Power simulation multi-agent game spot electricity price probability prediction calibration method and system

Abstract

The present application relates to the cross field of power market simulation and power system automation, and discloses a kind of power simulation multi-agent game spot electricity price probability prediction calibration method and system.The present application is based on grid, unit, transaction settlement basic data to build the simulation base containing grid constraint, clearing engine, settlement module;Differentiated market agent is constructed, and near-end strategy optimization is used to complete decentralized adversarial training, and a stable game model is obtained;Through parameter random disturbance combined with Monte Carlo simulation, the electricity price probability distribution and risk index are obtained by statistics;A Bayesian variational inference update framework is built, the model is fine-tuned online from the microcosmic parameter level of agent, and a multi-scenario calibration trigger mechanism is set to continuously reduce the simulation and real market deviation.The present application is adapted to the domestic spot market node electricity price and deviation evaluation mechanism, solves the problem that traditional scheme lacks efficient online calibration and prediction accuracy is easy to decay, and guarantees the long-term accuracy and reliability of electricity price probability prediction.

Description

Technical Field

[0001] This invention relates to the interdisciplinary field of power market simulation and power system automation, specifically to a method and system for calibrating the probability prediction of spot electricity prices in a multi-agent game simulation. Background Technology

[0002] As the reform of the electricity spot market continues to deepen, the electricity pricing mechanism is gradually shifting from traditional government-set prices to a competitive bidding model jointly determined by market supply and demand and the bidding behavior of major players. Accurately predicting electricity spot prices and quantifying uncertainties is of significant engineering value for power generation companies' bidding decisions, electricity sales companies' power purchase planning, and energy storage operators' charging and discharging scheduling. Electricity spot prices are influenced by multiple factors, including grid topology constraints, fuel price fluctuations, random output from new energy sources, and strategic game theory among market participants. They exhibit highly nonlinear, non-stationary, and strongly stochastic characteristics, posing significant technical challenges to accurate electricity price prediction and market simulation.

[0003] Current technologies for electricity market simulation and price forecasting have significant shortcomings, making them difficult to adapt to domestic market-based trading scenarios. First, existing simulation methods are incompatible with China's unique market rules. Rule-based agent models pre-set fixed pricing logic, failing to simulate the dynamic learning behavior of market participants. Mainstream open-source simulation frameworks, adapted to European auction mechanisms, have clearing and settlement modules that are incompatible with my country's complex rule system of nodal marginal pricing, segmented pricing, and deviation assessment. Supply function equilibrium models rely on idealized Nash equilibrium assumptions, deviating from the actual market's characteristics of information asymmetry and bounded rationality, resulting in poor simulation realism.

[0004] Secondly, existing technologies have not established a complete closed loop from simulation to prediction. Traditional simulation systems can only output deterministic electricity price results and cannot provide quantitative indicators of uncertainty such as probability distribution and confidence interval, making it difficult to support risk decision-making. At the same time, they lack a real-time synchronization mechanism with actual market data, the prediction accuracy degrades continuously over time, and the operation process is not compatible with the daily operation system of the domestic power market, resulting in insufficient engineering practicality.

[0005] Finally, online model updates face significant technical bottlenecks. Traditional online learning methods are only applicable to single prediction models. When applied to multi-agent game systems, updating individual parameters can easily disrupt the overall game equilibrium. The computational complexity of algorithms such as Kalman filtering and particle filtering increases exponentially with the state dimension, making them unsuitable for large-scale market simulation scenarios. Existing parameter calibration only supports global parameter adjustments, making it difficult to achieve fine-grained calibration of the agent's micro-behavior.

[0006] While existing research can independently employ deep reinforcement learning (DRL) to model agent behavior and Monte Carlo (MC) methods to conduct uncertainty analysis, there is currently no publicly available technical solution that organically integrates DRL, proximal policy optimization (PPO), Monte Carlo simulation, and variational inference online calibration, and specifically adapts to unique market rules in China, such as bias assessment and blocking surplus allocation. Existing simulation schemes are mostly based on simplified market models and do not accurately replicate the actual SCED clearing engine and domestic settlement rules, thus failing to address the long-term degradation of prediction accuracy.

[0007] In summary, the industry urgently needs an integrated technical solution that adapts to the rules of China's electricity spot market, can generate probabilistic electricity price forecasts, and supports efficient online calibration. Summary of the Invention

[0008] In order to overcome the shortcomings of the existing technology, the purpose of this invention is to provide a method and system for probabilistic prediction and calibration of spot electricity prices in power simulation multi-agent game, so as to solve the technical problems in the existing technology that cannot be adapted to the marginal electricity price plus deviation assessment system of the spot electricity market node and lack an efficient online calibration mechanism to maintain prediction accuracy.

[0009] This invention is achieved through the following technical solution: In a first aspect, the present invention provides a method for calibrating the probability prediction of spot electricity prices in a multi-agent game simulation for power grids, comprising: The system collects basic data on grid parameters, generating unit parameters, market trading and settlement rules for the target electricity spot market. Based on this basic data, it constructs a simulation environment adapted to the domestic electricity spot market mechanism and power system. The simulation environment provides a simulation foundation that can realistically replicate the operating mechanism of the electricity market. The power system simulation environment includes a grid physical constraint model, a clearing optimization engine, and a market settlement module. Based on the obtained simulation foundation, we construct multiple types and differentiated market agents for the diverse trading entities in the electricity market, and configure the state space, action space and comprehensive reward mechanism of each market agent. We use the near-end strategy optimization algorithm to conduct multi-stage decentralized adversarial training on all market agents, constrain the diversity of agent strategies, and iteratively converge to obtain a stable multi-agent game model with dynamic learning and strategy game capabilities. Based on the trained and stable multi-agent game model, random perturbations are applied to the agent policy parameters. The random game process in the market is simulated through Monte Carlo simulation. The simulation results are statistically analyzed to quantify the electricity price probability distribution and risk prediction indicators. We collect real-world electricity market operation data, construct a Bayesian variational inference update framework, perform refined online calibration and update of the trained stable multi-agent game model from the perspective of agent micro-behavioral parameters, and configure a multi-scenario calibration trigger mechanism to continuously correct the deviation between the simulation model and the real market.

[0010] Preferably, the clearing optimization engine is a safety-constrained economic dispatch clearing engine, which constructs an objective function with the goal of minimizing the total system electricity purchase cost:

[0011] In the formula: For the first The quoted prices for each price range are in yuan per megawatt-hour. For the first The winning bid output for each bidding segment is in megawatts; The safety constraints, economic dispatch clearing engine configuration, and generator output upper and lower limit constraints. Branch flow constraints System power balance constraints ; The power transfer distribution factor matrix is ​​calculated based on the power grid topology and branch impedance parameters to realize the conversion between node injected power and branch power flow. The node marginal electricity price, which includes energy components, congestion components, and loss components, is obtained by solving the dual decomposition. The settlement module is used to implement the functions of medium- and long-term contract deviation settlement, deviation assessment fee calculation, and blocking surplus allocation.

