A multi-agent-based power system transaction simulation method

Abstract

The present invention provides a method for simulating power system transactions based on multi-agent agents, which belongs to the technical field of power system simulation, and includes: collecting historical operation data of the power system, constructing a multi-agent system including power generation, flexible load and power sales agents, and performing Monte Carlo simulation and scenario generation. The method adopts a deep reinforcement learning algorithm, defines the profit function and decision model of each agent, performs market clearing calculations, and performs transaction simulation by constructing a verification scenario library. The system can evaluate agent strategies, calculate key market indicators, and feed back market information to the agent, realizing intelligent and dynamic optimization of power transaction decisions, and effectively improving the operational efficiency and transaction accuracy of the power market. The method of the present invention solves the technical problem that the existing technology generally does not fully consider the interaction and learning mechanism between market participants, resulting in a large deviation between the simulation results and the actual market performance.

Description

Technical Field

[0001] The present invention belongs to the technical field of power system simulation, and in particular relates to a power system transaction simulation method based on multi-agent agents. Background Art

[0002] With the deepening of electricity market reform and the large-scale access of new resources such as new energy, energy storage, and demand response, the transaction structure of the power system has become highly complex.

[0003] With their distributed and adaptive nature, intelligent agent systems are ideal tools for simulating electricity markets, capable of emulating the behavior and decision-making of market participants such as power generation companies and large users. Intelligent agent technology has been successfully applied in power producer bidding, demand-side management, and reliability management in electricity market transactions. Intelligent algorithms, such as Q-learning and DDPG, are used to simulate the self-learning and optimal decision-making processes of market participants to adapt to dynamic market changes. Electricity market simulation platforms, such as ElecSim, provide functions such as simulated clearing and intelligent bidding, support simulation of various market models, help market participants understand the market system, and build intelligent trading decision-making support systems. The development of these technologies has provided a solid theoretical and technical foundation for electricity market simulation and promoted the progress of electricity market reform.

[0004] Traditional methods for simulating power market transactions are primarily based on game theory and equilibrium theory. These methods typically assume that market participants are fully rational and have complete information. However, in real power markets, participants often face constraints such as incomplete information and bounded rationality, making it difficult for traditional methods to accurately simulate real market behavior. Currently, scholars at home and abroad have conducted extensive research on power market transaction modeling methods. One type of method uses Nash equilibrium theory to construct the optimal bidding strategies of market participants, such as proposing a clearing model that considers unit ramping constraints or establishing a market equilibrium model that considers transmission constraints. Another type of method uses evolutionary game theory to study the dynamic game process of market participants, such as analyzing the evolutionary stability of power producers' bidding strategies. However, these methods generally suffer from the technical problem of not fully considering the interactions and learning mechanisms between market participants, resulting in significant deviations between simulation results and actual market performance. Summary of the Invention

[0005] In view of this, the present invention provides a power system transaction simulation method based on multi-agent agents, which can solve the technical problem that the existing technology generally does not fully consider the interaction and learning mechanism between market participants, resulting in a large deviation between the simulation results and the actual market performance.

[0006] The present invention is implemented as follows: The present invention provides a multi-agent agent-based power system transaction simulation method, comprising the following steps:

[0007] S01. Collect historical operation data of the power system, wherein the historical operation data includes power generation output data, load demand data, electricity price data, and network topology data;

[0008] S02. Establishing an electricity trading environment, setting the number of market participants, the number of quotation periods, the number of capacity segments, contract data, the upper and lower quotation limits based on the historical operation data;

[0009] S03. Perform Monte Carlo simulation to generate a set of wind power output scenarios, a set of load demand scenarios, and a set of market price scenarios, and verify the set of scenarios through a probability density distribution test;

[0010] S04. Establish a multi-agent system, wherein the multi-agent system includes a power generation agent, a flexible load agent, and a power sales agent, and defines the state space, action space, strategy space, and reward function of each agent based on a partially observable Markov decision process;

[0011] S05. Define a profit function for the power generation agent, divide the power generation agent into a thermal power generation agent, a hydropower generation agent, and a wind power generation agent according to power generation type, and use the historical operation data to train a cost coefficient for each power generation agent;

[0012] S06. Construct a demand response model for a flexible load intelligent agent, and optimize the model by solving a flexible load optimization equation group, wherein the flexible load optimization equation group includes an objective function equation, a load balance equation, a comfort constraint equation, and a transfer constraint equation;

[0013] S07. Calculating input parameters of the flexible load optimization equation group through regression analysis, including a curtailment compensation cost coefficient, a transfer scheduling cost coefficient, a power purchase and sales cost coefficient, a user discomfort cost coefficient, a load adjustment upper limit, a load adjustment lower limit, and a user satisfaction threshold;

[0014] S08. Establishing a joint decision-making model for electricity purchase and sales by the electricity sales intelligent entity, wherein the joint decision-making model includes a spot market quotation strategy and a retail market pricing strategy, and sets a minimum electricity value and a maximum electricity value based on the historical operation data;

[0015] S09. Execute market clearing calculations, perform unconstrained clearing and constrained clearing based on the principle of supply and demand balance, and obtain the market clearing electricity price and market clearing electricity quantity;

[0016] S10, training the multi-agent system using a deep reinforcement learning algorithm, using the market-clearing electricity price and market-clearing electricity quantity as training data, and updating the strategy parameters of each agent;

[0017] S11. Build a verification scenario library and generate test samples based on the wind power output scenario set, load demand scenario set, and market price scenario set;

[0018] S12. Execute transaction simulation in the verification scenario library, evaluate the strategy parameters of the multi-agent system, and calculate the total market transaction volume, node electricity value, agent income value, agent profit value, and agent market share value;

[0019] S13. Publish market information and feed back the total market transaction volume, node electricity value, agent income value, agent profit value, and agent market share value to the multi-agent system for optimization of the next round of transaction decisions.

[0020] Based on the above technical solution, the multi-agent agent-based power system transaction simulation method of the present invention can also be improved as follows:

[0021] Among them, the objective function equation is used to minimize the total system cost, and the input includes reduction compensation cost, transfer scheduling cost, electricity purchase and sales cost, and user discomfort cost, and the output is the optimal load response plan. The load balance equation is used to ensure the balance of power supply and demand, and the input includes the total power on the demand side, renewable energy output, grid input power, and energy storage system output, and the output is the power balance state of each time period. The comfort constraint equation is used to ensure the user's electricity experience, and the input includes the user's original electricity demand, load adjustment amount, and user satisfaction threshold, and the output is the load adjustment range that meets the user's comfort requirements. The transfer constraint equation is used to control the rationality of load transfer, and the input includes the upper and lower limits of the load allowed to be transferred in and out in each time period, and the original load curve, and the output is a load scheduling plan that meets the transfer constraints.

