Power market risk monitoring method and system based on multiple agents

Abstract

The invention provides a multi-agent-based power market risk monitoring method and system, relates to the field of power market risk control, and integrates information perception, inference analysis and execution feedback functions by constructing a three-layer progressive multi-agent collaborative network composed of an information monitoring layer, a collaborative analysis layer and a supervision execution layer. A hybrid communication protocol based on semantic similarity is adopted, structural data transmission and a natural language collaborative discussion mechanism are combined, deep collaboration between agents is comprehensively supported from quantitative analysis to qualitative reasoning, and in the process of generating a power market decision scheme, the decision scheme of the power market is generated based on an antagonism debate mechanism. The efficiency optimization agent and the risk control agent are debate for multiple rounds, the weight is dynamically adjusted through the reinforcement learning algorithm to generate a balance decision scheme, the limitation of a traditional single-target optimization method is broken through, multiple targets of supervision efficiency and risk control can be completely considered, and self-adaptive decision support is provided for intelligent supervision and risk prevention and control of the power market.

Description

Technical Field

[0001] This application relates to the field of power market risk control, and in particular to a power market risk monitoring method and system based on multi-agent systems. Background Technology

[0002] Electricity market regulatory decisions are a crucial link in maintaining fair competition in the electricity market and ensuring the safe and stable operation of the power system. They are of great significance for promoting the optimal allocation of electricity resources and preventing systemic risks. Against the backdrop of deepening electricity market reform, electricity market regulation faces complex challenges such as diversified market participants, complex trading models, and heterogeneous data sources, exhibiting a high degree of dynamism and uncertainty.

[0003] Traditional manual decision-making and single-model analysis methods suffer from problems such as insufficient information processing capabilities, lack of decision-making collaboration mechanisms, difficulty in balancing multiple objectives, and weak adaptive learning capabilities. These problems can lead to delayed regulatory responses and inaccurate risk identification, which not only affect the healthy development of the electricity market but also easily trigger systemic risks.

[0004] Therefore, how to regulate the risks in the electricity market is a technical problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] The purpose of this application is to provide a method and system for monitoring power market risks based on multi-agent systems, which can effectively improve the level of intelligence in power market supervision and provide important technical support for the standardized operation of the power market and the safe and stable operation of the power system.

[0006] To address the aforementioned technical problems, this application provides a multi-agent-based method for monitoring electricity market risks, the specific technical solution of which is as follows:

[0007] Acquire multi-source heterogeneous regulatory data and preprocess it to generate a standardized regulatory dataset containing market status characteristics; the multi-source heterogeneous regulatory data includes at least one of electricity trading data, price fluctuation information and market participant behavior data; the market status characteristics include real-time trading prices, trading volumes, market participant identity information, historical violation records, electricity supply and demand balance indicators, market risk assessment indicators and market relationship characteristics;

[0008] A three-layer progressive multi-agent collaborative network consisting of an information monitoring layer, a collaborative analysis layer, and a regulatory execution layer is constructed. Based on a pre-trained language model, the agents at each layer are fine-tuned with professional domain knowledge. A hybrid communication protocol supporting structured data transmission and semantic-level collaborative discussion is established. A four-layer progressive knowledge management architecture is constructed through a graph neural network to obtain a multi-agent collaborative system.

[0009] The multi-agent collaborative system is invoked to perform multi-layer collaborative analysis on real-time data. An adversarial debate mechanism is introduced at the collaborative analysis layer to enable the efficiency optimization agent and the risk control agent to engage in multiple rounds of debate. Based on the reinforcement learning algorithm, the debate weights are dynamically adjusted to generate a balanced power market decision scheme.

[0010] Optionally, acquiring and preprocessing multi-source heterogeneous regulatory data to generate a standardized regulatory dataset containing market state characteristics includes:

[0011] A unified data acquisition interface is established, and raw data is obtained from multiple data sources by calling the unified data acquisition interface. A multi-source data fusion algorithm based on timestamp synchronization is adopted to convert the raw data into an enhanced JSON format and then store it in a standardized manner.

[0012] Linear interpolation is used to repair missing values ​​in time series data in the original data. Forward and backward filling algorithms based on sliding windows are used to process missing segments in continuous data. Kalman filtering is used to smooth noisy real-time data.