[0012] Preferably, the market intelligence agent includes a power generation intelligence agent, a power sales intelligence agent, and an independent energy storage intelligence agent. The power generation intelligence agent includes a thermal power intelligence agent, a photovoltaic intelligence agent, a wind power intelligence agent, a hydropower intelligence agent, and an energy storage intelligence agent. The differentiated comprehensive reward function configured for the market intelligence agent is as follows:

[0013] In the formula: This represents the agent's overall reward value, expressed in yuan. The revenue from clearing the electricity market is equal to the sum of the products of the winning bid output and the marginal electricity price at each node; The cost of electricity generation includes fuel costs and start-up and shutdown costs; The penalty for deviation assessment is equal to the deviation between actual output and the current day's plan multiplied by the deviation assessment coefficient. This is a risk penalty item; The aforementioned risk penalty items The conditional value at risk (VaF) function is used to measure the volatility risk of returns. The specific calculation formula is as follows:

[0014] In the formula: This is the risk aversion coefficient, with a value range of [value missing]. to ; For confidence level of The condition is in the risk value function; Net profit equals cleared revenue minus power generation costs; Value at risk; This indicates the operation of taking positive values; The state space of the thermal power intelligent agent additionally includes parameters such as the unit's minimum technical output, ramp rate, and cold / hot start-up time. The state spaces of the photovoltaic intelligent agent and the wind power intelligent agent include meteorological forecast data and the prediction uncertainty range; The state space of the hydropower intelligent agent additionally includes reservoir water level, water inflow forecast, and water abandonment penalty coefficient; The state space of the energy storage intelligent agent additionally includes state of charge and charge / discharge efficiency parameters, and its action space includes charging, discharging and standby operation modes.

[0015] Preferably, the near-end policy optimization algorithm performs iterative training in three stages: pre-training, adversarial training, and refinement, as follows: The pre-training phase involves independently initializing the policy network parameters for various agents using historical market data, conducting independent training for at least one thousand rounds, and setting the learning rate to [value missing]. The discount factor γ is set to 0.99 to fit the benchmark pricing behavior; The adversarial training phase: all agents participate in market games synchronously, a single training round covers the complete day-ahead clearing cycle, the training is no less than 1,500 rounds, the pruning coefficient ε is set to 0.2, and the advantage function is calculated using generalized advantage estimation; The refining phase involves reducing the learning rate and continuously training for at least 1,500 rounds, while introducing an entropy regularization term to constrain policy diversity. The formula for the entropy regularization term is:

[0016] In the formula: Here is the entropy regularization coefficient, and its value range is... to ; For the policy network in state Select action The probability of; This is the action space.

[0017] Preferably, the risk aversion coefficient and deviation assessment weight of each market intelligent agent are configured differently, specifically: risk aversion coefficient of 0.1 to 0.3 and deviation assessment weight of 0.8 to 1.2 for thermal power intelligent agents; risk aversion coefficient of 0.05 to 0.15 and deviation assessment weight of 1.0 to 1.5 for new energy intelligent agents; and risk aversion coefficient of 0.2 to 0.4 and deviation assessment weight of 0.6 to 1.0 for energy storage intelligent agents.

[0018] Preferably, each market agent supports type expansion, and can construct virtual power plant agents and demand response agents; The state space of the virtual power plant intelligent agent includes the total capacity of aggregated resources, availability, and coordination cost; The state space of the demand response agent includes load elasticity coefficient, adjustment cost, and user satisfaction constraints.

[0019] Preferably, the agent policy parameters are simulated using Gaussian random perturbation, and the perturbation expression is:

[0020] In the formula: For the first The policy parameter vector after perturbation of each agent; For the first The original policy parameter vector after the agent has been trained; It is a random perturbation vector that follows a normal distribution; For the first The perturbation variance of the policy parameters of each agent; The disturbance variance is determined by back-calculation based on the mean square error between the Monte Carlo simulated electricity price and the actual electricity price within the historical rolling window. The calculation formula is as follows:

[0021] In the formula: This is the variance scaling factor, and its value range is... to ; For the first The mean square error of historical electricity prices for each node corresponding to an intelligent agent; The Monte Carlo simulation is run at least 1,000 times, and the output probability prediction indicators include expected electricity price, decimal electricity price, median electricity price, ninetieth digit electricity price, 95% value at risk, and 95% conditional value at risk.

[0022] Preferably, real-world electricity market operation data is collected, and the agent's bid level offset is selected in the Bayesian variational inference update framework. Price elasticity coefficient Risk preference coefficient As a calibration parameter, the expression for the calibration parameter is: ; The posterior distribution of the calibration parameters is approximated using a Gaussian distribution, as shown in the following expression:

[0023] In the formula: A variational approximation for the posterior distribution of the calibration parameters; The variational inference is solved by maximizing the lower bound of evidence as the optimization objective. The formula for the lower bound of evidence is:

[0024] In the formula: For observational data; It is the likelihood function; It is the prior distribution; The Kullback-Leibler divergence; The mean vector and covariance matrix are updated iteratively using the stochastic gradient descent method. The calibration triggering conditions include three mechanisms: the first is periodic triggering, with a default cycle of once a week; the second is event triggering, which automatically triggers calibration when the prediction error exceeds a set threshold for three consecutive trading days; and the third is rule change triggering, which immediately triggers calibration when a market rule modification signal is detected, and adopts an enhanced calibration mode with increased sampling frequency on the first trading day after the rule change.

[0025] Preferably, in the variational inference process, the covariance matrix adopts a diagonal matrix structure to reduce computational complexity, and the specific expression is as follows:

[0026] In the formula: , , These are the variance parameters for the price level offset, price elasticity coefficient, and risk preference coefficient, respectively.

[0027] Secondly, the present invention also provides a power simulation multi-agent game-based spot electricity price probability prediction calibration system, comprising: The simulation base construction module is used to collect basic data on grid parameters, unit parameters, market trading and settlement rules of the target electricity spot market. Based on the basic data, a simulation environment adapted to the domestic electricity spot market mechanism and power system is constructed. According to the power system simulation environment, a simulation base that can realistically replicate the operation mechanism of the electricity market is obtained. The power system simulation environment includes a grid physical constraint model, a clearing optimization engine and a market settlement module. The agent training module is used to construct various types and differentiated market agents based on the obtained simulation base for the diverse trading entities in the electricity market, and to configure the state space, action space and comprehensive reward mechanism of each market agent. The module uses a proximal policy optimization algorithm to conduct multi-stage decentralized adversarial training on all market agents, constrains the diversity of agent policies, and iteratively converges to obtain a stable multi-agent game model with dynamic learning and policy game capabilities. The probability prediction module is used to apply random perturbations to the agent policy parameters based on a trained and stable multi-agent game model, simulate the random game process in the market through Monte Carlo simulation, perform statistical analysis on the simulation results, and quantitatively output the probability distribution of electricity prices and risk prediction indicators. The online calibration module is used to collect real power market operation data, construct a Bayesian variational inference update framework, perform fine-grained online calibration and update of the trained stable multi-agent game model from the dimension of agent micro-behavioral parameters, and configure a multi-scenario calibration trigger mechanism to continuously correct the deviation between the simulation model and the real market.

