of the present invention;

[0070] Figure 5 This is a schematic diagram of the structure of a multi-agent intelligent decision-making system for the electricity market based on graph neural networks and blockchain, provided in an embodiment of the present invention.

[0071] Figure 6 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0072] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and not intended to limit it. Furthermore, it should be noted that, for ease of description, the accompanying drawings show only the parts relevant to the present invention, and not all of the structures.

[0073] Before discussing the exemplary embodiments in more detail, it should be mentioned that some exemplary embodiments are described as processes or methods depicted as flowcharts. Although the flowcharts describe the operations (or steps) as sequential processes, many of the operations (or steps) can be performed in parallel, concurrently, or simultaneously. Furthermore, the order of the operations can be rearranged. The process can be terminated when its operation is completed, but it may also have additional steps not included in the figures. The process may correspond to a method, function, procedure, subroutine, subroutine, etc.

[0074] Example 1

[0075] like Figure 1 The diagram shows a flowchart of a multi-agent intelligent decision-making method 100 for the electricity market based on graph neural networks and blockchain, as provided in Embodiment 1 of the present invention. The method 100 specifically includes the following steps:

[0076] S110. Based on the types of market entities stored in the blockchain, define the relationships between market entities, and construct an initial undirected weighted relationship graph with each market entity as a node and the relationships between market entities as edges.

[0077] In some embodiments, the core types of entities in the blockchain-based distributed power market are first identified, including distributed generation project entities (including new energy entities such as photovoltaic and wind power), power users (industrial users, commercial users, and residential users), power grid companies (including incremental distribution network operators), power trading institutions, and regulatory bodies.

[0078] The relationships between market participants are defined based on the types of core entities, such as transaction cooperation relationships (supply and demand relationships under models such as direct transactions, agency transactions, and full acquisition), transmission service relationships (power transmission relationships between power grid companies and power generation and consumption entities), and regulatory constraint relationships (compliance supervision relationships between regulatory bodies and other market participants). The data on entity attributes (power generation capacity, power load, transmission capacity, credit rating), historical transaction data (past quotations, transaction volume, settlement price), and market environment data (grid constraints, fluctuations in new energy output, policy parameters) are integrated.

[0079] Based on the above-mentioned market entities as nodes and the relationships between them as edges, an initial undirected weighted relation graph is constructed, including:

[0080] Obtain the historical interaction frequency between market entities recorded in the blockchain distributed ledger, and assign corresponding relationship type weights according to the relationships between market entities; weight and merge the historical interaction frequency and relationship type weights to obtain the edge weights of the relationship graph; construct an initial undirected weighted relationship graph with each market entity as a node and the relationships between market entities as edges, combined with the edge weights of the relationship graph.

[0081] Specifically, the initial relational graph is defined as follows: ,in:

[0082] For a collection of market entities, each node For a given market entity, its attribute vector is: ; The total number of market entities, i.e., the total number of nodes in the relationship diagram; :main body The attribute vector contains the core feature information of the subject and is used to distinguish the characteristics of different subjects. Attribute vector The value of a single attribute dimension in; Attribute vector The number of dimensions, that is, the number of attribute feature types contained in each subject;

[0083] For a set of inter-subject relationships, if the subject and If there is a transaction, transmission, or regulatory relationship, then ; Main node and The node pairs formed;

[0084] This is the initial edge weight matrix. The strength of the relationship is quantified, and the calculation formula is as follows:

[0085]

[0086] in, Edge weight matrix Corresponding node pairs The initial weight values ​​reflect the degree of initial correlation between the two. : Normalization coefficient, used to adjust the contribution ratio of historical interaction frequency in edge weight calculation; : Normalization coefficient, used to adjust the contribution ratio of relation type weights in edge weight calculation, and satisfies . Main node and The frequency of historical interactions between them is calculated based on statistics of past transactions, service collaborations, and other records stored on the blockchain; Main node and The weights for the relationship types are assigned based on the relationship dimension: 1.0 for transaction cooperation relationships, 0.8 for power transmission service relationships, and 0.6 for regulatory constraint relationships.

[0087] Furthermore, to ensure that the data source for constructing the relationship graph is immutable, it is necessary to hash and encrypt the static attributes and dynamic interaction data of each subject based on the distributed ledger characteristics of blockchain, and then store them on the blockchain for evidence. At the same time, the authenticity of the data should be verified through the consensus mechanism of blockchain nodes to avoid the relationship modeling deviation caused by false data, thereby providing a reliable data foundation for subsequent feature optimization and game decision-making.

[0088] S120. Input the initial undirected weighted relationship graph into a graph neural network model adapted to the multi-subject characteristics of the power market, and perform dimensional attention weighted aggregation and global-local feature fusion through the graph neural network model to output optimized node relationship features. The graph neural network model includes a dimensional attention weighted aggregation module and a global-local feature fusion module.

