A cross-platform power transaction data interaction optimization method

Abstract

The application discloses a cross-platform power transaction data interaction optimization method, and particularly relates to the technical field of power transaction data digital twinning, and generates simulation transaction data of market participants based on an adversarial neural network model, identifies the causal relationship between market variables from the simulation transaction data of the digital twin through a causal discovery algorithm, and constructs a causal diagram; the spatial dependency and causal conduction path between market participants are modeled by using a graph neural network model, counterfactual reasoning in a single external shock scenario is carried out in the digital twin by intervening in key variables in the causal diagram; the dynamic game behavior of the market participants under an extreme weather parameter distribution scenario is simulated; power market failure evaluation is carried out in combination with the causal conduction path, power market failure evaluation indexes are obtained, early warning is carried out based on the power market failure evaluation indexes, and the adaptability and stability of the power market under an extreme weather scenario are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of power transaction data digital twinning, more specifically, the present application relates to a cross-platform power transaction data interaction optimization method. BACKGROUND

[0002] As the core platform for energy distribution and transaction, the power market needs to maintain stable operation under extreme weather scenarios to ensure power supply reliability and price reasonableness. The power market regulatory department relies on data-driven risk assessment methods to predict market behavior and develop response strategies. However, market fluctuations caused by extreme weather events (such as droughts and storms) have high uncertainty, and real transaction data is often scarce, making it difficult to be directly used for analysis. Digital twinning technology simulates power market behavior through a virtual environment, providing a potential solution for assessing market failure risks. However, existing methods have limitations in data generation, causal modeling, and dynamic simulation, limiting their application in complex scenarios.

[0003] Existing power market risk assessment methods are limited by the scarcity of real transaction data, making it difficult to accurately simulate market behavior under extreme weather scenarios in a digital twinning environment, leading to difficulties in predicting market failure risks, including price volatility, supply gap probability, and transmission congestion rate. In addition, existing methods mostly use static analysis, lack modeling of dynamic game behavior of market participants, and risk indicators are not comprehensive enough to capture the systemic impact of extreme scenarios, limiting the management efficiency of the power market regulatory department. SUMMARY

[0004] To overcome the above-mentioned defects of the prior art, the present application provides a cross-platform power transaction data interaction optimization method, which realizes the optimization of cross-platform power transaction data interaction through the combination of digital twinning, causal reasoning, and reinforcement learning, to solve the problems raised in the background art.

[0005] To achieve the above-mentioned purposes, the present application provides the following technical solution: a cross-platform power transaction data interaction optimization method, comprising the following steps:

[0006] Step 1: Building a digital twin of the power market

[0007] Generating simulated transaction data of market participants based on an adversarial neural network model, including bids, power generation, and electricity demand, the generator of the adversarial neural network model generates simulated transaction data with a multi-dimensional random vector as input;

[0008] Step 2: Causal discovery and counterfactual inquiry

[0009] A causal graph is constructed by identifying causal relationships between market variables from simulation transaction data of the digital twin through a causal discovery algorithm; a graph neural network model is used to model spatial dependencies and causal transmission paths between market participants; counterfactual reasoning under a single external shock scenario is performed in the digital twin by intervening in key variables in the causal graph;

[0010] Step three: market failure risk assessment

[0011] A multi-agent reinforcement learning model is deployed in the digital twin, and an extreme weather parameter distribution scenario is generated by combining Monte Carlo simulation to simulate the dynamic game behavior of market participants under the extreme weather parameter distribution scenario; the power market failure assessment index is obtained by combining the causal transmission path to evaluate the power market failure; and early warning is performed based on the power market failure assessment index.

[0012] Preferably, in the construction of the digital twin, physical constraint terms are embedded in the training process of the adversarial neural network model, the physical constraint terms include DC power flow equations and dynamic line capacity constraints, the DC power flow equations are used to balance the node injection power and line power flow, and the generated simulation transaction data meets the node energy conservation condition in the actual operation of the power grid; the dynamic line capacity constraint function is introduced to compare the power value of each transmission line in the generated data with the maximum allowed capacity under its current working condition, and a penalty term is used to constrain the line power not to exceed the real-time capacity threshold; to ensure that the generated simulation transaction data meets the transmission capacity limit and node energy balance.

[0013] Preferably, the spatial dependencies and causal transmission paths between market participants are modeled using a graph neural network model, and counterfactual scenarios under external shocks are simulated in the digital twin by intervening in key variables in the causal graph to predict market price and supply and demand changes, including the following steps:

[0014] Constructing a causal graph: extracting market variables and performing standardization processing; searching for a causal graph structure through the Bayesian information criterion of the GES algorithm; embedding prior knowledge of power network topology to constrain the direction of causal edges; determining key variables through Granger causality test and N-1 safety criterion verification;

[0015] Building a graph structure based on a causal graph: market participants are nodes of the graph structure, and the trading relationship and geographical location relationship of market participants are edges of the graph structure; node features include historical prices, supply and demand situations, geographical locations, line impedances, and weather sensitivities;

[0016] Spatial dependencies and causal transmission paths between nodes are learned through an attention mechanism; spatial dependencies and causal transmission paths between nodes are learned;

[0017] By intervening in the key variables through the do-calculus rule, the counterfactual scenarios under external shocks are simulated in the digital twin to predict market price fluctuations, supply-demand balance and transmission congestion risks.

[0018] Preferably, the multi-agent reinforcement learning model comprises a graph attention network module and a strategy network of each agent.

[0019] The graph attention network module is configured to generate node embeddings of market participants to support dynamic game simulation, and the node embeddings capture dynamic interactions in the electricity market through a dynamic attention mechanism, thereby improving the quantitative accuracy of market failure risk assessment.