[0028] Compared with the prior art, the present invention has the following beneficial technical effects: This invention provides a method for probabilistic prediction and calibration of spot electricity prices in a multi-agent game simulation of power. It is the first to build a high-fidelity digital twin simulation environment for power systems that is adapted to the rules of the electricity spot market. Through three core components—a finely modeled power grid physical constraint model, a security-constrained economic dispatch clearing engine, and a market settlement module—it accurately reproduces the three-component clearing mechanism of marginal electricity prices at nodes in the domestic electricity spot market, the settlement of deviations in medium- and long-term contracts, deviation assessment, and the full-process business rules of congestion surplus allocation. It completely overcomes the technical defects of traditional international open-source simulation frameworks that are only compatible with European auction mechanisms and cannot be compatible with domestic composite market rules, and achieves a true replication of the operating mechanism of the power market.

[0029] Furthermore, this invention breaks through the limitation of traditional electricity price forecasting, which can only output deterministic results. Relying on a multi-agent game model that has converged through training, and combining random perturbation of strategy parameters with large-scale Monte Carlo simulation, it effectively simulates the random game process of electricity market participants. Through statistical analysis, it quantifies and outputs multi-dimensional risk prediction indicators such as electricity price probability distribution, quantile electricity price, value at risk, and conditional value at risk, thereby achieving accurate quantification of the uncertainty of electricity price forecasting. This provides comprehensive and hierarchical data support for risk decision-making by various market participants such as power generation, electricity sales, and energy storage.

[0030] Furthermore, this invention innovatively introduces a variational inference online calibration mechanism to construct a complete Bayesian update framework. It can perform fine-grained parameter calibration from three micro-behavioral dimensions: agent bidding level shift, bidding elasticity, and risk preference coefficient. Combined with a more scenario-triggered mechanism that includes periodic, event, and rule-based variations, and combined with a variational inference algorithm with low computational complexity to efficiently approximate the Bayesian posterior distribution, it continuously synchronizes with real market operation data and corrects simulation model biases. This effectively curbs the problem of the continuous degradation of accuracy of traditional simulation prediction models over time, ensuring the prediction stability and accuracy of the system in the long term.

[0031] Furthermore, this invention constructs a closed-loop operation system covering the entire process from digital twin simulation of power systems to probabilistic electricity price prediction, achieving deep integration of simulation and prediction functions. Relying on standardized interfaces and automated scheduling mechanisms, it supports unattended continuous operation of the system, significantly improving the convenience and feasibility of engineering applications.

[0032] Furthermore, this invention deeply integrates near-end strategy optimization game theory algorithms, agent reward mechanisms, and core power technologies such as power system SCED clearing solutions, power transfer distribution factor calculation, and nodal marginal price component decomposition. Relying on the physical mechanisms of the power system and market trading rules to constrain agent game behavior, it forms a customized and practical technical solution for the electricity spot market. This effectively improves the professionalism, realism, and adaptability of electricity market simulation and electricity price probability prediction, and has extremely high industry promotion and engineering application value. Attached Figure Description

[0033] Figure 1 This is a flowchart of the power simulation multi-agent game-theoretic spot electricity price probability prediction calibration method in an embodiment of the present invention; Figure 2 This is a flowchart of the overall method in the embodiments of the present invention, which shows the complete technical process from power system simulation environment modeling, multi-agent modeling and training, Monte Carlo probability prediction to online calibration; Figure 3 This is a schematic diagram of modeling multiple types of intelligent agents in an embodiment of the present invention, showing the differentiated state space, action space and reward function structure of six types of intelligent agents: thermal power, photovoltaic, wind power, hydropower, energy storage and electricity sales. Figure 4 This is a diagram of the near-end policy optimization adversarial training architecture in an embodiment of the present invention, showing the data flow and control flow of the three training stages: pre-training, adversarial training, and refining. Figure 5 This is a schematic diagram of Monte Carlo probability prediction in an embodiment of the present invention, illustrating the calculation process of strategy parameter perturbation, independent simulation operation, and probability distribution statistics; Figure 6 The flowchart of the online calibration mechanism in this embodiment of the invention illustrates the triggering conditions and execution flow of three calibration mechanisms: periodic triggering, event triggering, and rule change triggering. Figure 7 This is a schematic diagram of the power simulation multi-agent game spot electricity price probability prediction calibration system in an embodiment of the present invention; In the diagram: 1. Simulation base construction module; 2. Agent training module; 3. Probability prediction module; 4. Online calibration module. Detailed Implementation

[0034] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0035] The purpose of this invention is to provide a method and system for calibrating the probability prediction of spot electricity prices in a multi-agent game simulation for power, in order to solve the technical problems in the prior art that it cannot adapt to the marginal electricity price plus deviation assessment system of the spot electricity market node and lacks an efficient online calibration mechanism to maintain prediction accuracy.

[0036] The present invention will now be described in further detail with reference to the accompanying drawings: See Figure 1 and Figure 2 This invention provides a method for calibrating the probability prediction of spot electricity prices in a multi-agent game simulation for power grids, comprising: Step 1: Collect basic data on grid parameters, generating unit parameters, market trading and settlement rules of the target electricity spot market. Based on the basic data, construct a power system simulation environment adapted to the domestic electricity spot market mechanism. Obtain a simulation base that can realistically replicate the operating mechanism of the electricity market based on the power system simulation environment. The power system simulation environment includes a grid physical constraint model, a clearing optimization engine, and a market settlement module. Specifically, based on the operating rules of the target electricity spot market, a digital twin simulation environment is constructed, comprising three core components: a network model, a security-constrained economic dispatch clearing engine, and a settlement module. The high fidelity of the simulation environment is a fundamental prerequisite for ensuring the reliability of subsequent agent training and probabilistic prediction results.

[0037] In terms of network model construction, node models are established based on the actual power grid topology, and the geographical location and electrical parameters of each generating node and load node are set. Branch impedance parameters, including resistance, are calculated for each transmission line. and reactance The power transfer distribution factor matrix is ​​calculated based on impedance parameters and network topology. The elements of this matrix represent the sensitivity coefficients of unit active power injection at nodes to active power flow in branches. The node model size is consistent with the actual power grid in the target market, ensuring that the simulation environment is physically consistent with the real power grid.