[0089] In some embodiments, the graph neural network model adapted to the multi-subject characteristics of the electricity market adopts the STP-UDGAT (Spatial-Temporal-Preference User Dimensional Graph Attention Network) graph neural network model to perform feature depth optimization on the initial undirected weighted relationship in order to solve the technical problem that the traditional model can only capture local relationships and cannot reflect the global market impact, including the following steps (1)-(4):

[0090] Step (1): Extract spatial dimension features, temporal dimension features, and preference dimension features from the node pairs in the initial undirected weighted relation graph. Specifically:

[0091] (a) Spatial dimensional characteristics :

[0092] ,topo ,

[0093] in, Main node and The spatial dimension feature vectors characterize the geographical distribution and power grid topology relationship between the two. Main node and Geographic distance quantification value; Euclid Euclidean distance calculation function, used to calculate the main body. geographic coordinates With the main body geographic coordinates The straight-line distance between them; Main node and The power grid topology correlation coefficient quantifies the connection relationship between the two in the power grid (1 for direct connection, 0.5 for indirect connection, and 0 for no connection).

[0094] (b) Time dimension characteristics :

[0095] period ,

[0096] in, Main node and The time dimension feature vector reflects the correlation and time period matching degree between the two transaction times; Main node and The transaction time synchronization metric is derived from the average of the differences in timestamps of all transactions. The larger the value, the more synchronized the transaction times are. : Average value calculation function, used to calculate the main body No. Next transaction timestamp With the main body No. Next transaction timestamp The average level of the absolute values ​​of the differences; period Main node and The peak and valley time matching degree is used to quantify the consistency of the two trading periods (1 for the same peak period, 0.8 for the same valley period, and 0.5 for the flat period).

[0097] (c) Preference Dimension Features :

[0098]

[0099] in, Main node and The feature value of the preference dimension is obtained by the cosine similarity between the two preference vectors. The larger the value, the more compatible the preferences are. Cosine similarity calculation function, used to calculate the similarity of the main body. Preference vector With the main body Preference vector The degree of similarity between them.

[0100] Step (2): Concatenate the spatial dimension features, temporal dimension features and preference dimension features to generate an initial associated feature vector.

[0101] Specifically, after concatenating and fusing spatial, temporal, and preference features, the resulting initial inter-subject association feature vector is as follows:

[0102]

[0103] in, Main node and The fused initial correlation feature vector integrates spatial, temporal, and preference-based features. 3D real space, The space of values; : Fusion feature vectors The number of dimensions, here (2-dimensional spatial dimension + 2-dimensional time dimension + 1-dimensional preference dimension, which are concatenated into 6 dimensions).

[0104] Step (3): Calculate the attention coefficients for the initial associated feature vector dimension by dimension according to the dimensional attention weighted aggregation module, specifically including the following steps (3.1)-(3.3):

[0105] Step (3.1): Perform linear projection on the initial associated feature vector and the attribute features of the corresponding nodes to convert them into hidden layer feature vectors adapted for attention calculation. Specifically:

[0106] The initial associated feature vector and the attribute features of the corresponding nodes are linearly projected using the following formula:

[0107] ,

[0108] in, Main node Hidden layer feature vectors after linear projection; : The projection weight matrix of node attributes, used to project the original attribute vector Mapped to the hidden layer dimension, the dimension is ( For the hidden layer dimension, (For the original attribute dimensions). The bias term for node attribute projection is used to adjust the feature distribution after projection, with a dimension of . ; Main node and Hidden layer feature vectors after linear projection; : Projection weight matrix of associated features, used to project the initial associated feature vector Mapped to the hidden layer dimension, the dimension is ; Main node and The initial associated feature vector after fusion; The bias term for the projection of associated features is used to adjust the distribution of the projected associated features, with a dimension of . ; Hidden layer dimension, i.e., the dimension of the projected feature vector, is used to balance the model's expressive power and computational complexity.

[0109] Step (3.2): Based on the hidden layer feature vector, calculate the feature semantic correlation degree between node pairs dimension by dimension to obtain the dimensional attention coefficient vector, and normalize the dimensional attention coefficient vector.

[0110] The following formulas are used to calculate the feature semantic correlation degree between node pairs dimension by dimension, obtain the dimensional attention coefficient vector, and normalize the dimensional attention coefficient vector:

[0111]

[0112] in, Main node For neighboring nodes The attention coefficient vector is a dimensionless vector that quantifies the importance of neighbor features dimension by dimension. ; : Linear unit activation function with leakage correction, used to introduce nonlinear features, alleviate the gradient vanishing problem, and retain some negative gradient information; Attention weight parameter matrix, dimension 1 (adaptation (Feature dimensions after concatenation) Neighbor nodes The hidden layer feature vector after linear projection, and Combine this with the calculation of the attention coefficient. Vector concatenation operation, used to concatenate vectors... The three vectors are concatenated along their dimensions to form the input features for attention computation; The bias term for attention calculation, with dimension 1. ; The exponential function is used to convert the attention coefficient into a non-negative value, which facilitates subsequent normalization calculations. : For nodes All neighboring nodes The summation of the attention coefficient exponents is used as the normalization denominator to ensure that the sum of the attention coefficients is 1. Main node The set of neighbors, that is, with A set of entity nodes that are related (transactions, power transmission, regulation).