[0020] Each agent is configured with a strategy network, which is a deep neural network-based decision model. The strategy network analyzes simulated trading data to obtain the state of the agent, including its own bid, power generation or electricity demand. The strategy network takes node embeddings and agent states as input, and outputs action decisions of market participants, including at least one of adjusting the magnitude of the bid and changing the plan of power generation or electricity demand. The strategy network is trained through reinforcement learning based on simulated trading data and causal transmission paths in the digital twin, and is optimized based on the reward function of the agent to adapt to market dynamics under extreme weather parameter distribution scenarios.

[0021] Preferably, the multi-agent reinforcement learning model is obtained in the following manner:

[0022] Simulated trading data of the digital twin is obtained, and a graph structure is constructed based on a causal graph.

[0023] A graph attention network module is used to aggregate node neighbor information and capture spatial dependencies. The strategy network of each agent is input to guide action decisions. The action decisions of the agents are fed back to the graph neural network model to update the market state sequence.

[0024] The digital twin is used as a simulation environment to provide a dynamic market state sequence. A reward function is designed for each agent.

[0025] The parameters of the graph neural network model and the strategy network are initialized. In the digital twin, multiple rounds of simulation are run, and the agents take actions based on the market state sequence output by the graph neural network model to obtain rewards. Reinforcement learning algorithms are used to update the parameters of the strategy network, and the graph neural network model is optimized to improve the accuracy of state prediction. A competition and cooperation mechanism is introduced to balance the goals of multiple agents.

[0026] verify the consistency of the power market failure assessment indicators output by the multi-agent reinforcement learning model with historical data or actual trends; deploy the multi-agent reinforcement learning model to the digital twin to assess the risk of power market failure in real time.

[0027] Preferably, the operation process of market failure risk assessment includes the following steps:

[0028] Step 301: Construct a multi-agent environment, define each market participant as an agent, and all agents form an agent set; output the state, action and reward function of each agent;

[0029] Step 302: Dynamic game simulation: run the multi-agent reinforcement learning model in the digital twin, combine Monte Carlo simulation to generate extreme weather parameter distribution scenarios, simulate the dynamic game behavior of market participants under extreme weather parameter distribution scenarios, calculate node embedding through graph attention network in each round of loop, and adjust actions according to node embedding. Verify the physical feasibility of the market state sequence, update the strategy and record the market state sequence;

[0030] Step 303: Quantify the power market failure assessment indicators: based on the market state sequence set and the causal transmission path, calculate the power market failure assessment indicators, including price fluctuation amplitude, supply gap probability and transmission congestion rate; aggregate all indicators of all scenarios to form a set of power market failure assessment indicators.

[0031] Preferably, the method further includes a power market transaction optimization step:

[0032] Define power transaction rule parameters and multi-objective optimization functions, use a model-independent meta-learning framework to generate multiple sets of extreme weather scenarios in the digital twin; solve the local optimal power transaction rule parameters through inner loop double game, and update the meta parameters to improve the cross-scene generalization ability of the optimal power transaction rule parameters; design a gradual constraint tightening mechanism, combine a graph neural network proxy model to accelerate the optimization process, output a set of Pareto optimal power transaction rule parameters, and improve the transaction stability of the power market under extreme weather.

[0033] Preferably, the operation process of power market transaction optimization includes:

[0034] Step 401: Define power transaction rule parameters and multi-objective optimization functions, the power transaction rule parameters of the power market include price cap, reserve capacity threshold and penalty coefficient, and set the multi-objective optimization function as the minimum amplitude of the deviation of the power market failure assessment indicators from the preset value,

[0035] Step 402: generating a plurality of sets of extreme weather scenario sets in the digital twin, each extreme weather scenario being generated by intervention of a key variable in the causal graph; inputting simulation transaction data, a causal graph, and power transaction rule parameters and a multi-objective optimization function of step 401; dividing the scenario sets into a training set and a test set, the training set being used for inner loop optimization, and the test set being used for outer loop verification, and outputting the extreme weather scenario sets, each set of scenarios containing scenario transaction data and market variable data;

[0036] Step 403: inner loop solving of locally optimal power transaction rule parameters: in the inner loop, locally optimal power transaction rule parameters of the training set scenario set are solved by a double-layer game method; the input is the training set scenario set and the power transaction rule parameters and the multi-objective optimization function; the processing process constructs a double-layer game model, the upper layer optimizes the power transaction rule parameters by the regulatory authority, and the lower layer maximizes the profit based on the parameter adjustment strategy by the market participants; after completing iteration, the output is a set of locally optimal power transaction rule parameters corresponding to each training set scenario;

[0037] Step 404: outer loop updating of meta-parameters and acceleration of optimization: in the outer loop, the meta-parameters are updated based on the test set scenario set to improve the cross-scenario generalization ability of the power transaction rule parameters, and the optimization is accelerated by combining a graph neural network proxy model and a progressive constraint tightening mechanism; the input includes the set of locally optimal power transaction rule parameters, the test set scenario set, and the power transaction rule parameters and the multi-objective optimization function, the meta-parameters are updated by a meta-learning method, and iteration optimization is performed until convergence, and the output is a set of Pareto optimal power transaction rule parameters.

[0038] Preferably, the graph neural network proxy model is a machine learning model based on a graph neural network model, which is used to quickly predict the power market failure evaluation index of the power market under different power transaction rule parameters; the power transaction rule parameters are taken as input, and the power market failure evaluation index is output; by learning the interaction relationship and historical data between market participants, the dynamic behavior of the power market is approximately simulated.

[0039] Preferably, the specific implementation of the progressive constraint tightening mechanism is to gradually narrow the adjustable range of the power transaction rule parameters in the optimization process, and the constraint range decays exponentially with the increase of the optimization round; the meta-parameters include the initial value, the adjustment step and the direction of the power transaction rule parameters, which are extracted from the optimization experience of a plurality of sets of extreme weather scenarios by a meta-learning method, so that the power transaction rule parameters under a new scenario are quickly converged to the Pareto optimal solution.