[0038] In terms of building the clearing engine, a safety-constrained economic scheduling solver based on linear programming is implemented, with the objective function shown below:

[0039] In the formula: For the first The quoted prices for each price range are in yuan per megawatt-hour. For the first The winning bid output for each bidding segment is in megawatts; The constraints include: upper and lower limits of generator output constraints. Branch flow constraints and system power balance constraints By employing a dual decomposition method, the Lagrange multipliers of branch power flow constraints are transformed into congestion component electricity prices at each node, which, together with energy and loss component electricity prices, constitute the complete node marginal electricity price. The clearing engine is implemented using a high-efficiency linear programming solver, ensuring that clearing calculations are completed in seconds under large-scale conditions.

[0040] In terms of settlement module construction, three core settlement functions are implemented. The first is medium- and long-term contract deviation settlement, which settles the deviation amount based on the deviation between the day-ahead planned output and the real-time actual output of each market participant, according to the nodal marginal electricity price. The second is deviation assessment fee calculation, which charges a penalty fee based on the deviation assessment coefficient for deviations exceeding the exemption threshold. The third is congestion surplus allocation, which allocates and returns the congestion surplus caused by the difference in nodal marginal electricity prices due to network congestion according to the load ratio of each node. The settlement module fully replicates the settlement rule system of the electricity spot market, providing accurate incentive signals for agent training.

[0041] Step 2: Based on the obtained simulation base, construct multiple types and differentiated market agents for the diverse trading entities in the electricity market, configure the state space, action space and comprehensive reward mechanism of each market agent; use the near-end strategy optimization algorithm to conduct multi-stage decentralized adversarial training on all market agents, constrain the diversity of agent strategies, and iteratively converge to obtain a stable multi-agent game model with dynamic learning and strategy game capabilities. Specifically, an independent policy optimization algorithm agent based on reward feedback is constructed for each type of market participant. Power generation agents are categorized into five types according to unit type: thermal power, photovoltaic, wind power, hydropower, and energy storage. Figure 3 As shown, an additional electricity sales agent and an independent energy storage agent are constructed. Each agent has a differentiated state space, action space, and reward function to accurately reflect the behavioral characteristics and decision constraints of different types of market participants.

[0042] In terms of state space design, the shared basic state variables for all agents include: the current time period. The system includes predicted load, predicted renewable energy output, fuel prices, historical electricity price sequences, and current contract positions. Based on this, different types of intelligent agents set differentiated state variables according to their own technical characteristics: thermal power agents additionally include minimum technical output, ramp-up rate, and start-up / shutdown time parameters; photovoltaic and wind power agents additionally include weather forecast data and output prediction uncertainty intervals; hydropower agents additionally include reservoir water levels, inflow forecasts, and water wastage penalty coefficients; energy storage agents additionally include state of charge and charge / discharge efficiency parameters; and electricity sales agents additionally include user load forecasts and medium- to long-term contract positions.

[0043] In terms of action space design, the action space for all power generation agents is unified as a 96-point segmented price curve, corresponding to the 96 scheduling periods of the day-ahead market. Each segment of the price curve is discretized into... Price tiers The value ranges from ten to fifty, with a default setting of twenty. The agent outputs a complete price and power quotation function for each scheduling period, consisting of a series of increasing price-power pairs, reflecting the supply curve characteristics of the power generation company. The action space of the energy storage agent, in addition to the quotation curve, includes the selection of three operating modes: charging, discharging, and standby.

[0044] Regarding the design of the reward function, the calculation formula is as follows: The differentiated comprehensive reward function for market agent configuration is as follows:

[0045] In the formula: This represents the agent's overall reward value, expressed in yuan. The revenue from clearing the electricity market is equal to the sum of the products of the winning bid output and the marginal electricity price at each node; The cost of electricity generation includes fuel costs and start-up and shutdown costs; The penalty for deviation assessment is equal to the deviation between actual output and the current day's plan multiplied by the deviation assessment coefficient. This is a risk penalty item; The reward functions for various intelligent agents comprehensively consider four factors: clearing revenue, power generation cost, deviation assessment penalty, and risk penalty. Introducing a risk penalty term into the reward function is a key design feature to ensure that the agent learns a robust pricing strategy. The risk penalty term uses a conditional value at risk function to measure the risk of return volatility, and its specific calculation formula is as follows:

[0046] In the formula: This is the risk aversion coefficient, with a value range of [value missing]. to ; For confidence level of The condition is in the risk value function; Net profit equals cleared revenue minus power generation costs; Value at risk; This indicates the operation of taking a positive value.

[0047] By adjusting the risk aversion coefficient, the risk preference of an agent can be controlled, allowing different types of agents to exhibit differentiated risk behavior characteristics.

[0048] In terms of the training process, a proximate strategy optimization algorithm is used for decentralized training, such as... Figure 3 As shown. The training process consists of three stages: The first stage is the pre-training stage, where historical market data is used to independently initialize the policy network parameters for each type of agent, with at least one thousand training rounds. In this stage, each agent learns independently without engaging in game interaction, only fitting its own bidding behavior baseline based on historical clearing results, so that the policy network obtains a reasonable initial parameter distribution. The learning rate in the pre-training stage is set to... Discount factor Set as The second phase is the adversarial training phase, where all agents simultaneously participate in market games within a simulated environment. Each training round includes a complete day-ahead market clearing cycle. Agents select bidding actions based on their own state observations, the clearing engine determines the winning bid and electricity price, calculate their respective rewards, and update the policy network parameters. This phase involves at least 1500 training rounds, with a pruning factor... Set as The advantage function is calculated using the generalized advantage estimation method. The third stage is the refinement stage, where the learning rate is reduced to one-tenth of that in the pre-training stage, and training continues for at least 1500 epochs to bring the policy network parameters to a stable Nash equilibrium, thus improving the robustness of the bidding policy. The refinement stage also introduces an entropy regularization term, the specific formula of which is as follows:

[0049] In the formula: Here is the entropy regularization coefficient, and its value range is... to ; For the policy network in state Select action The probability of; This is the action space.

[0050] The introduction of entropy regularization prevents policies from prematurely degenerating into deterministic policies and ensures sufficient policy diversity during Monte Carlo sampling.

[0051] Step 3: Based on the trained stable multi-agent game model, apply random perturbations to the agent policy parameters, simulate the market random game process through Monte Carlo simulation, perform statistical analysis on the simulation results, and quantify the output of electricity price probability distribution and risk prediction indicators. Specifically, Gaussian random perturbations are applied to the policy parameters of each trained agent, as shown in the following formula:

[0052] In the formula: For the first The policy parameter vector after perturbation of each agent; For the first The original policy parameter vector after the agent has been trained; It is a random perturbation vector that follows a normal distribution; For the first The perturbation variance of the policy parameters of each agent.

[0053] Disturbance variance The price is determined by inversely calculating the mean square error between the Monte Carlo simulation price and the actual price within the historical rolling window, using the following formula:

[0054] In the formula: This is the variance scaling factor, and its value range is... to ; For the first The mean square error of historical electricity prices for each node corresponding to an agent.