[0022] Furthermore, the time period is set as follows: one day is divided into 96 time periods, corresponding to one time period every 15 minutes. The number of capacity segments is 3 to 5, and each capacity segment corresponds to a different quotation strategy. The upper and lower quotation limits are set based on historical transaction prices, with the upper limit being 1.2 times the historical highest price and the lower limit being 0.8 times the historical lowest price. The user satisfaction threshold is set to 0.8. The upper and lower limits of the load adjustment are set to plus or minus 30% of the rated load.

[0023] Furthermore, in step S03, considering the correlation between wind power, load and electricity price, the scenario generation process is specifically expressed as follows:

[0024] Wind power output scenario generation equation:

[0025] P w,s (t) = P w,f (t)+σ w (t)·ξ s (t)·nho wl (t)·L d,s (t)+σ w (t)·ξ s (t)·nho wp (t)·λ s (t);

[0026] Where, P w,s (t) is the wind power output value (MW) of scenario s in period t; P w,f (t) is the predicted wind power output value in period t (MW); σ w (t) is the standard deviation of the forecast error in period t (MW); ξ s (t) is a random variable that obeys the standard normal distribution; nho wl (t) is the correlation coefficient between wind power output and load; nho wp (t) is the correlation coefficient between wind power output and electricity price; L d,s (t) is the load demand; λ s (t) is the electricity price.

[0027] Probability constraints:

[0028] P{|P w,s (t)-P w,f (t)|≤δ w}≥α w ;

[0029] Where, δ w is the allowable prediction error range (MW); α w is the confidence level, which is generally taken as 0.95.

[0030] In step S05, the cost function of the power generation agent takes into account factors such as the start and stop status of the unit and the climbing constraint:

[0031] Cost function of thermal power generation agent:

[0032]

[0033] Where, u g (t) is the start / stop status of the unit (0 / 1); ST g is the startup cost (yuan); T on is the continuous running time (h); β g is the efficiency reduction coefficient; ΔP g is the output change (MW); γg is the climbing cost coefficient.

[0034] Constraints:

[0035] P g (t)-P g (t-1)≤RU g ·u g (t);

[0036] P g (t-1)-P g (t)≤RD g ·u g (t);

[0037] Where, RU g ,RD g They are upper and lower ramp rate limits (MW / h) respectively.

[0038] In step S06, the flexible load optimization equations take into account the mutual influence between loads:

[0039] Objective function equation:

[0040]

[0041] Where, μ i is the interaction coefficient of load i; is the adjustment amount of load i in time period t (MW); θ is the discomfort sensitivity coefficient.

[0042] Load balance equation:

[0043]

[0044] Where, η res (),η grid (),η ess () are efficiency functions of renewable energy, power grid and energy storage respectively; SOC t The state of charge of the energy storage system.

[0045] In step S08, the decision-making model of the electricity sales agent takes into account market competition and user response:

[0046] Quotation strategy:

[0047]

[0048] Where, ω k is the impact factor of competitor k; MS k is the market share of competitor k; For adjustable load capacity; is the total load capacity.

[0049] Retail market pricing strategy:

[0050] Where, PD t is the current electricity demand density; PD ref is the reference demand density.

[0051] Principal component constraints:

[0052] Q T Q = I;

[0053] Where Q is the orthogonal transformation matrix; z1 and z2 are the principal component scores.

[0054] In step S09, the market clearing process takes network constraints into account:

[0055] Objective function:

[0056] Where, is the quote variance.

[0057] Power flow constraints:

[0058] P mn =B mn (δ m -δ n );

[0059]

[0060] Where, P mn is the power flow between nodes m and n; B mn is the node admittance; δ m ,δ n is the node phase angle; are the node power generation and load respectively.

[0061] Probabilistic constraints:

[0062]

[0063] Where, is the line transmission capacity; α f is the confidence level.

[0064] Variable coupling constraints:

[0065]

[0066] Where H is the coupling matrix, which describes the mutual influence relationship among power generation, electricity price and load.

[0067] The variables or functions involved in each equation are explained in detail below.

[0068] 1. In the wind power output scenario generation equation:

[0069] Correlation coefficient nho wl (t),nho wp How to obtain (t):

[0070]

[0071] Where, P w,i ,L d,i is the historical data sample point; is the average value of historical data; n is the number of samples.

[0072] Confidence level α w Determination of:

[0073] Based on the statistical analysis of historical data, the probability distribution function F(δ) of the forecast error is calculated so that F(δ w )=α w .

[0074] 2. In the cost function of thermal power generation agent:

[0075] Efficiency reduction coefficient β g How to obtain:

[0076] Where, η t is the power generation efficiency at time t; η0 is the rated power generation efficiency; T is the statistical period.

[0077] Climbing cost coefficient γ g Determination of:

[0078] Where, ΔC g The additional cost caused by ramping; ΔP g is the output change; P g,rated For rated output.

[0079] 3. In the flexible load optimization equation:

[0080] Interaction influence coefficient μ i How to obtain:

[0081] Where, is the cross elasticity of the load response; For rated load.

[0082] Efficiency function η res (),η grid (),η essThe expression of ():

[0083]

[0084] Where, is the rated efficiency; α res ,β grid ,α ess is the efficiency attenuation coefficient; is the rated power; SOC ref is the reference state of charge.

[0085] 4. In the electricity sales agent decision model:

[0086] Competitor influence factorω k How to obtain:

[0087] Where, is the sensitivity of market share to electricity price; MS i The market share of the retailer; k Competitor electricity prices.

[0088] Construction of orthogonal transformation matrix Q:

[0089]

[0090] Where, σ 12 is the covariance between the quoted price and the retail price; are the variances of the quoted price and the retail price, respectively.

[0091] 5. During the market clearing process:

[0092] Quote variance Calculation:

[0093] Where, C i,j,t is the historical quotation data; It is the average of historical quotes.

[0094] Construction of coupling matrix H:

[0095]

[0096] The matrix elements are obtained through regression analysis:

[0097] Where, y i is the output variable (power generation); x j are input variables (electricity price, load); For the working point.

[0098] Node admittance B mn Calculation:

[0099] Where, X mn is the line reactance, obtained through actual measurement or equipment parameters.

[0100] Compared with the existing technology, the beneficial effects of the power system transaction simulation method based on multi-agent agents provided by the present invention are as follows: First, by constructing a scenario generation model that considers the correlation between wind power output, load demand and market price, the coupling relationship of multiple uncertain factors is accurately portrayed. Second, by introducing dynamic characteristics such as unit efficiency decay and ramping cost, a more realistic power generation agent cost function is established. Third, by designing an optimization equation group that considers the interaction between loads, accurate modeling of flexible load agents is achieved. Fourth, by constructing a power sales agent decision-making model based on market competition and user response, the rationality of the retail market pricing strategy is improved. Fifth, by introducing probability constraints and principal component constraints, the reliability of market clearing results is guaranteed. It effectively solves the technical problem that the existing technology generally does not fully consider the interaction and learning mechanism between market participants, resulting in a large deviation between simulation results and actual market performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0101] Figure 1 A flow chart of the method provided by the present invention;

[0102] Figure 2 Schematic diagram of the composition of the multi-intelligent agent system in an embodiment of the present invention. DETAILED DESCRIPTION

[0103] like Figure 1 FIG. 1 is a flow chart of a multi-agent agent-based power system transaction simulation method provided by the present invention. The specific implementation methods of the steps of the present invention are described in detail below:

[0104] The specific implementation of step S01 is to collect power system operating data in real time through a data acquisition system. First, distributed data collectors are used to collect real-time output data from each power plant, with a sampling period of 5 minutes. The data includes information such as active power, reactive power, and generator set operating status. Secondly, smart meters are deployed to collect user load data with a sampling period of 15 minutes, recording parameters such as user power consumption, power factor, voltage and current. Thirdly, real-time electricity price data for each time period is obtained from the power trading center, including day-ahead market prices, real-time market prices, and frequency modulation market prices. Finally, topological data of the transmission network is collected, including busbar node data, line impedance parameters, and transformer ratio data. The purpose of this step is to establish a historical database to provide data support for subsequent scenario generation and model training.