[0040] Technical effects and advantages of the present application:

[0041] (1) The cross-platform power transaction data interaction optimization method provided by the application simulates power market behavior through a digital twin, evaluates market failure risks under extreme weather scenarios, facilitates timely early warning, and provides support for decision-making of the power market; based on a generative adversarial neural network model, simulation transaction data of market participants are generated, a causal graph is constructed through a causal discovery algorithm and a graph neural network model, and counterfactual reasoning is performed, and the power market failure evaluation index is evaluated by combining a Monte Carlo simulation with a multi-agent reinforcement learning.

[0042] (2) The cross-platform power transaction data interaction optimization method provided by the application optimizes power transaction rule parameters of the power market based on the power market failure evaluation index, outputs a Pareto optimal rule parameter set by combining a gradual constraint tightening mechanism and a graph neural network agent model through inner loop double-layer game and outer loop meta-parameter updating, and improves the adaptability and stability of the power market under extreme weather scenarios. BRIEF DESCRIPTION OF DRAWINGS

[0043] Figure 1 The cross-platform power transaction data interaction optimization flowchart of the application.

[0044] Figure 2 The power market transaction optimization flowchart of the application. DETAILED DESCRIPTION

[0045] Exemplary embodiments of the present disclosure will be described more fully hereinafter with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it is understood that the present disclosure can be embodied in various forms and should not be limited by the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the present disclosure to those skilled in the art.

[0046] At the same time, it should be understood that the sizes of the various parts shown in the drawings are not drawn in accordance with the actual proportional relationship for the sake of description.

[0047] The following description of at least one exemplary embodiment is merely illustrative in nature and is in no way limiting of the application or its applications or uses.

[0048] Techniques, methods, and devices known to those of ordinary skill in the relevant art can not be discussed in detail herein, but should be considered as part of the specification in appropriate circumstances.

[0049] Example 1, refer to Figure 1 The cross-platform power transaction data interaction optimization flowchart of the application provides a cross-platform power transaction data interaction optimization method as shown in Figure 1 The cross-platform power transaction data interaction optimization method provided by the application includes the following steps:

[0050] Step one: building a digital twin of the electricity market

[0051] Generate simulated transaction data of market participants in the electricity market based on an adversarial neural network model, including quotes, power generation, and electricity demand. The generator of the adversarial neural network model generates simulated transaction data with a multi-dimensional random vector as input. The discriminator evaluates the distribution consistency of simulated transaction data and real data, and outputs the authenticity probability value corresponding to the simulated transaction data.

[0052] Explain and provide information on electricity market variables and power network topology to build the foundation for the digital twin.

[0053] Step two: causal discovery and counterfactual inquiry

[0054] Identify causal relationships between market variables from simulated transaction data of the digital twin through causal discovery algorithms (PC algorithm or GES algorithm) to build a causal graph. Use a graph neural network model to model the spatial dependence and causal transmission path between market participants. Conduct counterfactual reasoning in the digital twin under a single external shock scenario by intervening in key variables in the causal graph. Predict price changes and supply-demand balance in the electricity market. Market variables include fuel prices, weather conditions, power generation costs, market clearing prices, etc. Explain and provide structured representation of causal relationships to clarify the transmission mechanism between key variables and provide a theoretical framework for risk assessment.

[0055] Step three: market failure risk assessment

[0056] Deploy a multi-agent reinforcement learning model in the digital twin, generate extreme weather parameter distribution scenarios using Monte Carlo simulation, and simulate the dynamic game behavior of market participants under extreme weather parameter distribution scenarios. Conduct electricity market failure assessment based on causal transmission paths to obtain electricity market failure assessment indicators such as quantitative price fluctuation amplitude, supply gap probability, and power transmission congestion rate. Provide decision support for the electricity market based on electricity market failure assessment indicators.

[0057] In one possible embodiment, on the basis of the conventional adversarial neural network model loss function, a physical constraint term is added. During the training process of the adversarial neural network model, after the generator generates simulated transaction data, the degree of violation of the physical constraint is calculated by an energy conservation loss function. The physical constraint loss and the conventional adversarial neural network model loss are added together as the overall loss function of the generator.

[0058] It needs to be further explained in the embodiments of the application that in the digital twin construction, the physical constraint term is embedded into the training process of the adversarial neural network model, and the physical constraint term includes the direct current power flow equation and the dynamic line capacity constraint, and specifically includes the following contents: based on the direct current power flow equation, the injection power of each node and the line power flow are balanced calculated to ensure that the generated simulation transaction data meets the node energy conservation condition in the actual operation of the power grid (i.e. the algebraic sum of power generation, load and line loss is zero); At the same time, the dynamic line capacity constraint function is introduced, the power value of each power transmission line in the generated data is compared with the maximum allowable capacity under its current working condition, and the line power is restricted not to exceed the real-time capacity threshold through the penalty term; Ensure that the generated simulation transaction data meets the transmission capacity limit and node energy balance, and the penalty term can be a loss function based on the line power overrun degree.

[0059] It needs to be further explained in the embodiments of the application that the spatial dependency and causal transmission path between market participants are modeled using a graph neural network model, and by intervening in the key variables in the causal graph, the counterfactual scenario under external impact is simulated in the digital twin, and the market price and supply and demand changes are predicted, including the following steps:

[0060] Step 101: Extract market variables and perform standardization processing; search the causal graph structure through the Bayesian information criterion of the GES algorithm; embed the prior knowledge of the power network topology to constrain the direction of the causal edge (such as power plant→transmission line); determine the key variables (such as regional power generation, line failure rate) through Granger causality test and N-1 safety criterion verification;

[0061] Step 102: Market participants (such as power suppliers, users, and power transmission companies) are nodes of the graph structure, and the trading relationship and geographical location relationship of the market participants are edges of the graph structure; node features include historical prices, supply and demand, geographical location, line impedance, and weather sensitivity;

[0062] In one possible embodiment, in the process of converting the causal graph into a graph structure, through multi-layer attention calculation, an attention coefficient is assigned to each pair of nodes based on node features and edge attributes, and neighbor node features are aggregated to generate node embeddings, representing the spatial dependency and causal transmission path between market participants;

[0063] Step 103: Learn the spatial dependency and causal transmission path between nodes through the attention mechanism; learn the spatial dependency and causal transmission path between nodes;

[0064] Step 104: Intervene in the key variables (such as setting the regional power generation to zero or adjusting the power generation cost) through the do-calculus rule to simulate the counterfactual scenario under external impact in the digital twin, and predict market price fluctuations, supply and demand balance, and power transmission congestion risk.