[0055] The default length of the historical scrolling window is thirty trading days, covering approximately one month of actual operational data. (Variance scaling factor) The default value is set to When the prediction accuracy is high, the value can be appropriately reduced to obtain a narrower confidence interval; when the prediction accuracy decreases, the value can be appropriately increased to obtain a wider confidence interval.

[0056] The Monte Carlo independent simulation should be run at least one thousand times. In practical engineering, the number of runs needs to be determined by comprehensively considering the requirements of computing resources and prediction accuracy. When the number of runs reaches one thousand, the standard error of the average electricity price estimate has converged to an acceptable level, and the marginal benefit of further increasing the number of simulations for improving prediction accuracy diminishes. Each simulation uses a different random seed to generate an independent perturbation vector to ensure the statistical independence between simulation results.

[0057] Statistical analysis of all simulation results yielded probabilistic forecasting indicators with six dimensions: the expected electricity price reflects the central trend of electricity prices; the decile and nintieth digits of the electricity price define the 80% confidence interval; the median electricity price serves as a robust central forecast; the 95% value at risk quantifies the risk of extreme high prices; and the 95% conditional value at risk further quantifies tail risk. These probabilistic forecasting indicators provide market participants with multi-level decision-making references ranging from conservative to aggressive.

[0058] Step 4: Collect real electricity market operation data, construct a Bayesian variational inference update framework, perform refined online calibration and update of the trained stable multi-agent game model from the dimension of agent micro-behavioral parameters, and configure a multi-scenario calibration trigger mechanism to continuously correct the deviation between the simulation model and the real market.

[0059] Specifically, a Bayesian update framework is used to feed real-world market data back to the simulation system, calibrating the agent's bid level offset, bid elasticity coefficient, and risk preference coefficient across three dimensions. A multidimensional normal distribution is used as a variational approximation of the posterior distribution, and the distribution parameters are iteratively updated using stochastic gradient descent, as shown in the following formula:

[0060] In the formula: This is a variational approximation of the posterior distribution of the calibration parameters.

[0061] The variational inference is solved by maximizing the lower bound of evidence as the optimization objective, and its mathematical expression is as follows:

[0062] In the formula: For observational data; It is the likelihood function; It is the prior distribution; The divergence is Kullback-Leibler.

[0063] ELBO consists of two parts: the first part is the expectation of the data likelihood, which encourages the variational distribution to approximate the true posterior; the second part is the KL divergence, which measures the difference between the variational distribution and the prior distribution. By optimizing ELBO, a good approximation of the true posterior can be obtained, provided it is computationally feasible.

[0064] In the variational inference process, the covariance matrix adopts a diagonal matrix structure to reduce computational complexity. The specific expression is as follows:

[0065] In the formula: , , These are the variance parameters for the price level offset, price elasticity coefficient, and risk preference coefficient, respectively.

[0066] Calibration triggering conditions include three mechanisms, such as Figure 6 As shown. The first is a periodic trigger mechanism, which automatically performs calibration at fixed time intervals, with a default cycle of once a week. Periodic triggering ensures the system's baseline update frequency, continuously tracking the slow drift of market behavior even when the prediction accuracy meets requirements. The second is an event-triggered mechanism, which automatically triggers calibration when the prediction error exceeds a set threshold for three consecutive trading days. The default error threshold for event triggering is set to 15% of the average absolute percentage error, which is dynamically adjusted based on market volatility characteristics. The third is a rule change trigger mechanism, which immediately triggers calibration upon detecting a market rule modification signal. Detection signals for rule change triggering include: changes in deviation assessment coefficients, modifications to congestion management rules, and adjustments to the clearing mechanism. On the first trading day after a rule change, the system automatically switches to enhanced calibration mode, increasing the sampling frequency to five times that of the normal mode, accelerating parameter convergence.

[0067] In constructing the simulation environment for this invention, the basic data of the target electricity spot market needs to be obtained in actual engineering. The network model construction requires the following input data: power grid topology data, including a list of nodes, a list of branches, and the relationships between nodes and branches; branch impedance parameters, including the per-unit values ​​of resistance and reactance of each branch; system reference capacity and reference voltage; generator parameters, including the rated capacity, minimum technical output, ramp rate, and node of each generator; and load node parameters, including the historical load curves and predicted load curves of each load node.

[0068] The safety-constrained economic dispatch clearing engine is implemented using a programming language, and the underlying linear programming solver can be either an open-source solver or a commercial solver. The input to the clearing engine consists of the bid curves submitted by all agents, network topology constraints, and load forecast data. The output is the marginal electricity price of each node and the winning output of each generating unit. In actual deployments, the clearing engine, for large-scale systems with more than two hundred nodes and more than one thousand bid segments, keeps the single clearing calculation time within two seconds.

[0069] The settlement module is implemented strictly in accordance with the settlement rules document of the target market. Taking the Shanxi electricity spot market as an example, the settlement module implements the following specific functions: the day-ahead market settles day-ahead electricity volume according to the cleared node marginal price; the real-time market settles real-time deviation electricity volume according to the actual running node marginal price; deviation assessment is carried out in accordance with the deviation exemption threshold and assessment coefficient of Shanxi Province; congestion surplus is proportionally distributed and returned according to the load electricity ratio of each node.

[0070] In the modeling and training of intelligent agents, both the policy network and value function network of various agents adopt a multilayer perceptron structure. The input layer dimension of the policy network is determined by the dimension of the state space. It has three hidden layers, each containing 256 neurons, with a linear rectified function as the activation function. The output layer uses a flexible maximum function to output the probability distribution of discrete actions. The value function network uses the same hidden layer structure as the policy network, and its output layer outputs a single scalar value representing the estimated state value. During training, the policy network and value function network share the parameters of the first two hidden layers to reduce the total number of parameters and accelerate training convergence.

[0071] The core hyperparameter settings for the proximal policy optimization algorithm are as follows: the learning rate is... The discount factor is The cutting factor is The generalized dominance estimation attenuation coefficient is Each update batch consists of 2048 time steps, and each batch of data is trained for ten epochs. During the pre-training phase, the learning rate remains constant. During the refining phase, the learning rate decreases to .

[0072] The convergence criterion for the training process is a comprehensive judgment based on the following three indicators: First, the average reward value of all agents fluctuates by no more than 5% over 500 consecutive training rounds; second, the average marginal electricity price of nodes changes by no more than 3 yuan per megawatt-hour over 500 consecutive training rounds; third, the norm of the parameter update gradient of the policy network is less than 1 / 300 over 200 consecutive training rounds. When all three conditions above are met, the training is considered to have converged, and the training process can be terminated early.