[0113] Step (3.3): Aggregate the normalized dimensional attention coefficient vector with the hidden layer feature vector of the corresponding neighbor node through Hadamard product to generate local node features.

[0114] Specifically, dimension-weighted aggregation is achieved through Hadamard product to generate local node features, using the following calculation formula:

[0115]

[0116] in, Main node The local feature vector after aggregating neighbor features is combined with important information from neighbor nodes; : For nodes All neighboring nodes Weighted feature summation is used to achieve local feature aggregation; Hadamard product (element-wise product) is used to combine the attention coefficient vector. and neighbor feature vectors Weighting is applied dimension by dimension to highlight the features of important dimensions.

[0117] Step (4): Based on the global-local feature fusion module, fuse the local feature vector with the global market features, and output the optimized node relationship feature vector.

[0118] First, global market characteristics are calculated based on the hidden layer feature vectors of each market entity type;

[0119] Next, linear projection and Sigmoid activation are performed on the hidden layer feature vectors of each market entity type to obtain global attention weights used to adjust the fusion ratio;

[0120] Finally, based on the local feature vectors, global market features, and global attention weights, an optimized node relationship feature vector is generated.

[0121] Specifically, a global attention module is introduced to capture global market impact and define global market characteristics. Through global attention weights By integrating local feature vectors and global market features, the optimized node relationship feature vector is finally obtained. Represented as:

[0122]

[0123] in, For all principal nodes to Projected attribute feature vector Summation provides the basis for calculating the global influence factor; Main node The global attention weight is used to adjust the fusion ratio between local and global features, and its value ranges from [value range missing]. ; The Sigmoid activation function is used to map linearly transformed features to... The interval is used to output the global attention weights. Main node The local feature vector, after aggregating neighbor features, reflects the local association information between the node and its surrounding entities.

[0124] Through the optimized node relationship feature vector Construct the optimized relationship diagram :

[0125]

[0126] in, The optimized blockchain electricity market multi-entity relationship graph includes a set of entities, a set of relationships, and an optimized feature matrix, which depicts more accurate entity relationship characteristics. The optimized node relationship feature matrix is ​​composed of the final relationship feature vectors of all main nodes, with a dimension of [missing information]. ; : Feature matrix The construction form of the single node feature vector to The rows are stacked to form a matrix.

[0127] Similarly, the optimized node relationship feature vector described above... After being encrypted with blockchain, the data is stored on the blockchain and serves as the core input for subsequent intelligent trading.

[0128] S130. Based on the optimized node relationship characteristics, and combined with real-time market parameters and power grid security constraints, a multi-agent non-cooperative game model is constructed.

[0129] In some embodiments, based on the optimized node relationship characteristics output in step S120 above, and combined with the real-time changing power market environment and grid operating conditions, a smart game model that closely resembles the actual market with "non-cooperative game + global constraints" is constructed to achieve equilibrium decision-making solutions for multiple stakeholders in the distributed power market. The key parameters of the smart game model are stored on the blockchain to ensure fairness and transparency at the starting point of the game.

[0130] Specifically, the optimized node relationship characteristics are used as the static relationship representation of each market participant; the real-time market parameters and grid security constraints are used as dynamic environment inputs; based on the static relationship representations and dynamic environment inputs, a multi-participant non-cooperative game model is constructed, with each market participant type as the game participants, the declared electricity volume and declared electricity price as the strategy space, and subject to the grid security constraints.

[0131] Among these measures, storing the key parameters of the intelligent game model on the blockchain includes, but is not limited to, storing the game model parameters (such as the initial graph) as evidence. Weight matrix and feature dimensions ), attributes of the game participants ( ), market rules (grid constraints) Deviation penalty coefficient The hash values ​​of key parameters are stored on the blockchain, and the hash calculation uses the SHA-256 algorithm.

[0132]

[0133] hash(param): The hash value of the key game parameter, calculated using the SHA-256 algorithm, used for on-chain storage to ensure that the parameter cannot be tampered with; The game parameters are JSON formatted strings, which serialize various parameters in a unified format to facilitate hash calculation and on-chain storage, ensure that the parameters cannot be tampered with, and provide a basis for tracing the game results.

[0134] S140. Based on the multi-agent non-cooperative game model, perform representation learning and equilibrium decision-making to obtain a Nash equilibrium strategy that satisfies global constraints.

[0135] The representation learning and equilibrium decision-making solution for the above multi-agent non-cooperative game model includes the following steps (4)-(8):

[0136] Step (4): Define the game state of each market participant at the decision-making moment. The game state, which includes at least the optimized node relationship characteristics, real-time market environment data, and power grid operation constraints, can be expressed as:

[0137]

[0138] in, Moment Subject The game state vector integrates the characteristics of the main parties' relationships, real-time market parameters, and power grid constraints to comprehensively depict the dynamic situation of the game participants. : Optimized node relationship feature vector; The moment-in-time electricity market parameter vector contains key market operation data such as the market electricity price, total electricity load, and total output of new energy sources at that moment. The constant-time power grid constraint parameter vector covers constraints that ensure the safe operation of the power grid, such as the upper limit of line transmission capacity and voltage and frequency stability thresholds.