[0013] Numerical transaction data is formatted and context-enhanced, converting raw values ​​into structured descriptions containing contextual information such as units, timestamps, and year-on-year change rates. Textual legal provisions are processed using semantic vectorization based on BERT. Time-series data is segmented according to business cycles and trend indicators are calculated to form multimodal data that is easy for large language models to understand and analyze.

[0014] A market relationship feature extraction algorithm based on graph neural networks is used to construct a market participant relationship graph and a transaction behavior relationship graph. The network topology features, centrality indicators, and community structure features of the market participant relationship graph and the transaction behavior relationship graph are extracted as multimodal data. The multimodal data is classified and integrated according to time series and business type to generate a standardized input dataset. The standardized input dataset includes the data content, processing status, timestamp, hierarchical identifier, and relationship features of the multimodal data.

[0015] Optionally, using linear interpolation to repair missing values ​​in the time series data in the original data includes:

[0016] Missing values ​​in the time series data in the original data are repaired using a linear interpolation formula; the linear interpolation formula is:

[0017] ;

[0018] in, As the reference time, As the base date, The length of the interpolation interval; Represented as day Continuous time series values ​​sampled at each time step. Represented as day Continuous time series values ​​sampled at each time step. Represented as day Continuous time series values ​​sampled at each time step; For adaptive weighting factors; ,in For adaptive parameters, This represents the standard deviation of the data.

[0019] Optionally, the smoothing process of the noisy real-time data using the Kalman filtering method includes:

[0020] Obtain the state equation and observation equation corresponding to the Kalman filter method;

[0021] An adaptive noise covariance matrix is ​​generated based on the state equation and the observation equation;

[0022] Based on the adaptive noise covariance matrix, iterative prediction and update steps are performed on the noisy real-time data to obtain the optimal estimate;

[0023] Wherein, the state equation is The observation equation is as follows: ;

[0024] The adaptive noise covariance matrix is:

[0025] , ;

[0026] in, Let A be the system state vector at time k; A is the state transition matrix, used to represent the transition law of the system state from time k-1 to time k. B is the system state vector at time k-1; B is the control input matrix. This is the control input vector at time k-1; This is the process noise vector at time k-1; Let be the observation vector at time k; The observation matrix; This is the noise matrix; Let I be the process noise covariance matrix at time k; I is the identity matrix. This is the predicted value of the system state at time k based on the state at time k-1. Let be the observation noise covariance matrix at time k; State prediction error covariance matrix traces; For process noise, To observe the noise, and All parameters are adaptively adjusted. It is an identity matrix.

[0027] Optionally, classifying and integrating the multimodal data according to time series and business type includes:

[0028] Statistical feature enhancement is performed on the numerical data in the multimodal data:

[0029] ;

[0030] in, The mean, Standard deviation For skewness, For kurtosis, Trend coefficient;

[0031] Semantic vectorization is performed on the textual data within the multimodal data:

[0032] ;

[0033] in, Indicates feature concatenation operation;

[0034] Data from different modalities are fused based on a multimodal feature fusion formula; the multimodal feature fusion formula is as follows:

[0035] ; This refers to time-series data within the multimodal data.

[0036] in, Both are weight matrices. This is the bias vector.

[0037] Optionally, the process of constructing the market participant relationship graph includes:

[0038] Construct a market participant relationship graph based on the participant node set, transaction relationship edge set, and adjacency matrix;

[0039] Market participant relationship diagram construction: ,in, For the set of participant nodes, For the set of transaction relationship edges, It is an adjacency matrix;

[0040] Feature extraction using graph convolutional neural networks: ;

[0041] in, It is an adjacency matrix. For degree matrix, For the first Layer weight matrix;

[0042] Centrality index calculation:

[0043] Betweenness centrality: ;

[0044] PageRank centrality: ;

[0045] in, The damping coefficient is... The total number of nodes. Pointing to a node A set of nodes.