[0073] according to Figure 5 As shown, the Monte Carlo probability prediction module operates as follows: First, load the policy parameters of various agents that have been trained; second, calculate the perturbation variance of various agents; third, set a random seed to generate multiple independent perturbation vectors; fourth, for each perturbation vector, load the perturbed policy parameters into the simulation environment, run a complete day-ahead market clearing simulation, and record the electricity price at each node and time period; fifth, perform statistical analysis on all simulation results and calculate the probability prediction index for each time period.

[0074] In practical deployments, by employing multi-process parallel simulation technology, multiple independent simulations are distributed across multiple computing cores for simultaneous execution, significantly reducing prediction computation time. For example, with 1,000 simulations, on a computing server equipped with a 32-core CPU, the total time for predicting 96 time periods per day using Monte Carlo simulations can be controlled within 30 minutes, meeting the time requirements of current market operations.

[0075] The visualization of the probabilistic forecast output includes: a time series curve of the expected electricity price, a band chart of confidence intervals, a heat map of extreme electricity price risks, and a contour map of the spatial distribution of electricity prices at each node. These visualizations are provided to downstream decision support systems in a standardized data interface format.

[0076] This invention first constructs a digital twin simulation environment based on target market rules, including a network model, a Security Constrained Economic Dispatch (SCED) clearing engine, and a settlement module. The SCED clearing engine integrates the calculation of the three components of the Power Transfer Distribution Factor (PTDF), the Nodal Marginal Price (LMP), and network constraint processing. Second, it constructs differentiated policy optimization algorithm agents based on reward feedback for power generation companies, electricity sales companies, and energy storage operators. It uses a near-end policy optimization algorithm for decentralized adversarial training in three stages: pre-training, adversarial training, and refinement, and deeply couples the deviation assessment and congestion surplus allocation rules unique to the Chinese electricity market. Then, it applies Gaussian random perturbations to the trained policy parameters and runs a large-scale Monte Carlo simulation to generate the complete probability distribution and confidence interval of electricity prices for each time period. Finally, it achieves Bayesian online calibration through variational inference, continuously synchronizing actual market data through three mechanisms: periodic triggering, event triggering, and rule change triggering, forming a complete closed-loop operation process from simulation to prediction. Through typical simulation verification in the Shanxi electricity spot market, the median prediction average absolute percentage error reached 10.23%, and the coverage of the 10% to 90% confidence interval reached 87.3%. The online calibration mechanism effectively curbed the long-term degradation of prediction accuracy.

[0077] This invention first constructs a high-fidelity digital twin simulation environment for the power system based on the actual rules of the target market. This environment deeply integrates the SCED clearing engine (including PTDF, LMP three-component, and network constraints) and China's unique settlement rules. Then, multiple types of reward-feedback-based policy optimization algorithm agents are deployed in the simulation environment to simulate the bidding game behavior of various market participants. The agents are trained using a proximal policy optimization algorithm to obtain bidding strategies that approximate Nash equilibrium. Next, a Monte Carlo stochastic simulation is used to generate a probability distribution prediction of electricity prices. Finally, a variational inference online calibration mechanism continuously synchronizes with actual market data, enabling the long-term adaptive operation of the simulation prediction system. This invention emphasizes that the closed-loop interaction between DRL and SCED is a technical solution unique to the power market, rather than a simple application of AI algorithms.

[0078] Example 1 This case study uses the Shanxi Province electricity spot market as an application example to verify the predictive performance of the method of this invention in a real market environment. The Shanxi Province electricity spot market adopts a nodal marginal price mechanism, with day-ahead and real-time markets clearing and settling separately, and implements a deviation assessment and congestion surplus sharing mechanism.

[0079] The simulation environment is configured as follows: 215 power grid nodes, 328 branch lines, and 200 generator units, including 140 thermal power units, 25 photovoltaic units, 20 wind power units, 5 hydropower units, and 10 energy storage units. There are five smart agents from the electricity sales company and ten smart agents from the independent energy storage operator. The maximum system load is approximately 85 MW, and the installed capacity of new energy sources accounts for approximately 35%.

[0080] The agent training configuration is as follows: the price quote curve has twenty price tiers, with a price range of -500 yuan per megawatt-hour to +1500 yuan per megawatt-hour. The pre-training phase consists of 1000 rounds, the adversarial training phase 1500 rounds, and the refinement phase 1500 rounds, for a total of 4000 training rounds. Training is performed on a computing server equipped with a graphics processing unit, and the total training time is approximately 72 hours.

[0081] The Monte Carlo probability prediction configuration is as follows: 1000 Monte Carlo simulations, a historical rolling window length of 30 trading days, and a variance scaling factor of 1.0. The probability prediction output contains a complete probability distribution for 96 time periods.

[0082] The forecast performance evaluation results are as follows: Using the median electricity price as the point forecast value, the average absolute percentage error of the daily average electricity price forecast on a test set of sixty consecutive trading days is 10.23%. Using the interval from the 10th to the 90th digit as the 80% confidence interval, the coverage rate of actual electricity prices falling within the confidence interval is 87.3%, higher than the nominal coverage rate of 80%, indicating good calibration of the confidence interval. Regarding extreme electricity price forecasts, the deviation between the 95% conditional value at risk and the actual extreme electricity price does not exceed 12%, providing a reliable reference for market risk management.

[0083] This case study verifies the effectiveness of the online calibration mechanism in maintaining prediction accuracy over the long term through comparative experiments. Two control groups were set up: the first group was the uncalibrated control group, where the Monte Carlo probability prediction system was not updated with any parameters after initial training; the second group was the calibrated experimental group, which underwent weekly calibration according to the online calibration mechanism proposed in this invention, while also enabling event-triggered and rule-change-triggered mechanisms.

[0084] The experiment lasted for thirty consecutive trading days, approximately six weeks. Both systems used the same initial training parameters and the same Monte Carlo simulation configuration.

[0085] The prediction accuracy of the uncalibrated control group degraded significantly over time: the mean absolute percentage error was 10.5% in the first week, rising to 16.9% in the sixth week, with a degradation rate of approximately 0.9 to 1.3 percentage points per week, showing an accelerating trend. In contrast, the calibrated experimental group experienced error fluctuations of only 0.4 percentage points over six weeks, with a difference of 6.2 percentage points between the calibrated and uncalibrated control groups in the sixth week.

[0086] Further analysis shows that the calibration mechanism mainly corrects three types of parameter drifts: the average correction for the price level offset is 3.5 yuan per megawatt-hour, the average correction for the price elasticity coefficient is 8.2 percent, and the average correction for the risk preference coefficient is 0.06 percent.

[0087] Validation of the rule change scenario shows that when the deviation assessment coefficient was adjusted from 0.15 to 0.20, the rule change triggering mechanism initiated enhanced calibration on the first trading day, completing parameter adaptation within one day. The error in the first week after the change was 11.5%, far lower than the expected degradation under the uncalibrated condition, verifying the rapid adaptation capability.

