A method, medium and system for determining behavioral trends of power market entities

Abstract

The present invention provides a method, medium and system for determining the behavioral trend of power market entities, which belongs to the technical field of power market entity behavioral trend analysis, including: realizing accurate prediction of complex behaviors of the power market by comprehensively applying reinforcement learning, deep neural networks and data analysis technology. The study first collects historical operating data of the power market, and performs multi-dimensional time series decomposition to extract the stability and fluctuation characteristics of prices and quotations. By constructing a five-tuple model without a reward function, combining expert knowledge and principal component analysis, iteratively optimizes the reward function model, and establishes a group of market entity behavior equations. Subsequently, a deep neural network with an encoder-decoder structure is used, combined with the maximum likelihood estimation method, to construct a probability distribution model of market entity behavior. Ultimately, this method can accurately predict the behavioral trends of market entities, solving the problem in the existing technology that it is difficult to accurately grasp the behavioral trends of power market entities.

Description

Technical Field

[0001] The present invention belongs to the technical field of power market subject behavior trend analysis, and specifically relates to a method, medium and system for determining the behavior trend of power market subjects. Background Art

[0002] Analyzing the behavior of power market participants is a crucial foundation for power market operation and regulation, and is crucial for maintaining market order, optimizing market structure, and improving market efficiency. With the continuous deepening of power market reform, the trading behaviors of market participants are becoming increasingly complex, and traditional methods for analyzing their behavior are no longer sufficient. Currently, the main methods used for analyzing the behavior of power market participants both domestically and internationally include game theory, econometrics, statistical analysis, and machine learning.

[0003] Game theory methods, based primarily on the assumptions of complete rationality and complete information, construct the utility functions and strategy spaces of market players and seek Nash equilibrium points. However, in real markets, the application of game theory methods is limited by factors such as incomplete information, complex market structures, and the dynamic and uncertain nature of player behavior. Econometric methods analyze the relationship between market player behavior and various influencing factors by establishing regression models, but this method struggles to capture the nonlinear characteristics and dynamic evolution of player behavior. Statistical analysis methods summarize the behavioral patterns of market players through statistical analysis of historical data, but lack in-depth exploration of the underlying mechanisms and have limited predictive capabilities. In summary, existing technologies struggle to accurately grasp the behavioral trends of power market players. Summary of the Invention

[0004] In view of this, the present invention provides a method, medium and system for determining the behavioral trends of power market entities, which can solve the problem in the prior art that it is difficult to accurately grasp the behavioral trends of power market entities.

[0005] The present invention is achieved in that:

[0006] The present invention provides a method for determining the behavior trend of power market entities, comprising the following steps:

[0007] S01. Collect historical data on power market operation, including market clearing price data, subject bidding strategy data, actual transaction volume data, system load demand data, unit technical constraint data, network security constraint data, and market share data;

[0008] S02. Perform multi-dimensional time series decomposition on the market clearing price data and the subject quotation strategy data to obtain price stability component data, price fluctuation component data, quotation stability component data, and quotation fluctuation component data;

[0009] S03. Constructing a five-tuple model without a reward function using the price stability component data, the price fluctuation component data, the quotation stability component data, and the quotation fluctuation component data, and establishing a state space matrix and an action space matrix;

[0010] S04. Performing Monte Carlo sampling based on the state space matrix and the action space matrix to obtain an initial state sample set;

[0011] S05. For the initial state sample set, construct an initial reward function model using expert knowledge, and calculate the cumulative reward expectation value and the characteristic expectation function;

[0012] S06. Performing dimensionality reduction processing on the market clearing price data, the subject quotation strategy data, and the actual transaction volume data using a principal component analysis method, extracting key feature vectors, and calculating actual feature expected values;

[0013] S07. Compare the actual feature expectation value with the feature expectation function to obtain a feature expectation deviation, train an inverse reinforcement learning model based on the feature expectation deviation, iteratively update the initial reward function model until the feature expectation deviation is less than a preset threshold, and obtain an optimized reward function model;

[0014] S08. Use the optimized reward function model to construct a market subject behavior equation group.

[0015] S09. Construct a deep neural network model, wherein the deep neural network model adopts an encoder-decoder structure, including a feature extraction module, a temporal encoding layer, an attention layer, and a strategy decoding layer;

[0016] S10. Based on the market entity behavior equations and the deep neural network model, a maximum likelihood estimation method is used to solve unknown parameters and establish a market entity behavior probability distribution model;

[0017] S11. Predicting the behavioral trends of market entities based on the market entity behavior probability distribution model, and generating a market entity behavior trend prediction report.

[0018] On the basis of the above technical solution, the method for determining the behavior trend of power market entities of the present invention can also be improved as follows:

[0019] Among them, the market subject behavior equation group includes utility maximization equation, strategy optimization equation, market clearing equation and constraint condition equation.

[0020] Furthermore, the utility maximization equation is used to calculate the profit function of the market entity, the input includes the entity's bidding strategy data, the market clearing price data, and power generation cost data, and the output is the expected profit value.

[0021] Furthermore, the strategy optimization equation is used to update the optimal bidding strategy of the market entity, the input includes the entity bidding strategy data, the competing entity bidding data, and the system load demand data, and the output is the optimal bidding strategy data.

[0022] Furthermore, the market clearing equation is used to calculate the market equilibrium price, the input includes the subject quotation strategy data, the system load demand data, and the network security constraint data, and the output is the market clearing price forecast value.

[0023] Furthermore, the constraint condition equation is used to ensure the feasibility of the strategy, the input includes the unit technical constraint data and the network security constraint data, and the output is the constraint judgment result data.

[0024] Furthermore, the feature extraction module is used to extract market subject behavior feature data, and the input is the market clearing price data, the subject quotation strategy data, and the actual transaction electricity data.