[0046] Optionally, a three-layer progressive multi-agent collaborative network consisting of an information monitoring layer, a collaborative analysis layer, and a supervisory execution layer is constructed. Based on a pre-trained language model, each layer of agents is fine-tuned using domain-specific knowledge. A hybrid communication protocol supporting structured data transmission and semantic-level collaborative discussion is established. A four-layer progressive knowledge management architecture is constructed using a graph neural network, resulting in a multi-agent collaborative system including:

[0047] Construct a three-layer progressive multi-agent collaborative network architecture that includes an information monitoring layer, a collaborative analysis layer, and a regulatory execution layer;

[0048] The system constructs intelligent agents corresponding to each of the information monitoring layer, the collaborative analysis layer, and the regulatory execution layer, and configures their functions. The information monitoring layer deploys a full-market information fusion intelligent agent, a compliance review intelligent agent, and a market memory and knowledge management intelligent agent. The collaborative analysis layer deploys a market regulatory analysis module and a regulatory strategy research module. The regulatory execution layer deploys a multi-level regulatory execution intelligent agent, a full-market risk monitoring intelligent agent, and a dynamic regulatory optimization intelligent agent.

[0049] A structured data transmission format is used to process quantitative regulatory information, a natural language collaborative discussion mechanism is designed to realize the communication of reasoning processes between intelligent agents, and a unified message routing system is built to support the dynamic switching between synchronous collaboration and asynchronous communication.

[0050] A four-layer progressive knowledge management architecture is constructed based on graph neural networks. A pre-trained language model based on attention mechanism is used to fine-tune the professional domain knowledge of each agent. A hybrid training strategy of supervised learning, reinforcement learning and contrastive learning is combined to realize the automatic abstraction of knowledge and the static construction of the agent collaboration system.

[0051] Optionally, a three-layer progressive multi-agent collaborative network architecture comprising an information monitoring layer, a collaborative analysis layer, and a supervisory execution layer may be constructed, including:

[0052] The information monitoring layer, as the lowest layer of the system, is used to perform real-time collection, preliminary processing, and anomaly identification of multi-source heterogeneous regulatory data;

[0053] The collaborative analysis layer, as the core layer of the system, is used to perform regulatory analysis and generate regulatory strategy recommendations. It includes a market regulatory analysis module, and the regulatory strategy research module is used to generate regulatory strategies through an adversarial debate mechanism.

[0054] As the top layer of the system, the regulatory execution layer is used to formulate and implement regulatory decisions, possesses regulatory response capabilities across multiple time scales, and establishes a comprehensive regulatory execution system covering everything from real-time emergency response to long-term institutional development.

[0055] The three-layer progressive multi-agent cooperative network architecture includes a bottom-up information enhancement transmission mechanism and a top-down instruction feedback mechanism.

[0056] Optionally, when constructing a three-layer progressive multi-agent collaborative network architecture that includes an information monitoring layer, a collaborative analysis layer, and a supervisory execution layer, the following may also be included:

[0057] A progressive information processing flow is added to the three-layer progressive multi-agent collaborative network architecture.

[0058] Optionally, the step of using a pre-trained language model based on an attention mechanism to fine-tune the domain knowledge of each layer of the agent includes:

[0059] The multi-head attention mechanism is used to fine-tune the domain knowledge of each layer of the intelligent agent; wherein, the attention calculation includes domain weights.

[0060] Optionally, the four-layer progressive knowledge management architecture based on graph neural networks includes:

[0061] A knowledge graph is constructed based on entity sets, relation sets, and triple sets; the knowledge graph includes a four-layer knowledge management strategy: real-time layer, short-term layer, long-term layer, and permanent layer;

[0062] Knowledge embedding learning is performed based on the aforementioned knowledge graph;

[0063] The importance of knowledge in the knowledge graph is assessed.

[0064] Optionally, the multi-agent collaborative system is invoked to perform multi-layer collaborative analysis on real-time data. An adversarial debate mechanism is introduced at the collaborative analysis layer to enable the efficiency optimization agent and the risk control agent to engage in multiple rounds of debate. Based on reinforcement learning algorithms, the debate weights are dynamically adjusted to generate a balanced electricity market decision-making scheme, including:

[0065] By setting dynamic trigger thresholds based on statistical learning and anomaly detection rules based on machine learning, standardized task instructions are sent to the collaborative analysis layer when a regulatory event is detected.

[0066] The design incorporates a multi-round debate process and a dynamic weight adjustment algorithm. Through a reinforcement learning mechanism, the decision weights are updated in real time during the debate, and a power market decision scheme is output.