[0065] For example, simulate the impact of an extreme weather event (such as severe drought) on a certain water-based region, by intervening in the power generation cost node of the region, and observe how this change propagates through the graph structure to other regions, affecting the price and supply-demand balance of the entire market; this method can help market managers to predict potential market risks in advance and develop corresponding response strategies.

[0066] Explanation, Do-calculus rules are a set of rules for handling interventions in causal inference proposed by Judea Pearl, based on causal graphs, helping to infer causal effects from observed data in the presence of hidden variables or unobserved confounding factors.

[0067] Further, the key variables in the causal graph refer to variables that have a significant causal impact on the failure risk of the electricity market (such as price fluctuations, supply gaps), including:

[0068] External shock variables: regional power generation affected by extreme weather, transmission line failure rate;

[0069] Market transmission variables: node price, cross-region transmission power, reserve capacity demand;

[0070] Physical constraint variables: dynamic line capacity, node power balance deviation.

[0071] Explanation, the key variables are obtained by extracting the causal dependence relationship between variables (such as weather → power generation → price) from the simulation trading data through causal discovery algorithms; conduct Granger causality test or Bayesian posterior probability analysis on the edge weights in the causal graph to filter out significant (p < 0.05) causal relationships; combine power system operation rules and expert experience to determine key variables (such as the transmission corridor capacity corresponding to the storm area).

[0072] Further, use the key variables in the causal graph (such as weather conditions, power generation costs) to design the parameter range of multiple extreme weather scenarios (such as rainfall 0-50 mm, power generation cost increase 5-15%), providing input templates for Monte Carlo simulation; for example, based on the causal transmission path of the drought scenario (weather conditions → power generation cost → market clearing price), set the intervention variables of the storm scenario (such as a 20% decrease in transmission line capacity).

[0073] Further, integrate the nodes of the causal graph (GAT-generated vectors representing the state of market participants) with the simulation trading data into the graph structure format required in step three (node features include offers, power generation, and edges represent transaction relationships); configure the data interface to ensure that the multi-agent reinforcement learning model in step 3 can directly call the causal graph and prediction results.

[0074] It needs to be further explained in the embodiments of the present application that the multi-agent reinforcement learning model comprises a graph attention network module and a strategy network of each agent,

[0075] The graph attention network module is used to generate node embedding of market participants to support dynamic game simulation, the node embedding is a way to represent the state and relationship of market participants by numbers, and is usually a fixed-length vector; the node embedding captures the dynamic interaction of the electricity market through a dynamic attention mechanism, improving the quantitative accuracy of market failure risk assessment;

[0076] A strategy network is configured for each agent, the strategy network is a decision-making model based on a deep neural network, analyzes simulation transaction data to obtain an agent state, and comprises self-bid, power generation or electricity demand; the node embedding and the agent state are inputted, and fuel prices, weather conditions, power generation costs and market clearing prices in a market variable set are outputted; action decisions of market participants are outputted, and the action decisions at least include one of an amplitude of adjusting a bid and a plan of changing power generation or electricity demand; the strategy network is obtained through reinforcement learning training based on simulation transaction data and causal transmission paths in a digital twin, and the decision-making behaviors of the agents are optimized based on a reward function of the agent to adapt to market dynamics under extreme weather parameter distribution scenarios.

[0077] It is explained and illustrated that a deep neural network is used to build a strategy network, initial dependencies are preset based on causal transmission paths (for example, a higher weight of a hydropower plant node on a neighboring load center node), training parameters of the multi-agent reinforcement learning model are set, including a learning rate (0.001) and a discount factor (0.95), and an initialized multi-agent reinforcement learning model is outputted; the update of the agent strategy network is calibrated by the transaction data collected in advance, the decision-making features of each participant are ensured to be consistent with historical data, a closed-loop feedback mechanism is formed, and a data foundation is laid for subsequent risk assessment;

[0078] Further, the present application does not limit the specific quantitative formula of the reward function, and the reward function is set based on the attributes of the agent, for example, the reward function of a power generator is based on profit, and is calculated as the market clearing price multiplied by the power generation minus the power generation cost multiplied by the power generation; the reward function of a user is based on the negative value of the electricity cost, and is calculated as the opposite number of the market clearing price multiplied by the electricity demand; the reward function of a power transmission company is based on the negative cost of power transmission congestion.

[0079] It needs to be further explained in the embodiments of the present application that the multi-agent reinforcement learning model is obtained in the following way:

[0080] Step 201: Obtain simulation transaction data of the digital twin, including offers, power generation, electricity demand, etc.; supplement external data such as market variables (fuel prices, weather conditions, power generation costs) and power network topology information (line impedance, capacity limitations); build a graph structure based on a causal diagram;

[0081] Step 202: Use a graph attention network module to aggregate node neighbor information and capture spatial dependencies; for example, the graph attention network module dynamically assigns edge weights through an attention mechanism to highlight key interactions; each agent uses an independent policy network (such as a deep neural network) to optimize decisions based on reinforcement learning algorithms; the graph neural network model outputs node embeddings (representing market state sequences) that are input into each agent's policy network to guide action decisions; the action decisions of the agents (such as adjusting offers) are fed back to the graph neural network model to update the market state sequence;

[0082] Step 203: Use the digital twin as a simulation environment to provide dynamic market state sequences (such as price changes and supply-demand balance); design a reward function for each agent;

[0083] Step 204: Initialize the graph neural network model and policy network parameters; run multiple rounds of simulation in the digital twin, with agents taking actions based on the market state sequences output by the graph neural network model to obtain rewards; use reinforcement learning algorithms to update the policy network parameters while optimizing the graph neural network model to improve state prediction accuracy; introduce competition and cooperation mechanisms (such as sharing part of the rewards) to balance the goals of multiple agents;

[0084] Step 205: Verify the consistency of the power market failure assessment indicators (such as price volatility thresholds and power transmission congestion rates) output by the multi-agent reinforcement learning model with historical data or actual trends; deploy the multi-agent reinforcement learning model to the digital twin to assess power market failure risks in real time.