[0025] Furthermore, the time series coding layer is used to capture the time series characteristics of market behavior and perform time series coding on the market subject behavior feature data; the attention layer is used to identify key market influencing factors and calculate feature weights; and the strategy decoding layer is used to generate market subject behavior prediction results.

[0026] Furthermore, multi-dimensional time series decomposition specifically adopts nonlinear time series decomposition:

[0027] ;

[0028] ;

[0029] Where, is the time series of market clearing prices; To stabilize the price; is the price fluctuation component; is the price decomposition error term; is the quote strategy time series; To stabilize the quotation weight; is the quotation fluctuation component; Decompose error terms for quotes; is the time series decomposition weight coefficient, which is obtained by least squares fitting.

[0030] Furthermore, the state space evolution equation:

[0031] ;

[0032] Where, is the state vector; is the state transfer matrix; is the nonlinear adjustment coefficient; is the state noise vector.

[0033] Among them, the state vector Defined as:

[0034] ;

[0035] Where, is the benchmark value of the corresponding variable, which is obtained by calculating the mean of historical data.

[0036] Furthermore, the reward function coupled equations:

[0037] ;

[0038] ;

[0039] Where, is the state value function; is the cumulative reward function; is the time discount factor; is the attenuation coefficient; is the noise term of the value function; is the state gradient operator; is the Laplace operator.

[0040] Furthermore, the market subject behavior coupling equations include:

[0041] (1) Utility maximization partial differential equation:

[0042] ;

[0043] Where, is the utility function; is the diffusion coefficient; is the nonlinear coupling term; is the utility function noise.

[0044] (2) Strategy optimization coupling equation:

[0045] ;

[0046] Where, For quotation strategy; The optimal strategy for the goal; is the optimization coefficient; Optimize noise for the strategy.

[0047] (3) Market clearing dynamics equation:

[0048] ;

[0049] Where, is the market price; is total demand; For the The output of each subject; is the market adjustment coefficient; is price noise.

[0050] (4) Constraint coupling equation:

[0051] ;

[0052] Where, is the transmission constraint function; is the network node coordinate; is the network sensitivity coefficient.

[0053] Method for obtaining equation coefficients:

[0054] 1. Obtained through regression analysis of historical data;

[0055] 2. The value range is [0.9, 0.99], determined by cross-validation;

[0056] 3. The value range is [0.01, 0.1], determined according to the market cycle characteristics;

[0057] 4. All noise terms are assumed to be zero-mean bounded noise.

[0058] Explanation of the principle of equation construction:

[0059] 1. Time series decomposition uses partial differential form, which can better characterize the dynamic coupling relationship between variables;

[0060] 2. Introducing nonlinear terms into state space evolution improves the model’s ability to describe abnormal states;

[0061] 3. The reward function adopts an integral form, taking into account the historical cumulative effect;

[0062] 4. The utility maximization equation introduces a second-order derivative term to characterize the accelerating change characteristics of utility;

[0063] 5. The strategy optimization equation includes diffusion terms, driving terms, and nonlinear coupling terms, which enhances the adaptability of the model;

[0064] 6. The market clearing equation considers the nonlinear coupling of supply and demand equilibrium and strategic interaction;

[0065] 7. The constraints are in the form of Poisson equations to better describe the spatial distribution characteristics of network constraints.

[0066] A second aspect of the present invention provides a computer-readable storage medium, wherein the computer-readable storage medium stores program instructions, and when the program instructions are run in a computer, they are used to execute the above-mentioned method for determining the behavioral trend of power market entities.

[0067] A third aspect of the present invention provides a system for determining behavioral trends of power market entities, which includes the above-mentioned computer-readable storage medium.

[0068] Compared with the prior art, the method, medium, and system for determining the behavior trend of power market entities provided by the present invention have the following beneficial effects:

[0069] First, the present invention employs a nonlinear time series decomposition equation, introducing partial differential and coupling terms to enable multi-timescale analysis of market behavior. This equation not only accurately separates long-term trends and short-term fluctuations in behavior, but also captures the interplay between different timescales.

[0070] Secondly, the state-space evolution equation constructed by this invention enhances its ability to describe the dynamic evolution of market states by introducing nonlinear geometric terms. This equation maintains the Markov characteristics of state transitions while capturing the geometric characteristics of the state space through nonlinear terms, enabling the model to better adapt to changes in the market environment.

[0071] Third, the reward function coupled equations proposed in this paper innovatively combine the continuous-time Bellman equations with nonlinear partial differential equations, enabling a precise description of the decision-making process of market entities. This system not only considers the gradient information of the immediate reward but also characterizes the curvature of the value function through second-order terms, significantly improving the model's generalization capabilities.

[0072] Fourth, the utility maximization partial differential equation designed by this invention effectively captures the complex dynamics of market participants' decision-making behavior by introducing multiple nonlinear coupling terms. This equation not only conforms to the utility maximization principle in economics, but also describes the acceleration effect and nonlinear characteristics of the decision-making process.

[0073] Therefore, the present invention solves the problem in the prior art that it is difficult to accurately grasp the behavioral trends of power market players. BRIEF DESCRIPTION OF THE DRAWINGS

[0074] Figure 1 A flow chart of the method provided by the present invention. DETAILED DESCRIPTION

[0075] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0076] like Figure 1 FIG. 1 is a flow chart of a method for determining the behavior trend of power market entities provided by the first aspect of the present invention. The method includes the following steps:

[0077] S01. Collect historical data on power market operation, including market clearing price data, subject bidding strategy data, actual transaction volume data, system load demand data, unit technical constraint data, network security constraint data, and market share data;

[0078] S02. Perform multi-dimensional time series decomposition on the market clearing price data and the subject quotation strategy data to obtain price stability component data, price fluctuation component data, quotation stability component data, and quotation fluctuation component data;