[0105] The specific implementation of step S02 is to initialize the configuration of the power trading environment based on the collected historical data. First, the number of market participants is set according to the number of power generation companies, power sales companies, and power users that have historically participated in the transaction, generally ranging from 50 to 200. Secondly, the number of quotation periods is set according to the actual trading rules, usually 96 periods, corresponding to one period every 15 minutes. Thirdly, the number of capacity segments is set, usually 3 to 5 segments, each capacity segment corresponds to a different quotation strategy. Then, contract data is set according to long-term agreements and bilateral contracts, including electricity consumption and electricity price information for monthly and annual contracts. Finally, the upper and lower quotation limits are set according to historical transaction prices, generally with the upper limit being 1.2 times the historical highest price and the lower limit being 0.8 times the historical lowest price. The purpose of this step is to build a trading environment that conforms to actual conditions.

[0106] The specific implementation of step S03 is to use the Latin hypercube sampling method to generate random scenarios. First, the probability distribution space of wind power, user load, and market electricity price is divided into equally probable subspaces, and the number of samples is generally 1000. Secondly, considering the correlation between variables, the Nataf transform is used to convert the correlated random variables into independent random variables. Then, the inverse transform method is used to generate samples that conform to the marginal distribution. The generated scenarios are then tested for probability density distribution using the chi-square test method, with a confidence level set to 0.95. Finally, valid scenarios are selected based on the test results to form the final scenario set. The purpose of this step is to generate a statistically representative scenario sample.

[0107] The specific implementation of step S04 involves constructing a multi-agent system based on a partially observable Markov decision process. First, the state space of each agent is defined, including its own state and observable environmental state. The state vector dimension is generally 10 to 20. Next, the action space is defined, including continuous and discrete variables such as power generation adjustment, load response, and electricity price quotes. Finally, the policy space is defined, using a random policy network to map the probability distribution of states to actions. Finally, a reward function is designed, taking into account economic benefits, technical constraints, and market rules. The weight coefficient of the reward function is determined through expert experience. The purpose of this step is to establish the decision-making model framework of the agent.

[0108] The specific implementation of step S05 involves establishing benefit function models for different types of power generation agents. First, a polynomial fitting method is used to train the fuel cost coefficients of thermal power generation units, taking into account startup and shutdown costs, ramping costs, and efficiency decay. Next, an opportunity cost model for hydropower generation is established, taking into account reservoir scheduling and cascade optimization. Finally, a forecast error cost model for wind power generation is constructed, accounting for the penalty costs associated with output uncertainty. Finally, a least squares method is used to perform regression analysis on the historical operating data of each type of generator to determine the coefficients of the cost function. The goal of this step is to accurately describe the economic characteristics of the power generation agent.

[0109] The specific implementation of step S06 involves building a flexible load demand response model that takes user experience into consideration. First, an objective function is established, including compensation reduction costs, transfer scheduling costs, electricity purchase and sales costs, and user discomfort costs. Next, power balance constraints are established to ensure a balance between generation and consumption during each time period. Comfort constraints are established to set the allowable range of load adjustment based on user type. Finally, load transfer constraints are established to ensure that the total amount of electricity is conserved before and after load transfer. The purpose of this step is to achieve optimal scheduling of flexible loads.

[0110] The specific implementation of step S07 is to determine the key parameters of the flexible load optimization equations through data analysis. First, a multivariate regression analysis method is used to calculate the curtailment compensation cost coefficient, which generally ranges from 0.5 to 2 yuan per kilowatt-hour. Second, the transfer scheduling cost coefficient is calculated based on historical response data, which generally ranges from 0.2 to 1 yuan per kilowatt-hour. Third, the purchase and sale cost coefficient is determined based on market transaction data, with reference to real-time electricity prices. Then, a questionnaire survey is conducted to obtain the user discomfort cost coefficient, which generally ranges from 1 to 5 yuan per kilowatt-hour. Finally, the upper and lower limits of the load adjustment are set according to the equipment parameters, generally within the range of plus or minus 30 percentage points of the rated load, and the user satisfaction threshold is set at 0.8. The purpose of this step is to ensure the practicality of the optimization model.

[0111] The specific implementation of step S08 involves establishing a market transaction decision-making model for the power retailer. First, a spot market bidding strategy is constructed that considers market competition, using adversary modeling to predict competitor bidding behavior. Second, a retail market pricing strategy is established that considers user response, using a nonlinear response function to describe user sensitivity to electricity prices. Finally, a floor and ceiling electricity value are set based on historical data, typically 0.8 and 1.2 times the average electricity price, respectively. Finally, principal component analysis is used to establish a coupling relationship between the bidding and pricing strategies. The goal of this step is to optimize the trading decisions of power retailers.

[0112] Step S09 involves performing a market-clearing calculation based on DC power flow. First, an unconstrained clearing calculation is performed, determining the initial clearing result based on the price-first principle. Second, a constrained clearing calculation is performed, taking network constraints into account, using the interior point method to solve the DC power flow optimization problem. Finally, the line power flow distribution is calculated based on the node injection power to check whether the line transmission capacity constraints are met. Finally, the final clearing price and cleared electricity quantity are determined based on the supply-demand balance principle. The goal of this step is to obtain a market-clearing result that satisfies physical constraints.

[0113] Step S10 is specifically implemented by training the agent using a deep reinforcement learning algorithm based on a dominance strategy. First, a value network and a policy network are constructed, using a multilayer perceptron as a function approximator. Next, an experience replay mechanism is used to store transaction data, with the replay pool capacity set to 10,000 entries. The policy network parameters are updated using a policy gradient method with a learning rate set to 0.001. Finally, the value network parameters are updated using a temporal difference algorithm with a discount factor set to 0.95. This step aims to improve the agent's decision-making ability.