[0085] Further explained in the embodiments of the present application is that the operation process of market failure risk assessment includes the following steps:

[0086] Step 301: Build a multi-agent environment, define each market participant (such as power generators, users, and power transmission companies) as an agent, and all agents form an agent set; output the state, action, and reward function of each agent;

[0087] Step 302: Dynamic game simulation: run the multi-agent reinforcement learning model in the digital twin, combine Monte Carlo simulation to generate extreme weather parameter distribution scenarios, simulate the dynamic game behavior of market participants under extreme weather parameter distribution scenarios, calculate node embedding in each round of circulation through graph attention network, and adjust actions according to node embedding. Verify the physical feasibility of the power market state sequence, update the strategy and record the market state sequence;

[0088] Step 303: Quantitative evaluation of power market failure indicators: Based on the market state sequence set and the causal transmission path, calculate the power market failure evaluation indicators, including price fluctuation amplitude, supply gap probability and power transmission congestion rate; Aggregate all scenario indicators to form a set of power market failure evaluation indicators.

[0089] In the embodiments of the present application, it needs to be further explained that the acquisition method of the price fluctuation amplitude is: for each set of market state sequence, extract the power market clearing price time sequence, calculate the price change range, define it as the difference between the maximum price and the minimum price, take the 95th percentile of the price change range of all scenarios as the price fluctuation amplitude, and ensure to capture extreme cases;

[0090] The acquisition method of the supply gap probability is: for each set of market state sequence, calculate the proportion of unsatisfied electricity demand in each simulation (electricity demand minus power generation, divided by electricity demand); Count the number of simulations with unsatisfied proportion greater than 0, and divide by the total number of simulations to get the supply gap probability, which is in percentage;

[0091] The acquisition method of the power transmission congestion rate is: for each set of market state sequence, extract the power transmission line load rate time sequence; Calculate the proportion of times when the load rate exceeds the dynamic line capacity constraint as the power transmission congestion rate, which is in percentage; Combined with the causal transmission path, the key lines (such as the lines connecting the high load area) are analyzed first.

[0092] It is explained that the relationship between steps two and three is that: the causal graph generated in step two is used to guide the variable intervention in step three (for example, the root node variable in the causal graph is intervened first), and the counterfactual reasoning result (such as "fuel price rise leads to increase in power generation cost") is used as the input parameter of Monte Carlo simulation in step three; Step two provides static causal relationship, step three simulates the cascade effect of causal chain through dynamic game simulation; The single shock analysis of step two verifies the causal hypothesis, and the multi-variable coupling simulation of step three evaluates the systemic risk.

[0093] Summary: The embodiments of the present application focus on market modeling and risk assessment, from data generation, causal analysis, to risk quantification, forming a complete analysis process; step one generates simulated trading data (quotes, power generation, electricity demand), step two builds a causal diagram and predicts single scenario market changes, step three quantifies power market failure assessment indicators through multi-agent reinforcement learning model, the three are progressive, forming a closed loop of market analysis.

[0094] Embodiment 2, the difference from embodiment 1 is that the method further comprises:

[0095] Step four: power market transaction optimization

[0096] Define power trading rule parameters and multi-objective optimization function, use model-independent meta-learning framework to generate multiple sets of extreme weather scenarios in digital twins; through inner loop double-layer game to solve local optimal power trading rule parameters, outer loop to update meta parameters to improve the cross-scenario generalization ability of optimal power trading rule parameters; design a progressive constraint tightening mechanism, combined with a graph neural network proxy model to accelerate the optimization process, output a set of Pareto optimal power trading rule parameters, improve cross-scenario generalization ability, and improve the transaction stability of the power market under extreme weather.

[0097] In the embodiments of the present application, it needs to be further explained that, referring to Figure 2 the power market transaction optimization flowchart, the running process of the power market transaction optimization includes:

[0098] Step 401: Define power trading rule parameters and multi-objective optimization function, the power trading rule parameters of the power market include price cap, reserve capacity threshold and penalty coefficient, set transaction price correlation constraints and dynamic reserve threshold constraints (adjustable range of parameters); the price cap is used to limit the market clearing price, the reserve capacity threshold is used to ensure power supply reliability, and the penalty coefficient is used to constrain the behavior of market participants, which is the fine standard for intelligent agents violating the rules, prompting market participants to act according to the rules; set the multi-objective optimization function as the minimum deviation of the power market failure assessment indicator from the preset value, in a possible embodiment, the multi-objective optimization function is the minimum weighted sum of price fluctuation amplitude, supply gap probability and power transmission congestion rate; the weights are set by the regulatory agency according to market priorities, for example, the weight of price fluctuation amplitude is 0.4, the weight of supply gap probability is 0.3, and the weight of power transmission congestion rate is 0.3;

[0099] Explanation and illustration, in order to ensure the rationality of the power trading rule parameters, the constraint conditions of the power trading rule parameters are applied, for example: the price cap needs to be higher than the market average clearing price in the simulated trading data, the reserve capacity threshold needs to meet the N-1 safety criterion (i.e. the system can still operate normally under single component failure), and the penalty coefficient needs to be sufficient to encourage market participants to comply with the rules;