[0079] S03. Using the price stability component data, the price fluctuation component data, the quotation stability component data, and the quotation fluctuation component data, a five-tuple model without a reward function is constructed, and a state space matrix and an action space matrix are established;

[0080] S04. Perform Monte Carlo sampling based on the state space matrix and the action space matrix to obtain an initial state sample set;

[0081] S05. For the initial state sample set, use expert knowledge to build an initial reward function model and calculate the cumulative reward expectation value and characteristic expectation function;

[0082] S06. Perform dimensionality reduction processing on the market clearing price data, the subject quotation strategy data, and the actual transaction volume data using the principal component analysis method, extract key feature vectors, and calculate the actual feature expected value;

[0083] S07. Compare the actual feature expectation value with the feature expectation function to obtain the feature expectation deviation, train the inverse reinforcement learning model based on the feature expectation deviation, iteratively update the initial reward function model until the feature expectation deviation is less than a preset threshold, and obtain the optimized reward function model;

[0084] S08. Use the optimized reward function model to construct the market subject behavior equation group.

[0085] S09. Build a deep neural network model. The deep neural network model adopts an encoder-decoder structure, including a feature extraction module, a temporal encoding layer, an attention layer, and a strategy decoding layer.

[0086] S10. Based on the market subject behavior equations and deep neural network model, the maximum likelihood estimation method is used to solve the unknown parameters and establish a market subject behavior probability distribution model;

[0087] S11. Based on the market subject behavior probability distribution model, predict the behavior trend of market subjects and generate a market subject behavior trend prediction report.

[0088] The specific implementation of the above steps is described in detail below:

[0089] The specific implementation method of step S01 is to comprehensively collect the historical operation data of the power market through multiple data acquisition modules. First, a market data acquisition interface is established, and the market clearing price data is collected through the real-time database system with a collection frequency of once every 15 minutes, and the data timestamp is recorded at the same time; then the quotation strategy data of each market player is collected through the application interface of the power trading platform, including quotation curve, quotation amount and time information; then the actual transaction power data is collected, including the planned power generation and actual power generation in different time periods; then the power system load demand data is collected, including the day-ahead forecast load and real-time load data; the technical constraint data of the generator set is continued to be collected, including parameters such as the maximum and minimum output of the unit, the climbing rate limit, and the minimum start and stop time; then the power grid safety constraint data is collected, including the transmission line limit, the node voltage limit, etc.; finally, the market share data of each market player is collected, including indicators such as the installed capacity share and the power generation share. All collected data are standardized and stored in the time series database. The purpose of this step is to provide complete historical data support for subsequent analysis.

[0090] The specific implementation method of step S02 is to perform multi-dimensional time series decomposition on the collected market clearing price data and subject quotation strategy data. This step adopts a nonlinear time series decomposition method, and describes the dynamic evolution process of prices and quotations by constructing a group of partial differential equations. First, the price time series is decomposed into a stable component and a fluctuating component, wherein the stable component reflects the long-term trend of the price, and the fluctuating component reflects the short-term price fluctuation characteristics; then a similar decomposition is performed on the quotation strategy time series to obtain the stable component and the fluctuating component of the quotation; then the least squares method is used to determine the weight coefficient in the time series decomposition equation, and the optimal weight value is determined by minimizing the decomposition error; then the decomposition result is subjected to error analysis to ensure that the decomposition error is within an acceptable range, and the relative error is generally required to be no more than 5%; finally, the decomposed component data are normalized and stored. The purpose of this step is to decompose complex market data into components that are easier to analyze.

[0091] The specific implementation method of step S03 is to construct a five-tuple model without a reward function, which includes a state space, an action space, a state transition probability, a time discount factor and a terminal state set. First, the state vector is defined, including dimensions such as price stability component, price fluctuation component, quotation stability component, quotation fluctuation component, demand and cost; then the state space matrix is ​​constructed, and the matrix dimension is determined by the state vector dimension; then the action space is defined, including an optional quotation strategy set; then a state transition probability matrix is ​​established to describe the probability distribution of the system transferring to the next state given the current state and action; then the time discount factor is set, with a value range of 0.9 to 0.99; finally, the terminal state set is determined, including situations such as market clearing failure or convergence to a stable state. The purpose of this step is to establish a mathematical description framework for the behavior of market entities.

[0092] The specific implementation method of step S04 is to perform Monte Carlo sampling based on the constructed state space matrix and action space matrix. First, a suitable sampling algorithm is selected. The importance sampling method can be used to improve sampling efficiency. Then the number of samples is determined, generally not less than 1000 samples to ensure statistical significance. Then, the state space is randomly sampled to generate initial state samples. Then, action sampling is performed on each initial state to form state-action pairs. The representativeness of the sampling is then evaluated to ensure that the samples cover the main areas of the state space. Finally, the sampling results are organized into a standard format and stored. The purpose of this step is to obtain a statistically significant set of initial state samples to provide data support for subsequent reward function learning.

[0093] The specific implementation method of step S05 is to use expert knowledge to construct an initial reward function model. First, based on the experience of electricity market experts, the key factors affecting the decision-making of market entities are determined; then the basic form of the reward function is constructed, including benefit terms, cost terms and constraint penalty terms; then the instantaneous reward value is calculated for each sample, and the cumulative reward expectation value is calculated considering the time discount effect; then the characteristic expectation function is constructed to describe the long-term performance of the system under a given strategy; then the characteristic expectation function is normalized so that its value range is between 0 and 1; finally, the rationality of the reward function is verified to ensure that it can correctly guide the decision-making behavior of market entities. The purpose of this step is to establish an initial reward evaluation mechanism.