[0088] Example 2 according to Figure 7 As shown, this embodiment also provides a power simulation multi-agent game-based spot electricity price probability prediction calibration system, including: Simulation base construction module 1 is used to collect basic data on grid parameters, unit parameters, market trading and settlement rules of the target electricity spot market, and to construct a simulation environment adapted to the domestic electricity spot market mechanism and power system based on the basic data. The simulation base that can realistically replicate the operation mechanism of the electricity market is obtained from the power system simulation environment. The power system simulation environment includes a grid physical constraint model, a clearing optimization engine and a market settlement module. The agent training module 2 is used to construct multiple types and differentiated market agents for the diverse trading entities in the electricity market based on the obtained simulation base, and to configure the state space, action space and comprehensive reward mechanism of each market agent. The near-end policy optimization algorithm is used to conduct multi-stage decentralized adversarial training on all market agents, constrain the diversity of agent policies, and iteratively converge to obtain a stable multi-agent game model with dynamic learning and policy game capabilities. The probability prediction module 3 is used to apply random perturbations to the agent strategy parameters based on the trained stable multi-agent game model, simulate the random game process in the market through Monte Carlo simulation, perform statistical analysis on the simulation results, and quantitatively output the probability distribution of electricity prices and risk prediction indicators. Online calibration module 4 is used to collect real power market operation data, construct a Bayesian variational inference update framework, perform refined online calibration and update of the trained stable multi-agent game model from the dimension of agent micro-behavioral parameters, and configure a multi-scenario calibration trigger mechanism to continuously correct the deviation between the simulation model and the real market.

[0089] In summary, this invention provides a method and system for probabilistic prediction and calibration of spot electricity prices in multi-agent game theory for power simulation. It is the first to achieve a high-fidelity digital twin simulation environment for power systems that adapts to the marginal electricity price plus deviation assessment rules of China's electricity spot market. Through three core components—a precise network model, a security-constrained economic dispatch clearing engine, and a settlement module—this invention can completely reproduce the clearing mechanism and settlement process of China's electricity spot market, solving the technical challenge of existing international simulation frameworks being unable to adapt to Chinese market rules.

[0090] This invention directly quantifies the uncertainty of electricity price forecasting using Monte Carlo stochastic simulation. Unlike existing methods that only output deterministic electricity price forecasts, this invention generates a complete probability distribution of electricity prices through large-scale Monte Carlo simulation, providing multi-dimensional probabilistic forecasting indicators such as expected electricity price, quantile electricity price, value at risk, and conditional value at risk, thus providing market participants with a hierarchical basis for risk quantification assessment.

[0091] The variational inference online calibration mechanism effectively curbs the long-term degradation of prediction accuracy. Unlike existing methods that rely on fixed simulation system parameters, this invention designs three calibration triggering mechanisms to perform refined online calibration of the agent from three dimensions: bid level shift, bid elasticity, and risk preference coefficient. It employs a computationally efficient variational inference method to approximate a Bayesian posterior distribution, ensuring continuous synchronization between the simulation system and the actual market.

[0092] A complete closed-loop operation process, from digital twin simulation of the power system to probabilistic electricity price prediction, has been established. This invention integrates the simulation system and the prediction system, and achieves unattended continuous operation of the system through standardized input / output interfaces and an automated operation scheduling mechanism.

[0093] This invention emphasizes deep integration with the technical characteristics of the power system, avoiding the risk of being identified as a pure AI algorithm. It closely combines a reward-feedback-based strategy optimization algorithm with core power system technologies such as the SCED clearing engine, PTDF calculation, and LMP three-component decomposition, forming a unique technical solution for the power market.

[0094] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A method for calibrating the probability prediction of spot electricity prices in a multi-agent game simulation of power systems, characterized in that... include: Collect basic data on grid parameters, unit parameters, market trading and settlement rules of the target electricity spot market, construct a power system simulation environment adapted to the domestic electricity spot market mechanism based on the basic data, and obtain a simulation base that can realistically replicate the operation mechanism of the electricity market based on the power system simulation environment; The power system simulation environment includes a power grid physical constraint model, a clearing optimization engine, and a market settlement module. Based on the obtained simulation foundation, we construct multiple types and differentiated market agents for the diverse trading entities in the electricity market, and configure the state space, action space and comprehensive reward mechanism of each market agent. We use the near-end strategy optimization algorithm to conduct multi-stage decentralized adversarial training on all market agents, constrain the diversity of agent strategies, and iteratively converge to obtain a stable multi-agent game model with dynamic learning and strategy game capabilities. Based on the trained and stable multi-agent game model, random perturbations are applied to the agent policy parameters. The random game process in the market is simulated through Monte Carlo simulation. The simulation results are statistically analyzed to quantify the electricity price probability distribution and risk prediction indicators. We collect real-world electricity market operation data, construct a Bayesian variational inference update framework, perform refined online calibration and update of the trained stable multi-agent game model from the perspective of agent micro-behavioral parameters, and configure a multi-scenario calibration trigger mechanism to continuously correct the deviation between the simulation model and the real market.

2. The method for calibrating the probability prediction of spot electricity prices in a multi-agent game simulation for power simulation, as described in claim 1, is characterized in that... The aforementioned clearing optimization engine is a safety-constrained economic dispatch clearing engine, with the objective function constructed to minimize the total system electricity purchase cost: In the formula: For the first The quoted prices for each price range are in yuan per megawatt-hour. For the first The winning bid output for each bidding segment is in megawatts; The safety constraints, economic dispatch clearing engine configuration, and generator output upper and lower limit constraints. Branch flow constraints System power balance constraints ; The power transfer distribution factor matrix is ​​calculated based on the power grid topology and branch impedance parameters to realize the conversion between node injected power and branch power flow. The node marginal electricity price, which includes energy components, congestion components, and loss components, is obtained by solving the dual decomposition. The settlement module is used to implement the functions of medium- and long-term contract deviation settlement, deviation assessment fee calculation, and blocking surplus allocation.

3. The method for calibrating the probability prediction of spot electricity prices in a multi-agent game simulation for power simulation, as described in claim 1, is characterized in that... The market intelligence agents include power generation intelligence agents, electricity sales intelligence agents, and independent energy storage intelligence agents. The power generation intelligence agents include thermal power intelligence agents, photovoltaic intelligence agents, wind power intelligence agents, hydropower intelligence agents, and energy storage intelligence agents. The differentiated comprehensive reward function configured for the market intelligence agents is as follows: In the formula: This represents the agent's overall reward value, expressed in yuan. The revenue from clearing the electricity market is equal to the sum of the products of the winning bid output and the marginal electricity price at each node; The cost of electricity generation includes fuel costs and start-up and shutdown costs; The penalty for deviation assessment is equal to the deviation between actual output and the current day's plan multiplied by the deviation assessment coefficient. This is a risk penalty item; The aforementioned risk penalty items The conditional value at risk (VaF) function is used to measure the volatility risk of returns. The specific calculation formula is as follows: In the formula: This is the risk aversion coefficient, with a value range of [value missing]. to ; For confidence level of The condition is in the risk value function; Net profit equals cleared revenue minus power generation costs; Value at risk; This indicates the operation of taking positive values; The state space of the thermal power intelligent agent additionally includes parameters such as the unit's minimum technical output, ramp rate, and cold / hot start-up time. The state spaces of the photovoltaic intelligent agent and the wind power intelligent agent include meteorological forecast data and the prediction uncertainty range; The state space of the hydropower intelligent agent additionally includes reservoir water level, water inflow forecast, and water abandonment penalty coefficient; The state space of the energy storage intelligent agent additionally includes state of charge and charge / discharge efficiency parameters, and its action space includes charging, discharging and standby operation modes.