[0100] Generate multiple sets of extreme weather scenario sets in the digital twin, and divide them into training sets and test sets; solve the local optimal power trading rule parameters for each training set scenario through inner loop double-layer game, where the upper layer optimizes the power trading rule parameters by the regulatory agency, and the lower layer adjusts the strategy based on the parameters by the market participants; in the outer loop, update the meta-parameters based on the power market failure evaluation index of the test set scenario, and improve the cross-scene generalization ability of the power trading rule parameters; design a gradual constraint tightening mechanism to gradually narrow the range of power trading rule parameters as the optimization rounds increase, to ensure the convergence stability of the optimization process; combine the graph neural network proxy model to predict the power market failure evaluation index with the power trading rule parameters as input, to accelerate the optimization process;

[0101] Step 402: Generate multiple sets of extreme weather scenario sets in the digital twin, each extreme weather scenario is generated by intervention of key variables (such as fuel price, weather conditions) in the causal diagram; input includes simulation trading data (offer, power generation, electricity demand) of step 1, causal diagram (causal relationship between market variables) of step 2, and power trading rule parameters, multi-objective optimization function of step 401; divide the scenario set into training set (80%) and test set (20%), the training set is used for inner loop optimization, and the test set is used for outer loop verification, the output is the extreme weather scenario set, each scenario contains scenario trading data and market variable data;

[0102] Explanation: The extreme weather scenario set includes drought, rainstorm, and high temperature; for example, the drought scenario adjusts the key variables (such as a 10% increase in the cost of water and electricity generation) through the causal diagram, simulates the market participant behavior in the digital twin, and generates scenario trading data (updated offer, power generation); the scenario generation refers to the counterfactual reasoning result of step 2, such as the causal chain of fuel price increase leading to increased power generation cost, to ensure consistency between the scenario and the market variables, and the market variable data refers to the corresponding market behavior simulated in the digital twin;

[0103] Step 403: Inner loop to solve local optimal power trading rule parameters: in the inner loop, solve the local optimal power trading rule parameters for the training set scenario set through the double-layer game method; input is the training set scenario set (contains scenario trading data) and power trading rule parameters, multi-objective optimization function; the processing process constructs a double-layer game model, the upper layer optimizes the power trading rule parameters by the regulatory agency, and the lower layer adjusts the strategy based on the parameters by the market participants to maximize profits; after completing the iteration, output is the local optimal power trading rule parameter set corresponding to each training set scenario;

[0104] Explanation: In the lower-level game, market participants (such as power generators) adjust their bids and generation according to the current electricity trading rule parameters (e.g., price cap 120 yuan / MWh), and the profit is calculated as the market clearing price multiplied by the generation minus the generation cost. The solution is the optimal strategy combination. In the upper-level optimization, the regulatory agency adjusts the parameters based on the lower-level results, aiming to minimize the electricity market failure evaluation index in step 3. The optimization uses the gradient descent method with a step size of 0.01 and a convergence condition of parameter change less than 0.001.

[0105] Step 404: Outer loop updates meta-parameters and accelerates optimization: In the outer loop, update the meta-parameters based on the test set scenario set to improve the cross-scenario generalization ability of the electricity trading rule parameters, combine the graph neural network agent model and the progressive constraint tightening mechanism to accelerate the optimization; input includes the local optimal electricity trading rule parameter set, the test set scenario set, and the electricity trading rule parameter, the multi-objective optimization function, update the meta-parameters by meta-learning method, iterate optimization until convergence, output the Pareto optimal electricity trading rule parameter set.

[0106] Explanation: The test set scenario is used to evaluate the performance of the meta-parameters, and the electricity market failure evaluation index under the current electricity trading rule parameters is calculated and compared with the local optimal parameters in step 403; the meta-parameter update direction is to minimize the average of the electricity market failure evaluation index of the test set, and the step size is 0.001; the progressive constraint tightening mechanism gradually narrows the parameter range; the graph neural network agent model takes the electricity trading rule parameter as input and predicts the electricity market failure evaluation index (e.g., input price cap 120 yuan / MWh, predict supply gap probability 5%), reducing the computational cost of digital twin simulation; the output Pareto optimal electricity trading rule parameter set contains multiple combinations, such as price cap 120 yuan / MWh, reserve capacity 100 MW, penalty coefficient 12 yuan / MWh; price cap 130 yuan / MWh, reserve capacity 90 MW, penalty coefficient 13 yuan / MWh, suitable for different extreme weather scenarios.

[0107] Further explained in the embodiments of the present application, the graph neural network agent model is a machine learning model based on a graph neural network model, which is used to quickly predict the electricity market failure evaluation index of the electricity market under different electricity trading rule parameters; taking the electricity trading rule parameter as input, outputting the electricity market failure evaluation index, approximating the dynamic behavior of the electricity market by learning the interaction between market participants and historical data, the core features include:

[0108] Input: Electricity trading rule parameters, such as price cap, reserve capacity threshold, penalty coefficient, etc.

[0109] Output: Electricity market failure evaluation index, such as price fluctuation amplitude, supply gap probability, transmission congestion rate, etc.

[0110] Function: Accelerate the optimization of power transaction rule parameters.

[0111] Further, the specific implementation of the progressive constraint tightening mechanism is to gradually narrow the adjustable range of the power transaction rule parameters in the optimization process, and the constraint range exponentially decays with the increase of the optimization round; the meta-parameters include the initial value, adjustment step and direction of the power transaction rule parameters, which are extracted from the optimization experience of multiple sets of extreme weather scenarios through a meta-learning method, so that the adjustment of the power transaction rule parameters under a new scenario quickly converges to a Pareto optimal solution.