[0094] The specific implementation method of step S06 is to extract features through the principal component analysis method. First, the market clearing price data, the main quotation strategy data and the actual transaction electricity data are standardized and preprocessed; then the covariance matrix of the data is calculated to analyze the correlation between the variables; then the eigenvalues ​​and eigenvectors of the covariance matrix are solved and sorted by the size of the eigenvalues; then the eigenvectors with a cumulative contribution rate of 95% are selected as the main features; then the original data is projected into the main feature space to obtain the feature representation after dimensionality reduction; finally, the actual feature expected value is calculated as a benchmark for evaluating the effectiveness of the strategy. The purpose of this step is to extract key features in the data and reduce the complexity of subsequent calculations.

[0095] The specific implementation method of step S07 is to train the inverse reinforcement learning model and iteratively optimize the reward function. First, the feature expectation deviation is calculated, that is, the difference between the actual feature expectation value and the feature expectation function; then the loss function is constructed based on the feature expectation deviation, and the reward function parameters are optimized using the gradient descent method; then the intelligent agent is trained using a reinforcement learning algorithm such as the policy gradient method to update the behavior strategy; then the effect of the updated strategy is evaluated and the new feature expectation deviation is calculated; then it is determined whether the feature expectation deviation is less than a preset threshold, generally set to 0.01; if it is greater than the threshold, the iterative optimization continues, and if it is less than the threshold, the iteration is stopped and the optimized reward function model is output. The purpose of this step is to obtain a reward function that accurately reflects the decision-making preferences of market entities.

[0096] The specific implementation method of step S08 is to construct a group of market subject behavior equations, including a utility maximization equation, a strategy optimization equation, a market clearing equation and a constraint condition equation. First, a utility maximization equation is constructed based on the optimized reward function to describe the behavior of market subjects in pursuit of profit maximization; then a strategy optimization equation is constructed to describe the process of market subjects adjusting their quotation strategies according to market conditions; then a market clearing equation is established to describe the market price formation mechanism; then a constraint condition equation is constructed to ensure that the strategy meets physical and market constraints; then the boundary conditions and initial conditions of the equation group are set; finally, a numerical solution method is used to solve the equation group to obtain the dynamic evolution process of the market subject behavior. The purpose of this step is to establish a mathematical model of the market subject behavior.

[0097] The specific implementation method of step S09 is to construct a deep neural network model using an encoder-decoder structure. First, a feature extraction module is designed, and a convolutional neural network is used to extract the spatiotemporal features of market data; then a temporal encoding layer is constructed, and a long short-term memory network is used to capture the temporal dependencies of the data; then an attention layer is implemented to calculate the importance weights of different features; then a strategy decoding layer is designed to map the encoded features into behavior prediction results; then the hyperparameters of the network are determined, including the number of layers, the number of neurons, the learning rate, etc.; finally, the backpropagation algorithm is used to train the network, and an early stopping strategy is used to avoid overfitting. The purpose of this step is to establish a deep learning model that can accurately predict the behavior of market entities.

[0098] The specific implementation method of step S10 is to use the maximum likelihood estimation method to solve the unknown parameters. First, a likelihood function is constructed to describe the probability of occurrence of observed data; then the logarithm of the likelihood function is taken to convert it into a log-likelihood function; then the gradient of the log-likelihood function with respect to the unknown parameters is calculated; then the stochastic gradient ascent method is used to optimize the parameter value; then the accuracy of the parameter estimation is verified and the confidence interval of the estimated value is calculated; finally, a market entity behavior probability distribution model is established to describe the probability of occurrence of different behavioral strategies. The purpose of this step is to obtain accurate parameter estimates of the behavioral model.

[0099] The specific implementation method of step S11 is to generate a market entity behavior trend forecast report. First, the future behavior trend of the market entity is predicted based on the probability distribution model; then the uncertainty of the prediction result is analyzed and the prediction interval is calculated; then the influence of different factors on the behavior trend is evaluated; then a visual chart is generated to intuitively display the prediction results; then a prediction report is written, including a description of the prediction method, result analysis, risk warnings, etc.; finally, corresponding decision-making recommendations are put forward to provide reference for market supervision and policy formulation. The purpose of this step is to convert the analysis results into understandable and actionable information.

[0100] By implementing the above steps, we can accurately predict the behavioral trends of power market participants, providing important decision-making support for market operations and regulation. This method, which integrates advanced algorithms such as time series decomposition, reinforcement learning, and deep learning, is highly practical and scalable. In practical applications, relevant parameters and thresholds can be adjusted according to specific market conditions to achieve better forecasting results.

[0101] The principles for establishing the following equations are explained in detail below.

[0102] 1. The process of establishing the time series decomposition equation:

[0103] First consider the traditional timing decomposition:

[0104] ;

[0105] Introducing time dynamics:

[0106] ;

[0107] Considering the coupling relationship between components, the final form is obtained:

[0108] ;

[0109] Derivation principle:

[0110] 1. Stable component The effect of price changes is linear express;

[0111] 2. The influence of the fluctuating component is through its rate of change reflect;

[0112] 3. How to obtain : Solve by minimizing the sum of squares of prediction errors:

[0113] .

[0114] 2. Derivation of state space evolution equation:

[0115] Starting from the linear state equation: ;

[0116] Considering nonlinear effects: ;

[0117] Introducing the geometric characteristics of the state space, we get:

[0118] ;

[0119] Derivation principle:

[0120] 1. The matrix structure is:

[0121] ;

[0122] 2. Estimate by Kalman filter method:

[0123] ;

[0124] 3. The value range is [0.1, 0.5] and is determined by cross-validation.

[0125] 3. Derivation of the reward function coupling equations:

[0126] Starting from the basic Bellman equation: ;

[0127] Introducing the continuous-time form: ;

[0128] Considering the second-order effects, we get: ;

[0129] Derivation principle:

[0130] 1. First-order terms Gradient information describing the immediate reward;

[0131] 2. Second-order terms Characterize the curvature characteristics of the value function;

[0132] 3. The integral form takes into account the historical cumulative effect.