[0114] The specific implementation of step S11 involves constructing a validation scenario library for testing. First, cluster analysis is performed on the set of wind power output scenarios, using a Gaussian mixture model to classify the scenarios into three to five categories. Next, feature extraction is performed on the set of load demand scenarios, using principal component analysis for dimensionality reduction. Finally, probability distribution analysis is performed on the set of market price scenarios, using kernel density estimation to fit the price distribution. Finally, test samples are selected based on the typicality and representativeness of the scenario features. The purpose of this step is to construct an effective validation dataset.

[0115] Step S12 is implemented by conducting a transaction simulation test in a verification scenario. First, the total market electricity volume is calculated, and the supply-demand matching for each time period is analyzed. Next, the electricity value of each node is calculated, and the spatial distribution of prices is analyzed. Finally, the revenue of each agent is calculated, including revenue from power generation, sales, and demand response. Finally, the profit and market share of each agent are calculated to assess its competitiveness. This step aims to verify the effectiveness of the agent's strategy.

[0116] The specific implementation of step S13 involves establishing a market information feedback mechanism. First, total market transaction volume information is released to all market participants for use in assessing the overall supply and demand situation. Second, node electricity value information is released to market participants in relevant regions for price signal analysis. Third, agent revenue and profit information is fed back to each agent for strategy evaluation and adjustment. Finally, market share information is released to regulatory agencies for market power monitoring. The goal of this step is to promote healthy market development.

[0117] In a specific implementation of step S01, data collection involves the following mathematical representation:

[0118] 1) Mathematical expression of active power acquisition:

[0119] Among them, P active (t) represents the three-phase active power at time t (MW), U L (t) represents the effective value of the line voltage (kV), I L (t) represents the effective value of the line current (kA), and cosφ(t) represents the power factor.

[0120] 2) Mathematical expression of load data collection:

[0121] Among them, P load (t) represents the total load at time t (MW), P i (t) represents the load power of the i-th user (MW), k i (t) represents the load participation factor, and n represents the total number of users.

[0122] 3) Statistical processing of electricity price data:

[0123] where λ avg (t) represents the weighted average electricity price at time t (yuan / MWh), λ i (t) represents the electricity price in the i-th market (yuan / MWh), V i (t) represents the corresponding transaction electricity volume (MWh), and m represents the number of market types.

[0124] In a specific implementation of step S02, the environment configuration involves the following mathematical representation:

[0125] 1) Calculation of market participant capacity:

[0126]

[0127] Among them, Cap total Indicates the total transaction capacity (MW), represents the capacity of the i-th generator (MW), represents the capacity of the jth load (MW), represents the k-th electricity retailer capacity (MW).

[0128] 2) Quote upper and lower limits: λ max =αmax(λ hist )+β·std(λ hist ),

[0129] λ min =γ·min(λ hist )-δ·std(λ hist ),

[0130] where λ max ,λ min Respectively represent the upper and lower limits of the quotation (yuan / MWh), λ his Represents the historical price series, α, β, γ, δ are adjustment coefficients, and std() represents the standard deviation function.

[0131] The specific implementation of step S03 also involves the following mathematical model:

[0132] 1) Latin Hypercube Sampling:

[0133] where x ij represents the jth sample point of the i-th variable, represents the inverse function of the marginal distribution, π j represents the permutation of the jth interval, u ij represents a uniformly distributed random number between [0,1], and n represents the number of samples.

[0134] 2) Nataf transformation: y i =Φ -1 (F i (x i )),

[0135] where y i represents the transformed standard normal variable, Φ -1 Represents the inverse function of the standard normal distribution, F i Represents the marginal distribution function of the original variable.

[0136] 3) Scenario probability density test:

[0137] Among them O i represents the observation frequency, E i represents the theoretical frequency, and k represents the number of intervals.

[0138] The specific implementation of step S04 also involves the following decision model:

[0139] 1) State transition probability:

[0140] Where N(s t ,a t ,s t+1 ) represents the state transition count, N(s t ,a t ) represents the number of occurrences of state-action pairs.

[0141] 2) Strategy function:

[0142] where Q θ (s t ,a t ) represents the state-action value function, and θ represents the policy network parameters.

[0143] 3) Reward calculation:

[0144]

[0145] profit tIndicates current income, penalty t Indicates technical breach penalty, compliance t Indicates compliance with market rules.

[0146] The specific implementation of step S05 also involves the following cost model:

[0147] 1) Calculation of efficiency attenuation coefficient:

[0148] where η t represents the power generation efficiency at time t, η0 represents the rated power generation efficiency, and T represents the statistical period.

[0149] 2) Determination of climbing cost coefficient:

[0150] where ΔC g represents the additional cost caused by ramping, ΔP g Indicates the change in output force, P g,rated Indicates rated output.

[0151] The specific implementation of step S06 also involves the following optimization model:

[0152] 1) Load interaction coefficient:

[0153] in represents the cross elasticity of the load response, Indicates rated load.

[0154] 2) Equipment efficiency function:

[0155]

[0156] in Indicates rated efficiency, α res ,β grid ,α ess Indicates the efficiency reduction coefficient, SOC ref Indicates the reference state of charge.

[0157] The specific implementation of step S07 also involves the following parameter model:

[0158] 1) Calculation of reduction compensation cost coefficient:

[0159] where c r represents the reduction compensation cost coefficient (yuan / MWh), α r represents the compensation coefficient, λ t represents the electricity price in period t (yuan / MWh), Indicates load reduction (MW).

[0160] 2) Transfer scheduling cost coefficient:

[0161] where c s represents the transfer scheduling cost coefficient (yuan / MWh), β s represents the adjustment coefficient, It represents the load amount (MW) transferred from period t to period τ.

[0162] 3) Discomfort cost coefficient:

[0163] where c d Discomfort cost coefficient (yuan / MW 2 ), γ d represents the sensitivity coefficient, w i Indicates the user importance weight.

[0164] The specific implementation of step S08 also involves the following decision model:

[0165] 1) Competitor influence factors:

[0166] in Indicates the sensitivity of market share to electricity price, MS i represents the market share of the retailer, λ k Indicates competitor electricity prices.

[0167] 2) Orthogonal transformation matrix:

[0168]

[0169] where σ 12 represents the covariance between the quote and the retail price, They represent the variances of the quoted price and the retail price respectively.

[0170] The specific implementation of step S09 also involves the following clearing model:

[0171] 1) Quote variance calculation:

[0172] Among them C i,j,t Represents historical quote data, Indicates the average value of historical quotes.

[0173] 2) Coupling matrix construction:

[0174]

[0175] The matrix elements are obtained through regression analysis:

[0176] where y i represents the output variable (power generation), x j represents input variables (electricity price, load), Indicates the working point.

[0177] The specific implementation of step S10 also involves the following learning algorithm:

[0178] 1) Policy gradient update:

[0179] where π θ represents a parameterized strategy, represents the action-value function.

[0180] 2) Value network loss function:

[0181] Where V φ represents the parameterized value function and γ represents the discount factor.

[0182] 3) Experience replay update:

[0183]

[0184] Where α represents the learning rate, N represents the batch size, and δ i Represents the timing difference error.