[0139] Step (5): Based on the deep dual simulation metric framework, perform low-dimensional representation learning on the game state, and construct a compact latent representation space by measuring the behavioral similarity between states in order to focus on the core information related to game decision-making.

[0140] Specifically, a dual-simulation metric is defined to represent the state. and Dual analog metrics Defined as:

[0141]

[0142] in, Moment Subject and The dual simulation measure of game states quantifies the similarity between two states in terms of behavioral decisions; the smaller the value, the higher the behavioral equivalence. : All possible action space of the subject Take the maximum value and filter out the actions that have the most significant impact on the difference between the two states; The weighting coefficients for the dual analog metric have a range of values. This is used to balance the contribution ratio between differences in immediate rewards and differences in future state transitions; The scope of action for market participants includes all the game-theoretic decisions that participants can make, such as adjusting prices, submitting electricity volume declarations, and choosing trading models. :main body In state Next action The reward value obtained is defined according to the subject type; The Wasserstein distance measures the distance between two state transition distributions, reflecting the similarity of their future state evolutions. Basic metric function; The subject is in a state Next action Then, transition to the next state. The probability distribution characterizes the randomness of state evolution; The next game state after the subject performs an action is the output result of the state transition distribution.

[0143] Furthermore, we define the representation learning objective and learn the encoder function. , making latent space The distance is equal to the above two-simulation metric, and its loss function is:

[0144]

[0145]

[0146] in, The encoder function for the game state, which encodes the high-dimensional game state. Mapping to a low-dimensional latent space, the output is a compact state representation. ; Encoder The training loss function is used to optimize encoder parameters so that the latent spatial distance is consistent with the dual analog metric. Mathematical expectation operator, for all possible game states Take the expected value to reduce the loss fluctuation caused by random samples; Subject in latent space and State representation Distance, used to approximate a two-analog metric; :main body In state Next action The reward value obtained, i.e. ; :main body In state Next action The reward value obtained, i.e. ; Discount factor, with a value range of: ; The learned probabilistic dynamic model outputs a Gaussian distribution to approximate the true state transition distribution. This reduces model complexity; Based on latent state representation and actions The predicted next state transition distribution is derived from the probabilistic dynamic model. Output.

[0147] Step (7): Design differentiated revenue functions based on the type of market participants, and introduce global constraints to ensure the safe and stable operation of the system:

[0148] Specifically, step (7.1): Define the power generation entity. payoff function

[0149]

[0150] Power generation main body The revenue function is used to calculate its reported electricity volume. and the declared electricity price Net income below; Power generation main body The power generation cost function describes its production Total cost of electricity (including fuel, operation and maintenance costs, etc.); Power generation main body The predicted output is the expected power generation derived from data such as new energy output forecasts and unit operating status. The power deviation penalty coefficient for power generation entities is used to penalize behaviors where the declared power output is inconsistent with the actual output, thereby urging entities to make accurate declarations. Electricity users The revenue function (net utility function) is used to calculate its value in the declared electricity volume. and transaction electricity price Net income below.

[0151] Step (7.2): Define electricity users

[0152]

[0153] in, Electricity users The power efficiency function describes its consumption. The utility value of electricity (such as industrial production benefits and the quantification of residential electricity convenience). Electricity users The predicted electricity consumption is the expected electricity consumption derived from historical electricity consumption data, production plans, etc. The power shortage (or over-purchase) penalty coefficient for electricity users is used to punish behaviors that are inconsistent with the declared power consumption and the actual power consumption.

[0154] Step (7.3): Design global constraints, including power balance constraints and grid security constraints. The power balance constraints are as follows:

[0155]

[0156] in, The total electricity volume declared by all power generation entities, i.e., the total electricity supply in the market; The collection of all power generation entities in the electricity market, including various power generation participants such as distributed photovoltaic, wind power, and small hydropower; The total electricity demand reported by all electricity users, i.e., the total market demand for electricity; The electricity market is a collection of all electricity users, including industrial, commercial, and residential electricity users.

[0157] The power grid security constraints are as follows:

[0158]

[0159] in, The minimum allowable power flow value of a power line, that is, the minimum power transmission limit for safe operation of the line (to avoid problems such as no-load overvoltage). The maximum permissible power flow value of a power line, i.e., the maximum power transmission limit for the safe operation of the line (to avoid problems such as line overload and burnout). The actual power flow value of the power lines is determined by the electricity volume declared by each power generation entity. User-reported electricity consumption The data is calculated based on the power grid topology parameters, reflecting the actual power transmission of the lines.