[0133] 4. Derivation of the utility maximization partial differential equation:

[0134] Starting from the static utility function: ;

[0135] Introducing dynamic features: ;

[0136] Consider the second-order effects of prices and quotes: ;

[0137] Finally, the nonlinear coupling term is introduced:

[0138] ;

[0139] Derivation principle:

[0140] 1. Obtained by maximum likelihood estimation:

[0141] ;

[0142] 2. Using neural network fitting:

[0143] ,in is the network parameter.

[0144] A second aspect of the present invention provides a computer-readable storage medium, wherein the computer-readable storage medium stores program instructions, and when the program instructions are run in a computer, they are used to execute the above-mentioned method for determining the behavioral trend of power market entities.

[0145] A third aspect of the present invention provides a system for determining behavioral trends of power market entities, which includes the above-mentioned computer-readable storage medium.

[0146] Specifically, the principle of this invention is to model the behavior of power market entities as a nonlinear dynamic system and describe the dynamic evolution of the system by constructing multiple coupled partial differential equations. The rationality of this modeling method is mainly reflected in the following aspects:

[0147] First, from an economic perspective, the decision-making behavior of market entities is essentially a dynamic optimization process. The utility maximization partial differential equation in this invention is derived based on the principle of utility maximization in economics. By introducing second-order derivatives and nonlinear coupling terms, this equation not only maintains the basic characteristics of economic principles but also describes the complex dynamic characteristics of real markets.

[0148] Secondly, from the perspective of information theory, the evolution of market player behavior is a process of continuous information accumulation and updating. The state-space evolution equations of this invention utilize nonlinear geometric characteristic terms to effectively capture the cumulative effect of state information. The integral form of the reward function coupled equations reflects the impact of historical information on current decisions.

[0149] Thirdly, from a systems science perspective, the behavior of market entities is a complex system with multiple timescales and multiple coupled levels. By constructing multiple coupled partial differential equations, this paper establishes connections between different timescales and levels, achieving an accurate description of the overall characteristics of the system.

[0150] The logic of this invention's approach is as follows: First, nonlinear time series decomposition is used to achieve multi-scale separation of behavioral characteristics; second, state-space evolution equations are used to describe the system's dynamic characteristics; third, reward function coupled equations are used to characterize the value assessment of the decision-making process; and fourth, the utility maximization equation is used to accurately model decision-making behavior. These four steps, layered and progressive, form a complete analytical framework.

[0151] A specific embodiment 1 of the present invention is provided below. The specific implementation methods of the steps in this embodiment 1 are described in detail as follows: The specific implementation method of step S01 is to achieve comprehensive collection of historical data on electricity market operation by establishing a distributed data acquisition system. The system adopts a three-layer architecture design, including a data acquisition layer, a data processing layer and a data storage layer; the data acquisition layer collects market clearing price data through a real-time data acquisition interface with a sampling period of 15 minutes. The interface adopts an open data access protocol and supports breakpoint resumption function; the data processing layer standardizes the original data and eliminates outliers. The judgment standard for outliers is 3 times the standard deviation of the price during the same period of the previous day. The processed data is stored in a unified data format; the data storage layer uses a distributed time series database to store the processed data, and the data retention period is 3 years; the purpose of this step is to establish a complete market data acquisition system to provide data support for subsequent analysis.

[0152] The specific implementation of step S02 is to perform multi-dimensional time series decomposition on the market clearing price data and the subject quotation strategy data, using a nonlinear time series decomposition method based on the following partial differential equation: ,in represents the market clearing price time series, represents the price stabilization component, represents the price fluctuation component, represents the price decomposition error term, and Decompose the weight coefficient for time series; decompose the quotation strategy data at the same time: ,in represents the quote strategy time series, Indicates the stable component of the quotation, Indicates the quotation fluctuation component, represents the quote decomposition error term, and Decompose the weight coefficients for the quotes; use the least squares method to determine the weight coefficients to minimize the decomposition error; the purpose of this step is to decompose complex market data into components that are easier to analyze.

[0153] The specific implementation of step S03 is to construct a five-tuple model without a reward function. First, a state space matrix is ​​constructed based on the state vector ,in is the baseline value of the corresponding variable; then construct the state space evolution equation: ,in is the state transition matrix, is the nonlinear adjustment coefficient, is the state noise vector; finally, the action space matrix is ​​established, which contains all optional quotation strategy combinations; the purpose of this step is to establish a mathematical description framework for the behavior of market entities.

[0154] The specific implementation of step S04 is to perform Monte Carlo sampling based on the constructed state space matrix and action space matrix, using the importance sampling method to improve sampling efficiency, with a sample size of no less than 1000; the state transition probability is considered during the sampling process to ensure the representativeness and diversity of the sample; the sampling results are statistically analyzed to calculate the mean, variance and distribution characteristics of the sample; the purpose of this step is to obtain a statistically significant set of initial state samples. The calculation formula for the importance sampling weight is: Where, For the The importance weight of each sample; is the target distribution probability density function; is the probability density function of the proposed distribution; is the sample point obtained by sampling.

[0155] Furthermore, the state transition probability matrix The specific expressions are as follows: Where, Indicates the slave state Transfer to state The probability of .

[0156] Calculation method of state transition probability matrix: Where, To observe from the state Transfer to state the number of times; is the state space dimension.

[0157] Calculation of state transition probability considering time decay effect: Where, For the moment The state transition probability of is the initial moment; is the time attenuation coefficient, and its value range is [0.01, 0.1].

[0158] Adjustment of state transition probability by introducing market volatility factors: Where, is the adjusted state transition probability; is the market sensitivity coefficient, and its value range is [0.1,1]; An indicator of market volatility.

[0159] Normalization of the state transition probability matrix: Where, is the normalized state transition probability.

[0160] Smoothing of the state transition probability matrix: Where, is the probability matrix after smoothing; is the smoothing coefficient, and its value range is [0.1, 0.3]; is a uniformly distributed matrix with all elements being .