4. The method for calibrating the probability prediction of spot electricity prices in a multi-agent game simulation for power simulation, as described in claim 1, is characterized in that... The near-end policy optimization algorithm sequentially performs iterative training in three stages: pre-training, adversarial training, and refinement. The specific process is as follows: The pre-training phase involves independently initializing the policy network parameters for various agents using historical market data, conducting independent training for at least one thousand rounds, and setting the learning rate to [value missing]. The discount factor γ is set to 0.99 to fit the benchmark pricing behavior; The adversarial training phase: all agents participate in market games synchronously, a single training round covers the complete day-ahead clearing cycle, the training is no less than 1,500 rounds, the pruning coefficient ε is set to 0.2, and the advantage function is calculated using generalized advantage estimation; The refining phase involves reducing the learning rate and continuously training for at least 1,500 rounds, while introducing an entropy regularization term to constrain policy diversity. The formula for the entropy regularization term is: In the formula: Here is the entropy regularization coefficient, and its value range is... to ; For the policy network in state Select action The probability of; This is the action space.

5. The method for calibrating the probability prediction of spot electricity prices in a multi-agent game simulation for power simulation, as described in claim 1, is characterized in that... The risk aversion coefficient and deviation assessment weight are configured differently for each market intelligent agent, specifically as follows: risk aversion coefficient of 0.1 to 0.3 and deviation assessment weight of 0.8 to 1.2 for thermal power intelligent agents; risk aversion coefficient of 0.05 to 0.15 and deviation assessment weight of 1.0 to 1.5 for new energy intelligent agents; and risk aversion coefficient of 0.2 to 0.4 and deviation assessment weight of 0.6 to 1.0 for energy storage intelligent agents.

6. The method for calibrating the probability prediction of spot electricity prices in a multi-agent game simulation for power simulation, as described in claim 1, is characterized in that... The types of intelligent agents in each market can be expanded to include virtual power plant agents and demand response agents. The state space of the virtual power plant intelligent agent includes the total capacity of aggregated resources, availability, and coordination cost; The state space of the demand response agent includes load elasticity coefficient, adjustment cost, and user satisfaction constraints.

7. The method for calibrating the probability prediction of spot electricity prices in a multi-agent game simulation for power simulation, as described in claim 1, is characterized in that... The agent policy parameters are simulated using Gaussian random perturbation, and the perturbation expression is as follows: In the formula: For the first The policy parameter vector after perturbation of each agent; For the first The original policy parameter vector after the agent has been trained; It is a random perturbation vector that follows a normal distribution; For the first The perturbation variance of the policy parameters of each agent; The disturbance variance is determined by back-calculation based on the mean square error between the Monte Carlo simulated electricity price and the actual electricity price within the historical rolling window. The calculation formula is as follows: In the formula: This is the variance scaling factor, and its value range is... to ; For the first The mean square error of historical electricity prices for each node corresponding to an intelligent agent; The Monte Carlo simulation is run at least 1,000 times, and the output probability prediction indicators include expected electricity price, decimal electricity price, median electricity price, ninetieth digit electricity price, 95% value at risk, and 95% conditional value at risk.

8. The method for calibrating the probability prediction of spot electricity prices in a multi-agent game simulation for power simulation, as described in claim 1, is characterized in that... Collect real-world electricity market operation data and construct a Bayesian variational inference update framework to select agent bid level offsets. Price elasticity coefficient Risk preference coefficient As a calibration parameter, the expression for the calibration parameter is: ; The posterior distribution of the calibration parameters is approximated using a Gaussian distribution, as shown in the following expression: In the formula: A variational approximation for the posterior distribution of the calibration parameters; The variational inference is solved by maximizing the lower bound of evidence as the optimization objective. The formula for the lower bound of evidence is: In the formula: For observational data; It is the likelihood function; It is the prior distribution; The Kullback-Leibler divergence; The mean vector and covariance matrix are updated iteratively using the stochastic gradient descent method. The calibration triggering conditions include three mechanisms: the first is periodic triggering, with a default cycle of once a week; the second is event triggering, which automatically triggers calibration when the prediction error exceeds a set threshold for three consecutive trading days; and the third is rule change triggering, which immediately triggers calibration when a market rule modification signal is detected, and adopts an enhanced calibration mode with increased sampling frequency on the first trading day after the rule change.

9. The method for calibrating the probability prediction of spot electricity prices in multi-agent game theory for power simulation according to claim 1, characterized in that, In the variational inference process, the covariance matrix adopts a diagonal matrix structure to reduce computational complexity. The specific expression is as follows: In the formula: , , These are the variance parameters for the price level offset, price elasticity coefficient, and risk preference coefficient, respectively.

10. A power simulation multi-agent game-based spot electricity price probability prediction calibration system, characterized in that, include: The simulation base construction module is used to collect basic data on grid parameters, unit parameters, market trading and settlement rules of the target electricity spot market, and to construct a simulation environment that adapts to the domestic electricity spot market mechanism and power system simulation environment based on the basic data. Based on the power system simulation environment, a simulation base that can realistically replicate the operation mechanism of the electricity market is obtained. The power system simulation environment includes a power grid physical constraint model, a clearing optimization engine, and a market settlement module. The agent training module is used to construct various types and differentiated market agents based on the obtained simulation base for the diverse trading entities in the electricity market, and to configure the state space, action space and comprehensive reward mechanism of each market agent. The module uses a proximal policy optimization algorithm to conduct multi-stage decentralized adversarial training on all market agents, constrains the diversity of agent policies, and iteratively converges to obtain a stable multi-agent game model with dynamic learning and policy game capabilities. The probability prediction module is used to apply random perturbations to the agent policy parameters based on a trained and stable multi-agent game model, simulate the random game process in the market through Monte Carlo simulation, perform statistical analysis on the simulation results, and quantitatively output the probability distribution of electricity prices and risk prediction indicators. The online calibration module is used to collect real power market operation data, construct a Bayesian variational inference update framework, perform fine-grained online calibration and update of the trained stable multi-agent game model from the dimension of agent micro-behavioral parameters, and configure a multi-scenario calibration trigger mechanism to continuously correct the deviation between the simulation model and the real market.

Images