[0160] Step (8): The payoff function is optimized by improving the gradient ascent method, and the strategy parameters of each market participant type are updated iteratively to obtain a Nash equilibrium strategy that satisfies global constraints.

[0161] Specifically, when solving for the Nash equilibrium policy using the improved gradient ascent method, the principal policy update formula is as follows:

[0162] in, Power generation main body The strategy parameters are the iteratively updated combination of declared electricity volume and declared electricity price; Power generation main body The current strategy parameters, i.e., the current combination of declared electricity volume and declared electricity price, are denoted as... ; The learning rate of policy iteration is used to control the step size of each policy update, balancing convergence speed and stability. Power generation main body The payoff function with respect to the current policy parameters The gradient indicates the direction of strategy adjustments to maximize returns; Except for the main power generation unit The set of strategies of all other market participants reflects the current decision-making status of the opposing players; : Constraint violation penalty coefficient, used to strengthen the enforcement of global constraints and impose penalties on strategies that violate power balance and grid security constraints; Constraint violation penalty term on current policy parameters The gradient indicates the direction of policy adjustments to reduce the degree of constraint violation; Penalty Constraint violation penalty: When the main strategy violates global constraints such as power balance and grid security, the corresponding penalty value is output to reduce the main strategy's violation benefits.

[0163] In addition, the convergence theorem of the DBC framework for Nash equilibrium strategy iteration must satisfy:

[0164]

[0165] in, The convergence criterion for strategy iteration, i.e., the convergence of the strategy with the number of iterations. The difference between the policy parameters in the two consecutive iterations approaches 0, indicating that the policy has converged. Time and The distance between policy parameters at different times (such as Euclidean distance) is used to measure the magnitude of policy adjustments.

[0166] When the Nash equilibrium strategy iteratively converges to the Nash equilibrium point and the equilibrium solution satisfies the global constraints, the feasibility and stability of the game outcome are ensured.

[0167] S150. The Nash equilibrium strategy is stored and automatically executed through a blockchain smart contract, and the entire decision-making process is recorded and traced.

[0168] In some embodiments, the Nash equilibrium strategy is automatically executed through a deployed smart contract to calculate and output the equilibrium quote, specifically...

[0169] Before each daily transaction report, the contract automatically reads the on-chain relationship feature matrix. With market parameters (Including real-time electricity prices, total load, and renewable energy output data), the contract uses an improved gradient ascent algorithm built into it to input the initial bids from each entity. Through iterative formula Penalty calculation of equilibrium pricing:

[0170]

[0171] in, :main body The initial pricing strategy parameters are denoted as... ,in This is the initial declared electricity volume. For the initial declared electricity price; :main body The initial declared electricity volume, that is, the initial transaction electricity volume submitted before the start of the game calculation; :main body The initial declared electricity price, i.e. the initial transaction price submitted before the start of the game calculation; : No. Main body after round of iteration The strategy parameters are the updated bid combination obtained by improving the gradient ascent algorithm; : No. Main body during round iteration The current strategy parameters serve as the basis for iterative updates; :main body The gradient of the payoff function indicates the direction in which the policy parameters are adjusted to maximize the payoff; Penalty: The gradient of the constraint violation penalty term, indicating the direction of policy parameter adjustment to reduce the degree of constraint violation; The set of equilibrium bidding strategies derived from game theory calculations, containing the optimal bids of all parties. This serves as the basis for transaction execution; :main body The equilibrium pricing strategy parameters are the optimal combination of declared electricity volume and declared electricity price after the game converges.

[0172] Next, the equilibrium price is verified using a preset consensus algorithm, and the verified equilibrium price is automatically matched according to a preset transaction matching logic.

[0173] The balanced price, after reaching a consensus among blockchain nodes (using the Kafka consensus algorithm), is made public to all parties. The consensus verification conditions are as follows:

[0174] count verify block

[0175] Where, count verify block The number of nodes in the blockchain that have verified the current block (including the equilibrium price) is used to determine whether the consensus condition is met; verify(block): The verification function of the blockchain node is used to verify the compliance of the game calculation process in the block (such as the correctness of algorithm execution and the fact that the parameters have not been tampered with). It returns True if the verification is successful, and False otherwise; block: A blockchain block containing the equilibrium price and the record of the game calculation process. It is the basic unit of consensus and data storage. The total number of nodes in the blockchain network that participate in the block verification and consensus process; The minimum number of nodes required for consensus verification is the threshold. Only when the number of nodes that have passed verification reaches or exceeds this threshold will the equilibrium offer be deemed valid and made public.

[0176] Combination Figure 3 As shown, blockchain nodes are deployed via Docker on servers belonging to different organizations, forming a blockchain network. Nodes communicate via a P2P protocol and achieve data consistency through a consensus mechanism. Each blockchain node possesses complete ledger information, and the node entity can be a power trading center, a participant in the distributed power trading market, or other social entities. A cloud-based deployment method is used, with 5-7 nodes deployed. The block information query interface is shown below. Figure 4 As shown, clicking on a re-node allows you to view the block information generated by that re-node.