[0185] The specific implementation of step S11 sets the following scene model:

[0186] 1) Gaussian mixture model:

[0187] where π k represents the weight of the kth Gaussian component, μ k ,∑ k denote the mean vector and covariance matrix respectively.

[0188] 2) Principal component analysis: X = U∑V T ,

[0189] Where X represents the original data matrix, U, V represent the left and right singular vectors, and ∑ represents the singular value matrix.

[0190] Steps S12 and S13 mainly involve result evaluation and information release, and also include the following indicator calculations:

[0191] 1) Market concentration index:

[0192] Where HHI represents the Herfindahl-Hirschman Index, MS i represents the market share of market entity i.

[0193] 2) Price discovery efficiency:

[0194] Among them E price represents the price discovery efficiency, represents the clearing price, Indicates the reference price, Represents the mean reference price.

[0195] Specifically, the principles of this invention are as follows: First, in the scenario generation phase, a conditional probability distribution is used to characterize the correlation between wind power, load, and electricity prices. By introducing a correlation coefficient matrix, a mapping relationship between variables is established, overcoming the flaw of traditional independent sampling methods that ignores variable correlations. Furthermore, probabilistic constraints are used to limit the prediction error range, ensuring the rationality of the generated scenarios. Second, in the modeling of the power generation agent, a cost function that considers efficiency decay is constructed based on thermodynamic principles and equipment characteristics. The efficiency decay coefficient is obtained through regression of historical operating data, and the ramp cost coefficient is calculated based on incremental costs. This method effectively reflects the dynamic characteristics of the unit. Third, in the modeling of flexible loads, a multi-objective optimization equation system is designed based on user utility theory. The load interaction influence coefficient is introduced to describe the correlation effect between different user groups. This coefficient is calculated by calculating the cross-elasticity of load response. The efficiency function adopts a piecewise nonlinear form, accurately reflecting the operating characteristics of the equipment under different operating conditions. Fourth, in the decision-making of the power sales agent, a competition model is constructed based on market microstructure theory. By calculating the sensitivity of market share to electricity price, the competitor influence factor is obtained, achieving a quantitative description of the market competition situation. The logistic function is used to characterize the nonlinear relationship between electricity price and demand, and principal component analysis is used to ensure the orthogonality of decision variables. Fifth, in the market clearing phase, an optimization model that considers network constraints is constructed based on the node admittance matrix. By introducing a coupling matrix, a mapping relationship between power generation, electricity price, and load is established. This matrix is ​​obtained through local linearization and regression analysis. Finally, a deep reinforcement learning algorithm is used to train the intelligent agent. Through value function approximation and policy gradient methods, online learning and policy optimization of the intelligent agent are achieved. This approach overcomes the low computational efficiency of traditional optimization algorithms and can adapt to dynamic changes in the market environment.

[0196] The following provides a specific embodiment 1 of the present invention. The specific implementation of each step in this embodiment 1 is described in detail as follows: The specific implementation of step S01 is to collect power system operation data in real time through a distributed data acquisition system, and use a SCADA system (data acquisition and supervision control system) to collect power generation unit data every 5 minutes, wherein the active power is collected using the three-phase power calculation formula Parameter P active(t) represents the three-phase active power at time t, in megawatts, U L (t) is the effective value of the line voltage, in kV, I L (t) is the effective value of the line current in kiloamperes, cosφ(t) is the power factor, and load data is collected through smart meters with a sampling period of 15 minutes. The load aggregation calculation formula is used. Data processing, where P load (t) represents the total load at time t, in megawatts, P i (t) represents the load power of the i-th user, k i (t) represents the load participation factor, n represents the total number of users, and the system obtains real-time electricity price information by subscribing to the data interface of the power trading center and adopts the weighted average calculation formula Handling multiple market electricity prices, where λ avg (t) represents the weighted average electricity price at time t, in RMB per MWh, λ i (t) represents the electricity price in the i-th market, V i (t) represents the corresponding transaction volume, m represents the number of market types, and network topology data is collected, including busbar node numbers, branch impedance parameters, transformer ratios, and other information. A relational database is used for storage and management. The collected data is quality checked, including data integrity, validity, and consistency tests. Outliers are removed and data is completed. Missing data is processed using linear interpolation. The purpose of this step is to build a complete historical database to provide data support for subsequent modeling analysis and scenario generation.

[0197] The specific implementation of step S02 is to initialize the configuration of the power trading environment based on historical data, and use the total balance calculation formula Determine the number of market participants, including Cap total represents the total transaction capacity, represents the capacity of the i-th generator, represents the j-th load quotient capacity, represents the capacity of the kth electricity seller. The number of participants is generally between 50 and 200. According to the trading rules, a day is divided into 96 periods for quotation. Each period is 15 minutes. The power generation capacity is divided into 3 to 5 quotation segments. The historical price statistics method is used to set the upper and lower limits of the quotation. The upper limit calculation formula is λ max =α·max(λ hist )+β·std(λ hist ), the lower limit calculation formula is λ min =γ·min(λ hist )-δ·std(λ hist ), where λ max ,λ minRespectively represent the upper and lower limits of the quotation, λ his represents the historical price series, α, β, γ, and δ are adjustment coefficients, which are generally 1.2, 0.2, 0.8, and 0.2 respectively. std() represents the standard deviation function. A fixed trading volume is set according to bilateral agreements and long-term contracts. The contract coverage rate is generally 60 to 80 percent of the total trading volume. The purpose of this step is to build a standardized and orderly trading environment.

[0198] The specific implementation of step S03 is to use the Monte Carlo method to generate transaction scenarios. First, the Latin hypercube sampling algorithm is used to generate the basic scenario. The scenario generation formula is: where x ij represents the jth sample point of the i-th variable, represents the inverse function of the marginal distribution, π j represents the permutation of the jth interval, u ij represents a random number uniformly distributed from 0 to 1, n represents the number of samples, and the Nataf transformation is used to process the variable correlation. The transformation formula is y i =Φ -1 (F i (x i )), where y i represents the transformed standard normal variable, Φ -1 Represents the inverse function of the standard normal distribution, F i Represents the marginal distribution function of the original variable. When generating wind power output scenarios, the correlation with load and electricity price is considered. The scenario generation formula is P w,s (t) = P w,f (t)+σ w (t)·ξ s (t)·ρ wl (t)·L d,s (t)+σ w (t)·ξ s (t)·ρ wp (t)·λ s (t), where the meanings of the parameters are the same as above. The chi-square test method is used to verify the probability distribution of the scenario. The test statistic calculation formula is: Among them O i represents the observation frequency, E i represents the theoretical frequency, k represents the number of intervals, and the significance level is set to 0.05. The purpose of this step is to generate representative trading scenarios.