[0112] Further explained in the embodiments of the present application is the obtaining method of the graph neural network agent model:

[0113] Based on the simulation transaction data of step one, the causal diagram of step two and the power market failure evaluation index of step three, a digital twin is constructed and trained to predict the power market failure evaluation index corresponding to the transaction rule parameters;

[0114] Firstly, simulation transaction data (bids, power generation, electricity demand) are obtained as initial node features, transaction relationships between market participants (power suppliers, users, power transmission companies) and power transmission lines are extracted from the causal diagram as edges of the graph structure, and a graph structure is constructed, wherein the node features are supplemented with the power market failure evaluation index;

[0115] Then, a graph attention network is designed as an agent model, the input is the transaction rule parameters (price cap, reserve capacity threshold, penalty coefficient), the node features are aggregated through multi-layer attention calculation, and the predicted power market failure evaluation index is generated;

[0116] The set of extreme weather scenarios generated in step four is used to train the model with the training set scenarios (80%), optimize the GAT parameters to make the predicted value close to the actual power market failure evaluation index, and verify the model accuracy with the test set scenarios (20%);

[0117] Finally, the graph neural network agent model is output to predict the effect of the transaction rule parameters and reduce the calculation cost of the digital twin simulation.

[0118] The model-independent meta-learning framework is used to help the electricity market quickly find electricity transaction rule parameters suitable for different extreme weather scenarios without the need to adjust from scratch each time; it is not limited to a specific algorithm, and in the inner loop, the electricity transaction rule parameters are quickly adjusted for each specific scenario (such as a drought scenario) to find the best parameter combination (locally optimal electricity transaction rule parameters) in this scenario; in the outer loop, the experience of multiple scenarios (such as drought, storm, and high temperature) is summarized to find a general adjustment method (meta-parameters) to enable the system to quickly find suitable parameters when facing new scenarios.

[0119] In the model-independent meta-learning framework, the meta-parameters are a set of parameters used to initialize and guide the optimization of the electricity transaction rule parameters, and the meta-parameters play a role in the following ways:

[0120] The initial values of the electricity transaction rule parameters are set (for example, the price cap is 120 yuan / MWh, the reserve capacity is 100 MW, and the penalty coefficient is 12 yuan / MWh), ensuring that the optimization does not start from zero in each new extreme weather scenario (such as high temperature or storm);

[0121] The adjustment direction and step size of the electricity transaction rule parameters are determined, and the optimization effect is evaluated according to the electricity market failure evaluation index (for example, the price fluctuation amplitude is reduced to below 50 yuan / MWh), ensuring that the adjustment process is efficient and stable;

[0122] By summarizing the optimization experience of multiple groups of extreme weather scenarios, a general adjustment strategy is constructed, enabling the electricity transaction rule parameters to quickly converge to the Pareto optimal electricity transaction rule parameter set in unseen scenarios (such as future new storms), maintaining the stability of the electricity market.

[0123] Summary: The embodiments of the present application focus on transaction rule optimization, use the simulation transaction data, causal diagram, and electricity market failure evaluation index output by Embodiment 1, and generate the Pareto optimal transaction rule parameter set (price cap, reserve capacity threshold, and penalty coefficient) through the model-independent meta-learning framework to improve the cross-scenario generalization ability and the transaction stability of the electricity market in extreme weather.

[0124] Finally: The above description is only for the preferred embodiments of the present application and is not intended to limit the present application. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. A cross-platform power transaction data interaction optimization method, characterized in that, The method comprises the following steps: generating simulated transaction data of market participants, including quotes, power generation and power demand, based on an adversarial neural network model; identifying causal relationships between market variables from the simulated transaction data of the digital twin through a causal discovery algorithm to construct a causal graph; modeling spatial dependencies and causal transmission paths between market participants using a graph neural network model; and simulating counterfactual scenarios under external shocks in the digital twin by intervening in key variables in the causal graph; In the construction of the digital twin, physical constraint terms are embedded in the training process of the adversarial neural network model, including direct current flow equations and dynamic line capacity constraints; the direct current flow equations are used to balance the injected power and line power flow at each node to ensure that the generated simulated transaction data meets the node energy conservation condition in the actual operation of the power grid; A dynamic line capacity constraint function is introduced to compare the power value of each transmission line in the generated data with the maximum allowable capacity under its current working condition, and a penalty term is used to constrain the line power not to exceed the real-time capacity threshold; ensure that the generated simulated transaction data meets the transmission capacity limit and node energy balance; In the digital twin, a multi-agent reinforcement learning model is deployed to generate extreme weather parameter distribution scenarios using Monte Carlo simulation, simulate the dynamic game behavior of market participants under extreme weather parameter distribution scenarios, evaluate power market failure based on causal transmission paths, and obtain power market failure evaluation indicators for early warning based on the power market failure evaluation indicators; Modeling spatial dependencies and causal transmission paths between market participants using a graph neural network model, simulating counterfactual scenarios under external shocks in the digital twin by intervening in key variables in the causal graph, and predicting market price and supply and demand changes, including the following steps: Constructing a causal graph: extracting market variables and performing standardization processing; searching for a causal graph structure through the Bayesian information criterion of the GES algorithm; embedding prior knowledge of power network topology to constrain the direction of causal edges; determining key variables through Granger causality test and N-1 safety criterion verification; Building a graph structure based on the causal graph: market participants are nodes in the graph structure, and the trading relationship and geographical location relationship of market participants are edges in the graph structure; Node features include historical prices, supply and demand conditions, geographical locations, line impedances, and weather sensitivities; Learning spatial dependencies and causal transmission paths between nodes through an attention mechanism; Learning spatial dependencies and causal transmission paths between nodes; 2. The cross-platform power transaction data interaction optimization method of claim 1, wherein, Intervening in key variables through do-calculus rules to simulate counterfactual scenarios under external shocks in the digital twin, and predicting market price fluctuations, supply and demand balance, and transmission congestion risks. The multi-agent reinforcement learning model includes a graph attention network module and a strategy network for each agent; The graph attention network module is used to generate node embeddings of market participants to support dynamic game simulation, and the node embeddings capture power market dynamic interactions through a dynamic attention mechanism to improve the quantitative accuracy of market failure risk assessment. A strategy network is configured for each agent, which is a deep neural network-based decision model analyzing the agent state obtained from the simulation trading data, including its own offer, power generation or electricity demand, taking the node embedding and agent state as input, and the fuel price, weather condition, power generation cost and market clearing price in the set of market variables as output, and outputting the action decision of the market participant, which at least includes one of adjusting the amplitude of the offer and changing the plan of power generation or electricity demand; the strategy network is obtained through reinforcement learning training based on the simulation trading data and the causal transmission path in the digital twin, and the decision behavior of each agent is optimized based on the reward function of the agent to adapt to the market dynamics under the scenario of extreme weather parameter distribution.