[0177] If the transaction is successfully matched, the entity will use the equilibrium price. Upon submitting an order, the contract is automatically matched and executed, generating an electronic contract and uploading it to the blockchain. The matching rules are as follows:

[0178]

[0179] in, :main body (Power Generator) and Main Body The transaction matching function for the electricity user returns 1 if the match is successful and 0 if the match fails. Power generation main body The equilibrium declared electricity price (electricity sales price) is the optimal electricity sales price after game convergence; Electricity users The equilibrium bid price (electricity purchase price) is the optimal electricity purchase price after game convergence; Power generation main body The equilibrium declared electricity volume (electricity sales volume) is the optimal electricity sales volume after game convergence; Electricity users The balanced declared electricity volume (purchased electricity volume) is the optimal electricity purchase scale after game convergence.

[0180] In some embodiments, the method further includes: monitoring and analyzing the strategy iteration trajectory stored on the blockchain based on preset abnormal behavior judgment rules; automatically triggering penalty clauses defined in the smart contract deployed on the blockchain when the monitoring and analysis results meet abnormal conditions; wherein, the abnormal conditions include: the change in strategy parameters of any market participant in adjacent decision rounds exceeds a preset mutation threshold.

[0181] Specifically, regulatory agencies can query the calculation records of each step of the game process (such as the strategy iteration trajectory) through blockchain nodes. If an anomaly is detected, the penalty clauses in the contract will be triggered (such as deducting credit points or restricting transaction permissions), and the violation record will be permanently recorded on the blockchain. The expression for the anomaly trigger condition is as follows:

[0182] Make

[0183] in, :main body The absolute difference between the strategy parameters of two adjacent rounds is used to detect whether the strategy changes abruptly. The strategy mutation threshold is a critical value for determining whether an entity is maliciously manipulating a strategy; exceeding this threshold is considered abnormal. There exists a certain subject and a certain iteration , is used to describe the triggering conditions for anomaly detection (i.e., the existence of a policy mutation in a subject during a certain iteration).

[0184] Based on the above embodiments, the present invention achieves the following beneficial effects:

[0185] (1) The graph neural network model proposed in this invention includes a dimension attention weighted aggregation module and a global-local feature fusion module. By fusing spatial, temporal and preference features and combining the global-local feature fusion mechanism, it captures the local correlation preferences between subjects and reflects the influence of the global market environment. It solves the problem that local relationships cannot reflect the global influence. The accuracy of the relationship feature characterization is improved by more than 30% compared with traditional methods.

[0186] (2) The Dimensional Attention (DGAT) mechanism adopted in this invention accurately distinguishes the semantic importance of features in different dimensions by assigning weights to each dimension. Compared with the traditional scalar attention mechanism, the recognition of features after aggregation is significantly improved, providing more effective input basis for subsequent game decision-making.

[0187] (3) The representation learning based on the DBC framework of this invention can filter out irrelevant interference in high-dimensional states and retain the core information of the game. Combined with the non-cooperative game + global constraint framework, the equilibrium decision scheme solved satisfies the demand for maximizing the benefits of each subject and strictly abides by the power grid safety constraints. The feasibility and stability of the equilibrium solution are improved by 40% compared with the traditional algorithm.

[0188] (4) This invention uses blockchain smart contracts to realize the on-chain storage of game parameters, automatic execution of decisions, and traceability and supervision of results, ensuring that game parameters are tamper-proof and the decision-making process is transparent and auditable, effectively preventing malicious manipulation of parameters, transaction repudiation and other behaviors, and improving the fairness and regulatory efficiency of electricity market transactions.

[0189] (5) The technical solution of the present invention can be compatible with the characteristics of multi-entity, high-frequency and decentralized transactions in the distributed power market, and can be flexibly adapted to the transaction needs of different types of distributed power sources (photovoltaic, wind power, etc.) and different users (industrial, residential, etc.), and has a wide range of application scenarios and promotion value.

[0190] Example 2

[0191] Figure 5 This is a schematic diagram of the structure of a multi-agent intelligent decision-making system for the power market based on graph neural networks and blockchain, as provided in Embodiment 2 of the present invention. Figure 5 As shown, the system includes:

[0192] The initial relationship graph construction module 510 is used to obtain the types of market entities and the relationships between the types of market entities in the blockchain distributed ledger notarization, and to construct an initial undirected weighted relationship graph with the types of market entities as nodes and the relationships between the types of market entities as edges.

[0193] The node relationship feature optimization module 520 is used to input the initial undirected weighted relationship graph into a graph neural network model adapted to the multi-subject features of the power market, and to perform dimensional attention weighted aggregation and global-local feature fusion through the graph neural network model to output optimized node relationship features. The graph neural network model includes a dimensional attention weighted aggregation module and a global-local feature fusion module.

[0194] The game model construction module 530 is used to construct a multi-agent non-cooperative game model based on the optimized node relationship characteristics and combined with real-time market parameters and power grid security constraints.