[0067] The electricity market decision-making scheme is converted into control instructions, and the control instructions are decomposed and resources are allocated according to the time dimension.

[0068] Collect feedback data after the implementation of the power market decision-making scheme, and optimize the debate weighting strategy and agent model parameters based on the feedback data.

[0069] Optionally, the dynamic trigger threshold based on statistical learning adopts the following mathematical model:

[0070] The dynamic trigger threshold is calculated based on the historical mean, historical standard deviation, confidence coefficient, and time decay constant.

[0071] Calculate the isolated forest anomaly score of the original data and compare it with the dynamic trigger threshold; if it exceeds the dynamic trigger threshold, an anomaly event is confirmed.

[0072] Calculate the local density of the anomalous event in the neighboring data, and determine whether the anomalous event is a local outlier.

[0073] If it is a local outlier, the abnormal event is confirmed as an individual abnormal event; otherwise, it is a persistent abnormal event.

[0074] The risk index of the abnormal event is calculated based on the risk type weight, severity, and corresponding threshold of the abnormal event; the risk index is used to characterize the impact of the abnormal event.

[0075] Optionally, the design of the multi-round debate process and dynamic weight adjustment algorithm, which updates the decision weights in real time during the debate through a reinforcement learning mechanism and outputs a power market decision scheme, includes:

[0076] Initialize the adversarial debate agent and set the corresponding debate parameters;

[0077] The structured multi-round debate process of the adversarial debate agent is executed; in which the first round is for problem analysis and position statement, the second round is for evidence presentation and data support, the third round is for policy recommendations and solution comparison, the fourth round is for impact assessment and risk analysis, and the fifth round is for position adjustment and consensus seeking. Each round of debate generates a complete argumentation process through natural language.

[0078] Based on the reward function, combined with decision-making effectiveness, market stability, and execution costs, a weight adjustment amount is calculated after each round of debate based on the strength of the argument and historical performance data, and a weight adjustment range limit is set for the weight adjustment amount.

[0079] The application performs mathematical weighted fusion on the analysis results of the intelligent agent based on the dynamically adjusted weights, generating a power market decision-making scheme that includes specific regulatory measures, quantitative indicators of expected effects, key points of risk control, and implementation timetable.

[0080] This application also provides a power market regulatory decision-making system based on multi-agent collaboration, including:

[0081] The data acquisition module is used to acquire multi-source heterogeneous regulatory data and preprocess it to generate a standardized regulatory dataset containing market status characteristics. The multi-source heterogeneous regulatory data includes at least one of electricity trading data, price fluctuation information, and market participant behavior data. The market status characteristics include real-time trading prices, trading volumes, market participant identity information, historical violation records, electricity supply and demand balance indicators, market risk assessment indicators, and market relationship characteristics.

[0082] The collaborative system construction module is used to build a three-layer progressive multi-agent collaborative network consisting of an information monitoring layer, a collaborative analysis layer, and a regulatory execution layer. Based on a pre-trained language model, the agents at each layer are fine-tuned with professional domain knowledge, a hybrid communication protocol supporting structured data transmission and semantic-level collaborative discussion is established, and a four-layer progressive knowledge management architecture is constructed through a graph neural network to obtain the multi-agent collaborative system.

[0083] The decision generation module is used to call the multi-agent collaborative system to perform multi-layer collaborative analysis on real-time data. An adversarial debate mechanism is introduced in the collaborative analysis layer to enable the efficiency optimization agent and the risk control agent to conduct multiple rounds of debate. Based on the reinforcement learning algorithm, the debate weights are dynamically adjusted to generate a balanced power market decision scheme.

[0084] This application provides a multi-agent-based method for monitoring power market risks, comprising: acquiring and preprocessing multi-source heterogeneous regulatory data to generate a standardized regulatory dataset containing market state characteristics; the multi-source heterogeneous regulatory data includes at least one of power trading data, price fluctuation information, and market participant behavior data; the market state characteristics include real-time trading prices, trading volumes, market participant identity information, historical violation records, power supply and demand balance indicators, market risk assessment indicators, and market relationship characteristics; constructing a three-layer progressive multi-agent collaborative network consisting of an information monitoring layer, a collaborative analysis layer, and a regulatory execution layer; fine-tuning the agents at each layer based on pre-trained language models using professional domain knowledge; establishing a hybrid communication protocol supporting structured data transmission and semantic-level collaborative discussion; constructing a four-layer progressive knowledge management architecture through graph neural networks to obtain a multi-agent collaborative system; calling the multi-agent collaborative system to perform multi-layer collaborative analysis on real-time data; introducing an adversarial debate mechanism at the collaborative analysis layer to enable the efficiency optimization agent and the risk control agent to engage in multiple rounds of debate; and dynamically adjusting the debate weights based on reinforcement learning algorithms to generate a balanced power market decision-making scheme.