3. The cross-platform power transaction data interaction optimization method of claim 2, wherein, The acquisition method of the multi-agent reinforcement learning model is: Obtain simulation trading data of the digital twin, and construct a graph structure based on a causal diagram; Use a graph attention network module to aggregate node neighbor information and capture spatial dependencies; Input into the strategy network of each agent to guide action decision; the action decision of the agent is fed back to the graph neural network model to update the market state sequence; Use the digital twin as a simulation environment to provide a dynamic market state sequence; design a reward function for each agent; Initialize the parameters of the graph neural network model and the strategy network; run multiple rounds of simulation in the digital twin, and the agent takes action according to the market state sequence output by the graph neural network model to obtain rewards; update the strategy network parameters using a reinforcement learning algorithm, and optimize the graph neural network model to improve state prediction accuracy; introduce competition and cooperation mechanisms to balance the goals of multiple agents; Verify the consistency of the power market failure assessment indicators output by the multi-agent reinforcement learning model with historical data or actual trends; Deploy the multi-agent reinforcement learning model to the digital twin to assess the risk of power market failure in real time.

4. The cross-platform power transaction data interaction optimization method of claim 2, wherein, The operation process of market failure risk assessment includes the following steps: Construct a multi-agent environment, define each market participant as an agent, and all agents form an agent set; output the state, action and reward function of each agent; Dynamic game simulation: run the multi-agent reinforcement learning model in the digital twin, generate extreme weather parameter distribution scenarios combined with Monte Carlo simulation, simulate the dynamic game behavior of market participants under the extreme weather parameter distribution scenario, calculate the node embedding in each round of loop through the graph attention network, and the agent adjusts the action according to the node embedding, verifies the physical feasibility of the power market state sequence, updates the strategy and records the market state sequence; Quantify the power market failure assessment indicators: based on the set of market state sequences and the causal transmission path, calculate the power market failure assessment indicators, including price fluctuation amplitude, supply gap probability and power transmission congestion rate; aggregate the indicators of all scenarios to form a set of power market failure assessment indicators.

5. The cross-platform power transaction data interaction optimization method according to any one of claims 1-4, characterized in that, It also includes the power market transaction optimization step: The power transaction rule parameters and the multi-objective optimization function are defined, and a model-independent meta-learning framework is adopted to generate multiple sets of extreme weather scenarios in the digital twin; a local optimal power transaction rule parameter is solved through an inner loop double-layer game, and an outer loop is used to update the meta parameter to improve the cross-scenario generalization ability of the optimal power transaction rule parameter; a gradual constraint tightening mechanism is designed to accelerate the optimization process in combination with a graph neural network proxy model, and a set of Pareto optimal power transaction rule parameters is output to improve the transaction stability of the power market under extreme weather.

6. The cross-platform power transaction data interaction optimization method of claim 5, wherein, The operation process of the power market transaction optimization includes: The power transaction rule parameters of the power market include the price cap, the reserve capacity threshold and the penalty coefficient, and the multi-objective optimization function is set as the minimum amplitude of the deviation of the power market failure evaluation index from the preset value, A set of multiple extreme weather scenarios is generated in the digital twin, and each extreme weather scenario is generated by intervention of key variables in the causal diagram; the input includes simulation transaction data, a causal diagram, power transaction rule parameters and a multi-objective optimization function of step 401; and the input is divided into a training set and a test set, the training set is used for inner loop optimization, and the test set is used for outer loop verification, and the output is a set of extreme weather scenarios, each scenario including scenario transaction data and market variable data; In the inner loop, a local optimal power transaction rule parameter is solved for the training set scenario set through a double-layer game method; the input includes the training set scenario set and the power transaction rule parameter and the multi-objective optimization function; the processing process constructs a double-layer game model, the upper layer optimizes the power transaction rule parameter by the regulatory agency, and the lower layer adjusts the strategy based on the parameter to maximize the profit; after iteration, a set of local optimal power transaction rule parameters corresponding to each training set scenario is output; In the outer loop, the meta parameter is updated based on the test set scenario set to improve the cross-scenario generalization ability of the power transaction rule parameter, and the optimization is accelerated in combination with the graph neural network proxy model and the gradual constraint tightening mechanism; the input includes a set of local optimal power transaction rule parameters, a test set scenario set, and power transaction rule parameters and a multi-objective optimization function, the meta parameter is updated through a meta-learning method, and iteration optimization is performed until convergence, and a set of Pareto optimal power transaction rule parameters is output.

7. The cross-platform power transaction data interaction optimization method of claim 6, wherein, The graph neural network proxy model is a machine learning model based on a graph neural network model, which is used to quickly predict the power market failure evaluation index of the power market under different power transaction rule parameters; the power transaction rule parameter is used as the input, and the power market failure evaluation index is output, the interaction relationship between the market participants and the historical data are learned, and the dynamic behavior of the power market is approximately simulated.

8. The cross-platform power transaction data interaction optimization method of claim 7, wherein, The specific implementation of the gradual constraint tightening mechanism is to gradually narrow the adjustable range of the power transaction rule parameter in the optimization process, and the constraint range decays exponentially with the increase of the optimization round; the meta parameter includes the initial value, the adjustment step and the direction of the power transaction rule parameter, which is extracted from the optimization experience of multiple sets of extreme weather scenarios through the meta-learning method, so that the power transaction rule parameter under the new scenario is quickly converged to the Pareto optimal solution.

Images