[0161] Uncertainty evaluation of state transition probability matrix: Where, For the The entropy value of the row is used to evaluate the uncertainty of the transition probability.

[0162] Update rule for the state transition probability matrix: Where, for The probability matrix of the moment; for The transition frequency matrix observed at each moment; To update the weight, the value range is [0.7, 0.9].

[0163] The principles of these matrix calculation methods are explained as follows: 1. The basic transition probability adopts the frequency statistical method to reflect the historical laws of state transition; 2. The time decay term takes into account the timeliness of historical data, and recent data has a higher reference value; 3. Market volatility adjustment reflects the impact of the external environment on state transition; 4. Normalization processing ensures the validity of probability; 5. Smoothing processing avoids the zero probability problem and improves the robustness of the model; 6. Entropy value evaluation provides a quantitative indicator of uncertainty; 7. The dynamic update mechanism realizes the adaptive adjustment of the model.

[0164] Parameter acquisition method: 1. Obtained through historical data statistics; 2. Determined based on market data update cycle; 3. Determined through market sensitivity analysis; 4. Determined through cross-validation; 5. Determined based on model performance evaluation.

[0165] The specific implementation of step S05 is to construct an initial reward function model based on the reward function coupling equation group: and ,in is the state value function, is the cumulative reward function, is the time discount factor, is the attenuation coefficient, is the noise term in the value function; calculate the expected value of cumulative rewards and the characteristic expectation function; the purpose of this step is to establish the initial reward evaluation mechanism. Calculation of the expected value of cumulative rewards: Where, is the expected value of the cumulative reward; is the time step; is the discount factor; is the joint probability distribution of state-action pairs; is the immediate reward function. The characteristic expectation function calculation formula is: Where, is the characteristic expectation function; is the characteristic function; For the initial state and action.

[0166] The specific implementation of step S06 is to extract features through principal component analysis, perform standardization preprocessing on the market data, calculate the covariance matrix, solve the eigenvalues ​​and eigenvectors; select the eigenvectors with a cumulative contribution rate of 95% to form the principal component space; project the original data into the principal component space to obtain the feature representation after dimensionality reduction; calculate the actual feature expected value; the purpose of this step is to extract key features and reduce computational complexity. Data standardization: Where, is the standardized data; is the original data; For the column mean; For the Column standard deviation.

[0167] Covariance matrix calculation: Where, is the covariance matrix; is the sample size; is the standardized data matrix.

[0168] Eigenvalue decomposition: Where, is the eigenvector matrix; is a diagonal matrix of eigenvalues.

[0169] Dimensionality reduction transformation: Where, is the data matrix after dimensionality reduction; is the original data matrix; For the selected A matrix of eigenvectors.

[0170] The specific implementation of step S07 is to train the inverse reinforcement learning model, use the feature expectation deviation as the optimization target, and use the policy gradient method to update the model parameters; during the iterative optimization process, when the feature expectation deviation is less than 0.01, the iteration is stopped; and the optimized reward function model is obtained; the purpose of this step is to obtain an accurate reward function. Among them, the policy gradient calculation formula is: Where, is the policy gradient; is a parameterized strategy function; is the discounted return. Parameter update rule: Where, for Strategy parameters at the moment; is the learning rate.

[0171] The specific implementation of step S08 is to construct a group of market subject behavior equations, including the utility maximization partial differential equation: ; Strategy optimization coupling equation: ; Market clearing dynamics equation: ; Constraint coupling equations: ; The purpose of this step is to establish a complete mathematical model of market entity behavior.

[0172] The specific implementation method of step S09 is to construct a deep neural network model, adopting an encoder-decoder structure, including a feature extraction module, a temporal coding layer, an attention layer and a policy decoding layer; the feature extraction module adopts a 4-layer convolutional neural network with a convolution kernel size of 3×3 and a step size of 1; the temporal coding layer adopts a bidirectional long short-term memory network with a hidden layer dimension of 128; the attention layer adopts a multi-head attention mechanism with 8 heads; the policy decoding layer adopts a fully connected network with an activation function of ReLU; the purpose of this step is to establish a behavior prediction model.

[0173] The specific implementation method of step S10 is to use the maximum likelihood estimation method to solve the unknown parameters, construct the log-likelihood function, and use the stochastic gradient ascent method to optimize the parameter values; the confidence interval of the parameter estimate is set to 95%; and a market entity behavior probability distribution model is established; the purpose of this step is to obtain accurate parameter estimates of the behavior model.

[0174] The specific implementation method of step S11 is to generate a market entity behavior trend forecast report, predict future behavior trends based on the probability distribution model, and the forecast period is 7 days; calculate the 95% confidence interval of the forecast results; analyze the sensitivity of influencing factors; generate visual charts; write a forecast report; and make decision recommendations. The purpose of this step is to provide actionable decision support information.

[0175] Explanation of the coefficient values ​​in each step: Time series decomposition weight coefficient Obtained by least squares fitting; diffusion coefficient The value range is 0.01 to 0.1; optimization coefficient The value range is 0.1 to 1; the market adjustment coefficient The value range is 0.01 to 0.5; network sensitivity coefficient The value range is 0.1 to 1; time discount factor The value range is 0.9 to 0.99; attenuation coefficient The value range is 0.01 to 0.1; all noise terms are assumed to be zero-mean bounded noise.

[0176] To better understand and implement the present invention, Example 2 of a specific application scenario of the present invention is provided below. The specific implementation process of this method will be illustrated using a regional electricity market as an example. This regional electricity market has 15 thermal power generation companies participating in day-ahead market quotation transactions. The market utilizes a marginal clearing price mechanism with a one-day trading cycle and time-slot quotation clearing. This example analyzes historical data from January 1, 2024, to March 31, 2024.