[0195] The Nash equilibrium strategy solving module 540 is used to perform representation learning and equilibrium decision solving based on the multi-agent non-cooperative game model to obtain a Nash equilibrium strategy that satisfies global constraints.

[0196] The Nash equilibrium strategy automated execution module 550 is used to perform the Nash equilibrium strategy through blockchain smart contract for notarization and automated execution, and to record and trace the entire decision-making process.

[0197] The multi-agent intelligent decision-making system for the power market based on graph neural networks and blockchain provided in this embodiment of the invention can execute the multi-agent intelligent decision-making method for the power market based on graph neural networks and blockchain provided in any of the embodiments of the invention above. It has the corresponding functions and beneficial effects of executing the multi-agent intelligent decision-making method for the power market based on graph neural networks and blockchain. For detailed process, please refer to the relevant operations of the multi-agent intelligent decision-making method for the power market based on graph neural networks and blockchain in the foregoing embodiments.

[0198] Example 3

[0199] Figure 6 This is a schematic diagram of the structure of an electronic device provided in Embodiment 3 of the present invention. The electronic device 10 is intended to represent various forms of digital computers, and may also represent various forms of mobile devices. The components shown herein, their connections and relationships, and their functions are merely examples and are not intended to limit the implementation of the invention described and / or claimed herein.

[0200] like Figure 6As shown, the electronic device 10 includes at least one processor 11 and a memory, such as a read-only memory (ROM) 12 or a random access memory (RAM) 13, communicatively connected to the at least one processor 11. The memory stores computer programs executable by the at least one processor. The processor 11 can perform various appropriate actions and processes based on the computer program stored in the ROM 12 or loaded into the RAM 13 from storage unit 18. The RAM 13 can also store various programs and data required for the operation of the electronic device 10. The processor 11, ROM 12, and RAM 13 are interconnected via a bus 14. An input / output (I / O) interface 15 is also connected to the bus 14.

[0201] Multiple components in electronic device 10 are connected to I / O interface 15, including: input unit 16, such as keyboard, mouse, etc.; output unit 17, such as various types of displays, speakers, etc.; storage unit 18, such as disk, optical disk, etc.; and communication unit 19, such as network card, modem, wireless transceiver, etc. Communication unit 19 allows electronic device 10 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0202] Processor 11 can be a variety of general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of processor 11 include, but are not limited to, central processing unit (CPU), graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various processors running machine learning model algorithms, digital signal processors (DSPs), and any suitable processor, controller, microcontroller, etc. Processor 11 executes the dynamic prediction method for disaster-prone processes based on a multiphysics coupling model described above.

[0203] It should be understood that the various forms of processes shown above can be used, with steps reordered, added, or deleted. For example, the steps described in this invention can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution of this invention can be achieved, and this is not limited herein.

[0204] The above embodiments are merely illustrative examples and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations. However, obvious variations or modifications derived therefrom are still within the scope of protection of this invention.

Claims

1. A multi-stakeholder intelligent decision-making method for the electricity market based on graph neural networks and blockchain, characterized in that, include: Based on the types of market entities stored in the blockchain, the relationships between market entities are defined, and an initial undirected weighted relationship graph is constructed with each market entity as a node and the relationships between market entities as edges. The initial undirected weighted relationship graph is input into a graph neural network model adapted to the multi-entity characteristics of the electricity market. The graph neural network model performs dimensional attention weighted aggregation and global-local feature fusion to output optimized node relationship features. The graph neural network model includes a dimensional attention weighted aggregation module and a global-local feature fusion module. Based on the optimized node relationship characteristics, and combined with real-time market parameters and power grid security constraints, a multi-agent non-cooperative game model is constructed. Based on the multi-agent non-cooperative game model, representation learning and equilibrium decision-making are performed to obtain the Nash equilibrium strategy that satisfies global constraints. The Nash equilibrium strategy is stored and automatically executed through a blockchain smart contract, and the entire decision-making process is recorded and traced.

2. The method according to claim 1, characterized in that, An initial undirected weighted relation graph is constructed using the aforementioned market entities as nodes and the relationships between them as edges, including: Obtain the historical interaction frequency between various market entities recorded in the blockchain, and assign corresponding relationship type weights based on the relationships between the market entities; The historical interaction frequency and the relation type weight are weighted and fused together to obtain the edge weights of the relation graph; Using each market entity as a node and the relationships between them as edges, an initial undirected weighted relationship graph is constructed by combining the edge weights of the relationship graph.

3. The method according to claim 1, characterized in that, The initial undirected weighted relationship graph is input into a graph neural network model adapted to the multi-entity characteristics of the electricity market. The graph neural network model performs dimensional attention weighted aggregation and global-local feature fusion to output optimized node relationship features, including: Based on the node pairs in the initial undirected weighted relation graph, extract spatial dimension features, temporal dimension features, and preference dimension features; The spatial dimension features, temporal dimension features, and preference dimension features are concatenated to generate an initial associated feature vector; The dimension attention weighted aggregation module calculates the attention coefficients of the initial associated feature vector dimension by dimension and performs dimension attention coefficient weighted aggregation to generate local node features. The global-local feature fusion module fuses the local feature vector with the global market features to output an optimized node relationship feature vector.