[0199] The specific implementation of step S04 is to construct a multi-agent system based on a partially observable Markov decision process, define the state space of each agent including the local observable state and the environment state, and use the state transition probability calculation formula Describe the state evolution, where N(st ,a t ,s t+1 ) represents the state transition count, N(s t ,a t ) represents the number of occurrences of state-action pairs, and a continuous action space is defined to describe power generation regulation, load response and electricity price quotation. A random strategy based on a strategy network is adopted, and the strategy function is where Q θ (s t ,a t ) represents the state action value function, θ represents the policy network parameters, and the design of the comprehensive reward function profit t Indicates current income, penalty t Indicates technical breach penalty, compliance t Indicates the degree of compliance with market rules. The weight coefficients w1, w2, and w3 are 0.5, 0.3, and 0.2 respectively. The purpose of this step is to establish the decision-making model framework of the intelligent agent.

[0200] The specific implementation of step S05 is to establish a profit function model for different types of power generation agents, taking into account the dynamic characteristics and constraints of the power generation unit. For thermal power generation units, a cost function that considers the start-stop state and climbing constraints is adopted. Among them, P g,t represents the power output at time t, u g,t Indicates the start and stop status of the unit at time t, a g ,b g ,c g represents the fuel cost coefficient, obtained by least squares fitting, ST g represents the startup cost, β g Indicates the efficiency attenuation coefficient, using the efficiency statistical formula Calculate, where η t represents the power generation efficiency at time t, η0 represents the rated power generation efficiency, and the ramp cost coefficient γ g Using the cost analysis formula Determine that for hydropower generating units, establish an opportunity cost model considering reservoir operation Among them, P h represents hydropower output, V represents the water storage capacity of the reservoir, V max represents the maximum storage capacity, α h ,β h Represents the opportunity cost coefficient. For wind turbines, a prediction error cost model C is constructed. w (P w )=α w ·|Pw -P w,f |+β w max(0,P w,f -P w ), where P w Indicates the actual wind power output, P w,f represents the wind power forecast value, α w ,β w represents the penalty cost coefficient. The operating constraints of each generator set include upper and lower output limits, ramp rate, and minimum start and stop time. The purpose of this step is to establish an accurate power generation cost model.

[0201] The specific implementation of step S06 is to build a flexible load demand response model that takes into account user experience, using a multi-objective optimization method, with the objective function being in Indicates the load reduction amount, represents the load transfer amount, represents the amount of electricity purchased, c r ,c s ,c p They represent the reduction, transfer, and electricity purchase cost coefficients, θ represents the discomfort sensitivity coefficient, and the load balance constraint is in Indicates the actual load, represents the base load, They represent the incoming and outgoing loads respectively. The equipment efficiency constraint is described by piecewise function. The renewable energy efficiency function is: The grid efficiency function is The energy storage efficiency function is The meaning of each parameter is the same as above. The purpose of this step is to achieve optimal load scheduling.

[0202] The specific implementation of step S07 is to determine the key parameters of the flexible load optimization equation group through data analysis, and use multiple regression analysis to calculate the reduction compensation cost coefficient. The calculation formula is: where α r It represents the compensation coefficient, which is generally between 1.2 and 1.5. The formula for calculating the transfer scheduling cost coefficient is: where β s It represents the adjustment coefficient, which is generally between 0.8 and 1.2. The calculation formula of the discomfort cost coefficient is: where γ d It represents the sensitivity coefficient, which is generally set at 1.5 to 2.5. The upper and lower limits of the load adjustment are set to plus or minus 30 percentage points of the rated load, and the user satisfaction threshold is set to 0.8. The purpose of this step is to determine the parameter values ​​of the optimization model.

[0203] The specific implementation of step S08 is to establish a market transaction decision model for the power selling intelligent body. The quotation strategy adopts a dynamic pricing method considering market competition. The quotation function is in Indicates the quoted value. represents the base electricity price, Indicates the quotation adjustment coefficient and the competitor influence factor ω k The elastic analysis method is used for calculation, and the calculation formula is: in Represents the sensitivity of market share to electricity prices. The retail market adopts a dynamic pricing strategy based on user demand. The pricing function is Among them PD t Indicates the current power demand density, PD ref represents the reference demand density, α represents the price response coefficient, which is generally between 2 and 5. The principal component analysis method is used to establish the coupling relationship between the quotation strategy and the pricing strategy. The coupling matrix is: in σ 12 Represents the covariance between the bid price and the retail price. The purpose of this step is to optimize the transaction strategy of the retailer.

[0204] The specific implementation of step S09 is to perform market clearing calculation considering network constraints, using a social welfare maximization method based on DC power flow, with the objective function being in It represents the quote variance, and the calculation formula is The power flow constraints include the power balance equation P mn =B mn (δ m -δ n ) and the node power injection equation Among them B mn represents the node admittance, δ m ,δ n represents the node phase angle, They represent the node power generation and load respectively. The interior point method is used to solve the optimization problem. The convergence accuracy is set to 0.001 and the maximum number of iterations is 1000. The purpose of this step is to obtain the market clearing result that meets the physical constraints.

[0205] The specific implementation of step S10 is to use the deep reinforcement learning algorithm to train the intelligent agent and use the policy gradient method to update the policy network parameters. The gradient calculation formula is: where π θ represents a parameterized strategy, Represents the action value function, the value network is updated using the temporal difference algorithm, and the loss function is Where V φ represents the parameterized value function, γ represents the discount factor, which is set to 0.95. The experience replay mechanism is used to store transaction data. The update formula is: Where α represents the learning rate, which is set to 0.001, and N represents the batch size, which is set to 32. The purpose of this step is to improve the decision-making ability of the agent.

[0206] The specific implementation of step S11 is to build a verification scenario library, cluster the wind power output scenarios using a Gaussian mixture model, and the probability density function is where π k represents the weight of the kth Gaussian component, μ k ,∑ k Denote the mean vector and covariance matrix respectively. The expectation maximization algorithm is used for parameter estimation. The principal component analysis is used to reduce the dimension of the load demand scenario. The matrix decomposition form is X = U∑V T , where X represents the original data matrix, U and V represent the left and right singular vectors, and the principal components with cumulative variance contribution greater than 85% are selected. The kernel density estimation is used to fit the probability distribution of the market price scenario. The purpose of this step is to construct an effective validation data set.

[0207] The specific implementation of step S12 is to conduct a transaction simulation test under the verification scenario and calculate the market concentration index using the Herfindahl Hirschman Index, which is calculated as follows: Among them, MS i Represents the market share of market entity i and evaluates price discovery efficiency. The calculation formula is: Among them E price represents the price discovery efficiency, represents the clearing price, Represents the reference price, calculates the agent's revenue value, profit value and market share value, and evaluates its competitiveness. The purpose of this step is to verify the effectiveness of the agent's strategy.

[0208] The specific implementation of step S13 involves establishing a market information feedback mechanism, categorizing and organizing information such as total market transaction volume, node electricity prices, agent revenue, and market share. This information is stored and published in a real-time database, with an update cycle of 5 minutes. Data security is ensured through encrypted transmission. Information access levels are set based on the permissions of market participants, enabling hierarchical management of market information. This step aims to promote the healthy development of the market. By organically combining these steps, a complete power system transaction simulation method is constructed, enabling intelligent decision-making by market participants and effective market operation.