[0085] This application constructs a three-layer progressive multi-agent collaborative network consisting of an information monitoring layer, a collaborative analysis layer, and a regulatory execution layer. It integrates information perception, reasoning analysis, and execution feedback functions, enabling a more systematic handling of complex regulatory tasks compared to traditional single-decision models. Employing a semantic similarity-based hybrid communication protocol, combined with structured data transmission and natural language collaborative discussion mechanisms, it comprehensively supports deep collaboration among agents from quantitative analysis to qualitative reasoning. Compared to conventional systems that rely solely on preset rules, it achieves a more flexible collaborative model. In the process of generating power market decision-making schemes, an adversarial debate mechanism allows efficiency optimization agents and risk control agents to engage in multiple rounds of debate. Reinforcement learning algorithms dynamically adjust weights to generate balanced decision schemes, overcoming the limitations of traditional single-objective optimization methods and comprehensively considering multiple objectives of regulatory efficiency and risk control. The optimized dynamic weight adjustment strategy outputs scientific and personalized regulatory measures, providing adaptive decision support for intelligent power market regulation and risk prevention.

[0086] This application also provides a power market regulatory decision-making system based on multi-agent cooperation, which has the above-mentioned beneficial effects, and will not be elaborated here. Attached Figure Description

[0087] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of this application. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0088] Figure 1 A flowchart illustrating a multi-agent-based power market risk monitoring method provided in this application embodiment;

[0089] Figure 2 A schematic diagram of a three-layer progressive multi-agent cooperative network structure provided in an embodiment of this application;

[0090] Figure 3 This is a schematic diagram of hybrid communication in an agent cooperative network provided in an embodiment of this application;

[0091] Figure 4 This is a schematic diagram of a power market supervision system based on multi-agent cooperation, provided as an embodiment of this application. Detailed Implementation

[0092] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0093] See Figure 1 , Figure 1 A flowchart of a multi-agent-based power market risk monitoring method provided in this application embodiment, the method including:

[0094] S101: Acquire multi-source heterogeneous regulatory data and preprocess it to generate a standardized regulatory dataset containing market status characteristics;

[0095] The multi-source heterogeneous regulatory data includes at least one of electricity trading data, price fluctuation information, and market participant behavior data; the market status characteristics include real-time trading prices, trading volume, market participant identity information, historical violation records, electricity supply and demand balance indicators, market risk assessment indicators, and market relationship characteristics.

[0096] S102: Construct a three-layer progressive multi-agent collaborative network consisting of an information monitoring layer, a collaborative analysis layer, and a regulatory execution layer. Based on a pre-trained language model, fine-tune the professional domain knowledge of each agent. Establish a hybrid communication protocol that supports structured data transmission and semantic-level collaborative discussion. Construct a four-layer progressive knowledge management architecture through a graph neural network to obtain a multi-agent collaborative system.

[0097] S103: Call the multi-agent cooperative system to perform multi-layer cooperative analysis on real-time data, introduce an adversarial debate mechanism in the cooperative analysis layer, so that the efficiency optimization agent and the risk control agent can conduct multiple rounds of debate, and generate a balanced power market decision scheme by dynamically adjusting the debate weights based on the reinforcement learning algorithm.

[0098] Specifically, when performing step S101 to preprocess the multi-source heterogeneous regulatory data, the following steps may be included:

[0099] A1: Establish a unified data acquisition interface, call the unified data acquisition interface to obtain raw data from multiple data sources, adopt a multi-source data fusion algorithm based on timestamp synchronization, convert the raw data into enhanced JSON format and then store it in a standardized manner;

[0100] A2: Linear interpolation is used to repair missing values ​​in the time series data in the original data. Forward and backward filling algorithms based on sliding windows are used to process missing segments of continuous data. Kalman filtering is used to smooth the noisy real-time data.