4. The method according to claim 3, characterized in that, The dimension-based attention weighted aggregation module calculates attention coefficients dimension-by-dimensionally on the initial associated feature vector and performs dimensional attention coefficient weighted aggregation to generate local node features, including: The initial associated feature vector and the attribute features of the corresponding nodes are linearly projected to convert them into hidden layer feature vectors adapted for attention calculation. Based on the hidden layer feature vector, the feature semantic correlation degree between node pairs is calculated dimension by dimension to obtain the dimensional attention coefficient vector, and the dimensional attention coefficient vector is normalized. The normalized dimensional attention coefficient vector is aggregated with the hidden layer feature vectors of the corresponding neighboring nodes through the Hadamard product to generate local node features.

5. The method according to claim 4, characterized in that, The global-local feature fusion module fuses the local feature vectors with global market features to output an optimized node relationship feature vector, including: Calculate global market characteristics based on the hidden layer feature vectors of each market participant; Linear projection and Sigmoid activation are performed on the hidden layer feature vector of each market entity to obtain global attention weights used to adjust the fusion ratio; Based on the local feature vector, global market features, and global attention weight, an optimized node relationship feature vector is generated.

6. The method according to claim 1, characterized in that, Based on the optimized node relationship characteristics, and combined with real-time market parameters and power grid security constraints, a multi-agent non-cooperative game model is constructed, including: The optimized node relationship features are used as the static relationship representation of each market entity; The real-time market parameters and power grid security constraints are used as dynamic environment inputs. Based on the static relationship representation and dynamic environment input, a multi-agent non-cooperative game model is constructed, with each market entity as the game participant, the declared electricity volume and declared electricity price as the strategy space, and subject to the power grid security constraints.

7. The method according to claim 6, characterized in that, Based on the aforementioned multi-agent non-cooperative game model, representation learning and equilibrium decision-making are performed to obtain Nash equilibrium strategies that satisfy global constraints, including: Define the game state of each market participant at the decision-making moment, wherein the game state includes at least the optimized node relationship characteristics, real-time market environment data and power grid operation constraints; Based on the deep dual-simulation metric framework, low-dimensional representation learning is performed on the game state. A compact latent representation space is constructed by measuring the behavioral similarity between states, so as to focus on the core information related to game decision-making. Design differentiated revenue functions based on the type of market entity, and introduce global constraints to ensure the safe and stable operation of the system; By improving the gradient ascent method to perform collaborative optimization of the payoff function, and iteratively updating the strategy parameters of each market participant type, a Nash equilibrium strategy that satisfies global constraints is obtained.

8. The method according to claim 7, characterized in that, The Nash equilibrium strategy is stored and automatically executed via a blockchain smart contract, and the entire decision-making process is recorded and traced, including: The Nash equilibrium strategy is automatically executed through the deployed smart contract to calculate and output the equilibrium price; The equilibrium price is verified by a preset consensus algorithm, and the verified equilibrium price is automatically matched according to a preset transaction matching logic. If the transaction is successfully matched, a legally binding electronic contract will be generated. The balanced quote, electronic contract content, and execution status are stored on the blockchain.

9. The method according to claim 1, characterized in that, The method further includes: The strategy iteration trajectory of each market participant in the game decision-making process is recorded and stored on the blockchain; Based on preset abnormal behavior judgment rules, the iterative trajectory of the strategy stored on the blockchain is monitored and analyzed; When the monitoring and analysis results meet the abnormal conditions, the penalty clauses defined in the smart contract deployed on the blockchain are automatically triggered. The abnormal conditions include: the change in strategy parameters of any market participant in adjacent decision-making rounds exceeds a preset mutation threshold.

10. A multi-stakeholder intelligent decision-making system for the electricity market based on graph neural networks and blockchain, characterized in that, include: The initial relationship graph construction module is used to define the relationship between market entities based on the types of market entities stored in the blockchain, and to construct an initial undirected weighted relationship graph with each market entity as a node and the relationship between each market entity as an edge. The node relationship feature optimization module is used to input the initial undirected weighted relationship graph into a graph neural network model adapted to the multi-subject features of the power market, and to perform dimensional attention weighted aggregation and global-local feature fusion through the graph neural network model to output optimized node relationship features. The graph neural network model includes a dimensional attention weighted aggregation module and a global-local feature fusion module. The game model construction module is used to construct a multi-agent non-cooperative game model based on the optimized node relationship characteristics and combined with real-time market parameters and power grid security constraints. The Nash equilibrium strategy solving module is used to perform representation learning and equilibrium decision solving based on the multi-agent non-cooperative game model to obtain the Nash equilibrium strategy that satisfies global constraints. The strategy automation execution module is used to execute the Nash equilibrium strategy automatically by storing and verifying it through a blockchain smart contract, and to record and trace the entire decision-making process.

Images