[0209] The following provides Example 2 of a specific application scenario of the present invention: This example uses a provincial power market as an example to illustrate the specific application process of the multi-agent-based power system transaction simulation method. This provincial power market includes 75 power generators, 120 power sales companies, and 3,000 large users. The total installed capacity of the units is 85,000 MW, including 45,000 MW of thermal power units, 25,000 MW of hydropower units, and 15,000 MW of wind power units. The annual electricity demand is 380 billion kWh.

[0210] First, the data acquisition system collected power system operating data. Data collected from the generating units showed an average utilization of 4,500 hours, an equivalent availability factor of 0.92 for thermal power units, 0.98 for hydropower units, and an average capacity factor of 0.25 for wind turbines. The dynamic characteristics of the generating units are shown in Table 1.

[0211] Table 1: Dynamic characteristic parameters of generator sets

[0212] Unit type Minimum technical output Maximum climbing rate (per minute) Minimum start-stop time (h) Cold start time (h) thermal power 40％ 2％ 4 6 hydropower 20％ 15％ 0.5 1 wind power 0 - - -

[0213] Data collected by smart meters show that industrial users account for 65 percent of total electricity consumption, commercial users account for 25 percent, and residential users account for 10 percent. User load characteristic data are shown in Table 2.

[0214] Table 2: User load characteristic data table

[0215] User Type Load rate Power Factor Adjustable load ratio Electricity price sensitivity Industrial users 75％ 0.92 30％ 0.8 Business users 65％ 0.85 20％ 0.6 Residential users 55％ 0.95 10％ 0.3

[0216] Historical price statistics show that the current market average electricity price is 485 yuan per megawatt-hour, the real-time market average electricity price is 520 yuan per megawatt-hour, and the frequency regulation market average electricity price is 580 yuan per megawatt-hour. The price fluctuation characteristics are shown in Table 3.

[0217] Table 3: Market price fluctuation characteristics

[0218]

[0219]

[0220] Next, we will build the electricity trading environment. The total number of market participants will be set at 200, including 75 power generation companies, 120 power sales companies, and 5 large-scale users. The day will be divided into 96 quotation periods, with power generation capacity divided into three quotation segments: base load, intermediate load, and peak load, with capacity percentages of 60%, 30%, and 10%, respectively. The quotation ceiling will be set at 850 yuan per megawatt-hour, and the floor at 280 yuan per megawatt-hour. The contracted electricity volume coverage ratio will be 70%.

[0221] Monte Carlo methods were used to generate trading scenarios. Latin hypercube sampling was used to generate 1,000 basic scenarios, considering a correlation coefficient of 0.65 between wind power output and load demand and a correlation coefficient of 0.45 with market electricity prices. Probability density distribution tests were used to select 300 valid scenarios for simulation analysis.

[0222] Construct a multi-agent system. The state space dimension of the power generation agent is 15, including information such as unit output, start and stop status, and operating time. The action space dimension is 4, including output adjustment and quotation adjustment. The state space dimension of the flexible load agent is 10, including information such as load power, electricity price signal, and comfort. The action space dimension is 3, including load reduction, transfer, and response time. The state space dimension of the power sales agent is 12, including information such as contracted electricity, market share, and competitor prices. The action space dimension is 2, including wholesale market quotations and retail market pricing. In this embodiment, transactions are all implemented in a centralized bidding clearing manner. In the centralized bidding market, each agent selects a bidding strategy with the goal of maximizing its own interests, and submits the selected bidding strategy to the trading center. The trading center processes the bidding and returns the clearing results to the agent. The agent then determines the next trading strategy based on the clearing results. The constructed multi-agent system is as follows: Figure 2As shown, the system master control module is the backend support module, capable of simulating various network topologies and setting up market environments based on simulation training needs. Environment settings include the number of bidding periods and capacity segments for generators, the number of bidding periods and capacity segments for buyers, the number of market participants, contract data, and bid and offer limits. It can simulate unilateral or bilateral bidding markets and provide default bids, taking into account contracts as needed, allowing for diverse market environments. The master control module also serves as the interface for interaction between the various submodules on the instructor's desk. The bid processing submodule controls the power market bidding process, which includes three phases: bid start, bid in progress, and bid end. After the backend opens the bidding market, multiple agents submit data and submit it to the designated server database via a communication interface. The market transaction submodule performs both unconstrained and constrained clearing of power market transactions. Unconstrained clearing follows the unified clearing market mechanism and the principle of supply and demand equilibrium. Generators queue up to sell electricity based on bids from lowest to highest, while buyers queue up to purchase electricity based on bids from highest to lowest. The calculated unconstrained clearing data is used to plot an unconstrained clearing curve, showing the unconstrained market-clearing price (MCP) and market-clearing quantity (MCQ), visually illustrating the algorithmic principles of unconstrained transaction clearing. Constrained clearing incorporates line transmission constraints on top of the unconstrained calculation. After the calculation is complete, the system annotates market data and system parameters, such as flow direction, node electricity prices, generator capacity utilization, line utilization, and transformer utilization, on the simulated system clearing network diagram. Furthermore, the network diagram uses different background colors at nodes based on the range of node electricity prices, visually highlighting the differences in system node electricity prices. The settlement and evaluation submodule performs post-market transaction settlement, including contract processing. Public settlement data includes total market electricity volume, social welfare, and balancing node transaction prices; private settlement data includes individual market participant revenue, profit, market share, surplus, and ranking. Distinguishing between unconstrained and constrained market transactions yields two sets of settlement results for evaluation. The information release submodule publishes the transaction, settlement and other information that are allowed to be made public by the backend to the frontend for reinforcement learning training of the multi-agent model.

[0223] The revenue function model is established for different types of power generation agents. The fuel cost coefficient a of the thermal power unit is 0.00325 yuan per kilowatt-hour squared, b is 0.285 yuan per kilowatt-hour, c is 21.5 yuan per kilowatt-hour, and the startup cost is 45 yuan per kilowatt of the unit's rated capacity. The opportunity cost coefficient α of the hydropower unit is h is 0.15, β h The wind turbine prediction error penalty coefficient α is 0.08. w is 0.35, β w is 0.25.

[0224] A flexible load demand response model was constructed. The curtailment compensation cost coefficient was 1.2 yuan per kilowatt-hour, the shift scheduling cost coefficient was 0.8 yuan per kilowatt-hour, and the discomfort cost coefficient was 2.5 yuan per kilowatt-hour. The upper limit of load adjustment was 30 percent of the rated load, and the lower limit was -30 percent of the rated load. The user satisfaction threshold was set at 0.8. Among the equipment efficiency coefficients, the renewable energy efficiency attenuation coefficient was 0.15, the grid efficiency attenuation coefficient was 0.12, and the energy storage efficiency attenuation coefficient was 0.18.