[0177] During the data collection phase, market-clearing price data was first collected from the power trading platform database, with a sampling period of 15 minutes. This data included market-clearing prices at 96 time points over a period of 91 days, ranging from 180 to 520 yuan per megawatt-hour. Bidding strategy data was collected for each power generation company, including segmented bidding curves and bid electricity. This bidding strategy data showed that companies generally adopted a three- to five-segment bidding strategy, with bids ranging from 160 to 550 yuan per megawatt-hour. Actual transaction volume data was collected, and the total transaction volume for each company over the 91-day period ranged from 850 million kWh to 1.26 billion kWh. System load demand data showed that daily maximum loads ranged from 28 million kilowatts to 32 million kilowatts, and daily minimum loads ranged from 16 million kilowatts to 19 million kilowatts. Generator unit technical constraint data includes parameters such as each power generation company's installed capacity, unit ramp rate, and minimum technical output. The maximum installed capacity is 3.2 million kilowatts, and the minimum is 800,000 kilowatts. Network security constraint data includes the limits on major transmission corridors, with a maximum transmission capacity of 5 million kilowatts. Market share data shows that the largest market player holds 15.8% of installed capacity, while the smallest player holds 3.2%.

[0178] The collected data was decomposed into a time series. For example, 91 days of clearing price data for a typical power generation company were analyzed. The price stability component showed an overall slow upward trend, rising from an average of 260 yuan per megawatt-hour in early January to 285 yuan per megawatt-hour by the end of March. The price volatility component revealed price fluctuations between weekdays and weekends, with fluctuations ranging between plus or minus 35 yuan per megawatt-hour. The decomposition of the bidding strategy data revealed a clear cyclical nature for the company's bidding strategy. The bid stability component showed an upward trend with fuel price fluctuations, while the bid volatility component was highly correlated with system load changes.

[0179] A five-tuple model without a reward function was constructed. The state vector consists of six dimensions: the ratio of the price stability component, the angle of the price fluctuation component, the ratio of the bid stability component, the angle of the bid fluctuation component, the hyperbolic sine of demand, and the hyperbolic tangent of cost. The price benchmark was 270 yuan per megawatt-hour, the bid benchmark was 265 yuan per megawatt-hour, the demand benchmark was 25 million kilowatts, and the cost benchmark was 200 yuan per megawatt-hour. The state transition matrix was derived through historical data regression, with a nonlinear adjustment coefficient set to 0.15. The action space includes five typical bidding strategy combinations: aggressive bidding, robust bidding, conservative bidding, follow-up bidding, and mixed bidding.

[0180] Using the Monte Carlo method, 1,500 initial state samples were collected, covering market conditions at different time periods, load levels, and competitive dynamics. Statistical analysis of these samples showed that during weekday peak hours, the mean of the price fluctuation component was high, with a standard deviation of approximately 28.5 yuan per megawatt-hour. During weekend trough hours, the mean of the bid price fluctuation component was low, with a standard deviation of approximately 12.3 yuan per megawatt-hour.

[0181] An initial reward function model was constructed based on expert knowledge, taking into account power generation revenue, operating costs, and opportunity costs. A time discount factor of 0.95 and a decay coefficient of 0.05 were used. The calculated expected value of cumulative rewards showed that an aggressive bidding strategy has higher expected returns during peak hours, while a conservative bidding strategy has a greater advantage during off-peak hours. The characteristic expected function reflects the long-term return performance of different strategy combinations.

[0182] Principal component analysis was used to reduce the dimensionality of market data and extract five principal eigenvectors, with a cumulative contribution rate of 96.8%. The first principal component reflects the positive correlation between system load and price, with a contribution rate of 45.3%; the second principal component reflects the relationship between competition intensity and bidding strategy, with a contribution rate of 23.5%; the third principal component reflects the impact of cost factors, with a contribution rate of 15.2%; the fourth principal component reflects time characteristics, with a contribution rate of 8.4%; and the fifth principal component reflects the impact of network constraints, with a contribution rate of 4.4%.

[0183] A reinforcement learning approach was used to train a market participant behavior model. After 2,000 rounds of iterative training, the feature expectation deviation gradually decreased from an initial 0.156 to 0.008, below the preset threshold of 0.01. The optimized reward function model better reflects the decision-making preferences of market participants, achieving a prediction accuracy of 92.5% on weekdays and 88.7% on weekends.

[0184] Table 1 below gives the prediction results of the optimal bidding strategy for a power generation company under different market scenarios:

[0185] Table 1 Prediction results

[0186]

[0187] Based on the forecast results, the power generation company adopted the model's recommended bidding strategy, achieving positive results in the following week's market transactions. For example, on April 1st, the company traded 8.55 million kWh of electricity in the day-ahead market at an average price of 292 yuan per megawatt-hour, a 5.8% increase from the previous week and a 0.6 percentage point increase in market share.

[0188] The deep neural network model was trained using three months of historical data, with 70% used for training, 20% for validation, and 10% for testing. The model achieved a root mean square error of 18.5 yuan per megawatt-hour and a mean absolute percentage error of 6.2% on the test set. The feature extraction module utilizes four convolutional layers with a 3×3 kernel size. The first layer has 6 input channels, corresponding to the 6 dimensions of the state vector. The number of channels in subsequent layers increases to 32, 64, 128, and 256, respectively. The temporal encoding layer utilizes a bidirectional long short-term memory network with a hidden layer dimension of 128 and a sequence length of 96, corresponding to 96 time points within a day. The attention layer utilizes an 8-head attention mechanism, with each head having a dimension of 32. The strategy decoding layer utilizes a three-layer fully connected network with 512, 256, and 5 neurons in each layer, corresponding to the five available quotation strategies.