[0101] Missing values ​​in the time series data in the original data are repaired using a linear interpolation formula; the linear interpolation formula is:

[0102] ;

[0103] in, As the reference time, As the base date, The length of the interpolation interval; Represented as day Continuous time series values ​​sampled at each time step. Represented as day Continuous time series values ​​sampled at each time step. Represented as day Continuous time series values ​​sampled at each time step; For adaptive weighting factors; ,in For adaptive parameters, This represents the standard deviation of the data.

[0104] Obtain the state equation and observation equation corresponding to the Kalman filter method;

[0105] An adaptive noise covariance matrix is ​​generated based on the state equation and the observation equation;

[0106] Based on the adaptive noise covariance matrix, iterative prediction and update steps are performed on the noisy real-time data to obtain the optimal estimate;

[0107] Wherein, the state equation is The observation equation is as follows: ;

[0108] The adaptive noise covariance matrix is:

[0109] , ;

[0110] in, For process noise, To observe the noise, and All parameters are adaptively adjusted. It is an identity matrix.

[0111] A3: Standardize the format and enhance the context of numerical transaction data, convert the raw values ​​into a structured description containing contextual information such as units, timestamps, and year-on-year change rates, use BERT-based semantic vectorization to process textual legal provisions, and segment time-series data according to business cycles and calculate trend indicators to form multimodal data that is easy for large language models to understand and analyze.

[0112] A4: A market relationship feature extraction algorithm based on graph neural networks is used to construct a market participant relationship graph and a transaction behavior relationship graph. The network topology features, centrality indicators, and community structure features of the market participant relationship graph and the transaction behavior relationship graph are extracted as multimodal data. The multimodal data is classified and integrated according to time series and business type to generate a standardized input dataset. The standardized input dataset includes the data content, processing status, timestamp, hierarchical identifier, and relationship features of the multimodal data.

[0113] The multimodal data is categorized and integrated according to time series and business type, including:

[0114] Statistical feature enhancement is performed on the numerical data in the multimodal data:

[0115] ;

[0116] in, The mean, Standard deviation For skewness, For kurtosis, Trend coefficient;

[0117] Semantic vectorization is performed on the textual data within the multimodal data:

[0118] ;

[0119] in, Indicates feature concatenation operation;

[0120] Data from different modalities are fused based on a multimodal feature fusion formula; the multimodal feature fusion formula is as follows:

[0121] ; This refers to time-series data within the multimodal data.

[0122] in, Both are weight matrices. This is the bias vector.

[0123] After the above three processes, multimodal data can be classified and integrated according to time series and business type.

[0124] The process of constructing a market participant relationship diagram includes:

[0125] Construct a market participant relationship graph based on the participant node set, transaction relationship edge set, and adjacency matrix;

[0126] Market participant relationship diagram construction: ,in, For the set of participant nodes, For the set of transaction relationship edges, It is an adjacency matrix;

[0127] Feature extraction using graph convolutional neural networks: ;

[0128] in, It is an adjacency matrix. For degree matrix, For the first Layer weight matrix;

[0129] Centrality index calculation:

[0130] Betweenness centrality: ;

[0131] PageRank centrality: ;

[0132] in, The damping coefficient is... The total number of nodes. Pointing to a node A set of nodes.

[0133] In step S102, the construction process of the three-layer progressive multi-agent cooperative network can be as follows:

[0134] B1: Construct a three-layer progressive multi-agent collaborative network architecture that includes an information monitoring layer, a collaborative analysis layer, and a regulatory execution layer;

[0135] B2: Construct the intelligent agents corresponding to each layer of the information monitoring layer, the collaborative analysis layer, and the regulatory execution layer, and configure their functions; the information monitoring layer deploys a full-market information fusion intelligent agent, a compliance review intelligent agent, and a market memory and knowledge management intelligent agent; the collaborative analysis layer deploys a market regulatory analysis module and a regulatory strategy research module; and the regulatory execution layer deploys a multi-level regulatory execution intelligent agent, a full-market risk monitoring intelligent agent, and a dynamic regulatory optimization intelligent agent.