[0225] A market transaction decision model for the electricity sales agent was established. The base electricity price was set at 500 yuan per megawatt-hour, with a bid adjustment coefficient range of plus or minus 15 percent. The competitor impact factor was derived through market share sensitivity analysis, with an average value of 0.35. The retail market price response coefficient was set at 3.5, with a floor price of 380 yuan per megawatt-hour and a ceiling price of 680 yuan per megawatt-hour.

[0226] Perform market clearing calculations. Solve the DC power flow optimization problem using the interior point method, with a convergence precision of 0.001 and a maximum number of 1000 iterations. Consider the 62 main transmission lines of the transmission network, with capacity constrained to within 90 percent of rated capacity. Balance the node injection power within a tolerance of 0.1 MW.

[0227] The agent was trained using a deep reinforcement learning algorithm. Both the policy network and the value network employed a three-layer neural network structure, with 128, 64, and 32 hidden neurons, respectively. The activation function used was the ReLU function. The learning rate was set to 0.001, the discount factor to 0.95, the experience replay pool to 10,000 entries, and the batch size to 32. Training was performed for 1,000 epochs, each consisting of 960 trading sessions.

[0228] Construct a validation scenario library. Use a four-Gaussian component mixture model to cluster wind power output scenarios, extract five principal components for feature expression in load demand scenarios, and use kernel density estimation to fit the probability distribution of market price scenarios. Select 30 typical scenarios to form a test set.

[0229] Trading simulations were conducted under a validation scenario. The market's total daily electricity transaction volume reached 120,000 MWh, of which 75,000 MWh were generated through contracts and 45,000 MWh were generated through the spot market. The average electricity price at each node was 510 RMB per MWh, with the largest node price difference reaching 75 RMB per MWh. The average rate of return for power generation agents was 12%, the average rate of return for power sales agents was 8%, and the average rate of return for flexible loads through demand response was 5%. The market concentration index was 0.15, and the price discovery efficiency was 0.92.

[0230] Establish a market information feedback mechanism. Market transaction data, including node electricity prices, traded electricity volume, and unit output, is updated every five minutes. Confidentiality is divided into three tiers, with corresponding access rights granted to market participants. This example demonstrates that the power system transaction simulation method proposed in this invention can effectively simulate the trading behavior of market participants and enable scientific evaluation of market mechanisms.

[0231] The above description is only a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with this technical field can easily think of changes or replacements within the technical scope disclosed by the present invention, which should be covered by the scope of protection of the present invention.

Claims

1. A multi-agent agent-based power system transaction simulation method, characterized in that: The following steps are involved: S01. Collect historical operation data of the power system, wherein the historical operation data includes power generation output data, load demand data, electricity price data, and network topology data; S02. Establishing an electricity trading environment, setting the number of market participants, the number of quotation periods, the number of capacity segments, contract data, the upper and lower quotation limits based on the historical operation data; S03. Perform Monte Carlo simulation to generate a set of wind power output scenarios, a set of load demand scenarios, and a set of market price scenarios, and verify the set of scenarios through a probability density distribution test; S04. Establish a multi-agent system, wherein the multi-agent system includes a power generation agent, a flexible load agent, and a power sales agent, and defines the state space, action space, strategy space, and reward function of each agent based on a partially observable Markov decision process; S05. Define a profit function for the power generation agent, divide the power generation agent into a thermal power generation agent, a hydropower generation agent, and a wind power generation agent according to power generation type, and use the historical operation data to train a cost coefficient for each power generation agent; S06. Construct a demand response model for a flexible load intelligent agent, and optimize the model by solving a flexible load optimization equation group, wherein the flexible load optimization equation group includes an objective function equation, a load balance equation, a comfort constraint equation, and a transfer constraint equation; S07. Calculating input parameters of the flexible load optimization equation group through regression analysis, including a curtailment compensation cost coefficient, a transfer scheduling cost coefficient, a power purchase and sales cost coefficient, a user discomfort cost coefficient, a load adjustment upper limit, a load adjustment lower limit, and a user satisfaction threshold; S08. Establishing a joint decision-making model for electricity purchase and sales by the electricity sales intelligent entity, wherein the joint decision-making model includes a spot market quotation strategy and a retail market pricing strategy, and sets a minimum electricity value and a maximum electricity value based on the historical operation data; S09. Execute market clearing calculations, perform unconstrained clearing and constrained clearing based on the principle of supply and demand balance, and obtain the market clearing electricity price and market clearing electricity quantity; S10, training the multi-agent system using a deep reinforcement learning algorithm, using the market-clearing electricity price and market-clearing electricity quantity as training data, and updating the strategy parameters of each agent; S11. Build a verification scenario library and generate test samples based on the wind power output scenario set, load demand scenario set, and market price scenario set; S12. Execute transaction simulation in the verification scenario library, evaluate the strategy parameters of the multi-agent system, and calculate the total market transaction volume, node electricity value, agent income value, agent profit value, and agent market share value; S13. Publish market information and feed back the total market transaction volume, node electricity value, agent income value, agent profit value, and agent market share value to the multi-agent system for optimization of the next round of transaction decisions.

2. A multi-agent agent-based power system transaction simulation method according to claim 1, characterized in that: The objective function equation is used to minimize the total system cost. The input includes reduction compensation cost, transfer scheduling cost, electricity purchase and sales cost, and user discomfort cost. The output is the optimal load response plan.

3. The multi-agent agent-based power system transaction simulation method according to claim 1, characterized in that: The load balance equation is used to ensure the balance between power supply and demand. The inputs include the total power on the demand side, the output of renewable energy, the power input to the grid, and the output of the energy storage system. The output is the power balance status at each time period.

4. The multi-agent agent-based power system transaction simulation method according to claim 1, characterized in that: The comfort constraint equation is used to ensure the user's electricity consumption experience. The input includes the user's original electricity demand, load adjustment amount, and user satisfaction threshold. The output is the load adjustment range that meets the user's comfort requirements.

5. The multi-agent agent-based power system transaction simulation method according to claim 1, characterized in that: The transfer constraint equation is used to control the rationality of load transfer. The input includes the upper and lower limits of loads allowed to be transferred in and out in each time period and the original load curve. The output is a load scheduling plan that meets the transfer constraints.

6. The multi-agent agent-based power system transaction simulation method according to claim 1, characterized in that: The time periods are set as follows: one day is set to 96 time periods, corresponding to one time period every 15 minutes.

7. The multi-agent agent-based power system transaction simulation method according to claim 1, characterized in that: The number of capacity segments is 3 to 5, and each capacity segment corresponds to a different quotation strategy.

8. The multi-agent agent-based power system transaction simulation method according to claim 1, characterized in that: The upper limit of the quotation is 1.2 times the historical highest transaction price, and the lower limit of the quotation is 0.8 times the historical lowest transaction price.

9. The multi-agent agent-based power system transaction simulation method according to claim 1, characterized in that: The user satisfaction threshold is set to 0.

8.

10. The multi-agent agent-based power system transaction simulation method according to claim 1, characterized in that: The upper and lower limits of the load adjustment are set to plus or minus 30% of the rated load.

Images