[0189] Maximum likelihood estimation results for the model parameters show that the diffusion coefficients α1 and α2 are 0.035 and 0.042, respectively; the strategy optimization coefficients β1, β2, and β3 are 0.28, 0.45, and 0.63, respectively; the market regulation coefficients κ1, κ2, and κ3 are 0.15, 0.22, and 0.31, respectively; and the network sensitivity coefficients σ1 and σ2 are 0.25 and 0.38, respectively. The 95% confidence intervals for these parameter estimates are all within reasonable ranges, indicating that the model has good stability.

[0190] The Market Player Behavior Trend Forecast Report provides a detailed analysis of market conditions over the next seven days. The forecast indicates that between April 2nd and 8th, the system's maximum load will reach 31.5 million kilowatts, with a minimum load around 17.5 million kilowatts. The 95% confidence interval for the market-clearing price is between 235 and 315 yuan per megawatt-hour. The report recommends that power generation companies adopt an aggressive bidding strategy during peak weekday hours and flexibly adjust their bidding strategies during other periods based on real-time market competition. The report also identifies potential risk factors, including load fluctuations due to weather changes and cost uncertainty caused by fuel price fluctuations.

[0191] It should be noted that the variables involved in the description of the present invention are explained in detail in Table 2 below.

[0192] Table 2 Variable Explanation Table

[0193]

[0194] The above description is only a specific embodiment of the present invention, but the protection scope 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 protection scope of the present invention.

Claims

1. A method for outputting a report on determining the behavior trend of power market entities, characterized in that: The following steps are involved: S01. Collect historical data on power market operation, including market clearing price data, subject bidding strategy data, actual transaction volume data, system load demand data, unit technical constraint data, network security constraint data, and market share data; S02. Perform multi-dimensional time series decomposition on the market clearing price data and the subject quotation strategy data to obtain price stability component data, price fluctuation component data, quotation stability component data, and quotation fluctuation component data; S03. Constructing a five-tuple model without a reward function using the price stability component data, the price fluctuation component data, the quotation stability component data, and the quotation fluctuation component data, and establishing a state space matrix and an action space matrix; S04. Performing Monte Carlo sampling based on the state space matrix and the action space matrix to obtain an initial state sample set; S05. For the initial state sample set, construct an initial reward function model using expert knowledge, and calculate the cumulative reward expectation value and the characteristic expectation function; S06. Performing dimensionality reduction processing on the market clearing price data, the subject quotation strategy data, and the actual transaction volume data using a principal component analysis method, extracting key feature vectors, and calculating actual feature expected values; S07. Compare the actual feature expectation value with the feature expectation function to obtain a feature expectation deviation, train an inverse reinforcement learning model based on the feature expectation deviation, iteratively update the initial reward function model until the feature expectation deviation is less than a preset threshold, and obtain an optimized reward function model; S08. Constructing a group of market subject behavior equations using the optimized reward function model; S09. Construct a deep neural network model, wherein the deep neural network model adopts an encoder-decoder structure, including a feature extraction module, a temporal encoding layer, an attention layer, and a strategy decoding layer; S10. Based on the market entity behavior equations and the deep neural network model, a maximum likelihood estimation method is used to solve unknown parameters and establish a market entity behavior probability distribution model; S11. Predicting the behavior trends of market entities based on the market entity behavior probability distribution model, and generating a market entity behavior trend prediction report; Among them, multi-dimensional time series decomposition specifically adopts nonlinear time series decomposition: ; ; Where, is the time series of market clearing prices; To stabilize the price; is the price fluctuation component; is the price decomposition error term; is the quote strategy time series; To stabilize the quotation weight; is the quotation fluctuation component; Decompose error terms for quotes; is the time series decomposition weight coefficient; The market subject behavior equations include: Utility maximization partial differential equation: ; Where, is the utility function; is the diffusion coefficient; is the nonlinear coupling term; is the utility function noise; Strategy optimization coupling equation: ; Where, For quotation strategy; The optimal strategy for the goal; is the optimization coefficient; Optimizing noise for strategies; Market clearing dynamics equation: ; Where, is the market price; is total demand; For the The output of each subject; is the market adjustment coefficient; is price noise; Constraint coupling equations: ; Where, is the transmission constraint function; is the network node coordinate; is the network sensitivity coefficient; The feature extraction module uses a convolutional neural network to extract the spatiotemporal features of market data; the temporal encoding layer uses a long short-term memory network to capture the temporal dependencies of data.

2. The method for outputting a report on determining the behavior trend of power market entities according to claim 1, characterized in that: The market subject behavior equation group includes a utility maximization equation, a strategy optimization equation, a market clearing equation and a constraint condition equation.

3. The method for outputting a report on determining the behavior trend of power market entities according to claim 2, characterized in that: The utility maximization equation is used to calculate the profit function of the market entity. The input includes the entity's bidding strategy data, the market clearing price data, and power generation cost data. The output is the expected profit value.

4. The method for outputting a report on determining the behavior trend of power market entities according to claim 2, characterized in that: The strategy optimization equation is used to update the optimal bidding strategy of the market entity. The input includes the entity bidding strategy data, the competing entity bidding data, and the system load demand data. The output is the optimal bidding strategy data.

5. The method for outputting a report on determining the behavior trend of power market entities according to claim 2, characterized in that: The market clearing equation is used to calculate the market equilibrium price. The input includes the subject quotation strategy data, the system load demand data, and the network security constraint data. The output is the market clearing price forecast value.

6. The method for outputting a report on determining the behavior trend of power market entities according to claim 2, characterized in that: The constraint condition equation is used to ensure the feasibility of the strategy, the input includes the unit technical constraint data and the network security constraint data, and the output is the constraint judgment result data.

7. A computer-readable storage medium, characterized in that The computer-readable storage medium stores program instructions, and when the program instructions are executed in a computer, they are used to execute the method for outputting a report on determining the behavior trend of power market entities according to any one of claims 1 to 6.

8. A system for outputting a report on the behavior trend of power market entities, characterized in that: Contains the computer-readable storage medium of claim 7.

Images