[0136] B3: Employ a structured data transmission format to process quantitative regulatory information, design a natural language collaborative discussion mechanism to enable communication of reasoning processes among intelligent agents, and construct a unified message routing system to support dynamic switching between synchronous collaboration and asynchronous communication;

[0137] B4: A four-layer progressive knowledge management architecture is constructed based on graph neural networks. A pre-trained language model based on attention mechanism is used to fine-tune the professional domain knowledge of each agent. A hybrid training strategy of supervised learning, reinforcement learning and contrastive learning is combined to realize the automatic abstraction of knowledge and the static construction of the agent collaboration system.

[0138] For step B1, see Figure 2 , Figure 2 This is a schematic diagram of a three-layer progressive multi-agent cooperative network structure provided in an embodiment of this application. The construction process of the three-layer progressive multi-agent cooperative network architecture can be carried out in the following manner:

[0139] Information monitoring layer architecture design: As the bottom layer of the system, it is responsible for the real-time collection, preliminary processing and anomaly identification of multi-source heterogeneous regulatory data, establishes 24 / 7 uninterrupted market monitoring capability, and provides standardized basic data support for upper-level analysis;

[0140] Collaborative Analysis Layer Architecture Design: As the core layer of the system, it is responsible for professional regulatory analysis and strategy research. It includes a market regulatory analysis module for multi-dimensional professional analysis and a regulatory strategy research module for forming balanced regulatory strategy recommendations through an adversarial debate mechanism.

[0141] Regulatory execution layer architecture design: As the top layer of the system, it is responsible for the formulation and execution of regulatory decisions, has the ability to respond to regulatory issues at multiple time scales, and establishes a comprehensive regulatory execution system covering everything from real-time emergency response to long-term institutional development;

[0142] Establishing inter-layer interaction relationships: Design a bottom-up information enhancement transmission mechanism and a top-down instruction feedback mechanism to ensure that regulatory information is gradually enriched and accurate during the inter-layer transmission process, forming a closed-loop collaborative decision-making process.

[0143] Figure 2 The structure shown is an example application structure of this application. Figure 2 Both the bottom and middle layers contain several intelligent agents. In the bottom layer, the full-market information fusion intelligent agent adopts a hybrid architecture. The data integration module is implemented based on Apache Kafka stream processing technology. The abnormal data identification intelligent agent is built based on a fine-tuned GPT-4o model, making full use of its multimodal capabilities to process the fusion analysis of chart data and text information. The compliance review intelligent agent is built based on a fine-tuned GPT-4o model, specifically optimizing the ability to understand regulations and determine violations. The market memory and knowledge management intelligent agent is built using traditional data management technology, establishing a four-layer progressive knowledge architecture based on vector databases and knowledge graphs.

[0144] In the collaborative analysis layer, the price regulation, competition regulation, and supply-demand balance agents in the market supervision and analysis module are built based on a fine-tuned GPT-4o model. The technology standard regulation agent is also built based on a fine-tuned GPT-4o model to handle complex technical charts and risk data visualization analysis. The risk prevention and control regulation agent is built based on a fine-tuned GPT-4o model to specifically identify and assess systemic market risks, including credit risk, liquidity risk, operational risk, and policy risk. In the regulatory strategy research module, the market efficiency optimization researcher is built based on a fine-tuned GPT-o1 model to enhance reasoning capabilities, and the risk prevention and control regulator is built based on a fine-tuned GPT-o1 model, forming regulatory consensus through a structured policy debate mechanism.

[0145] In the regulatory execution layer intelligent agent, the multi-level regulatory execution intelligent agent adopts a hybrid architecture. The decision reasoning module is built based on the fine-tuning of the GPT-o1 model to handle complex multi-factor decision reasoning, and the execution scheduling module is implemented based on the fine-tuning of the GPT-4 model combined with rule engine technology. The full market risk monitoring intelligent agent is mainly built based on time series prediction model and machine learning anomaly detection model. The dynamic regulatory optimization intelligent agent adopts a hybrid architecture with reinforcement learning algorithm as the main method and GPT-4o model as the auxiliary method.

[0146] Furthermore, based on the above embodiments, a progressive information processing flow can be added to the three-layer progressive multi-agent cooperative network architecture.

[0147] First, information gain is used for initial screening. Before information is transmitted from the lower layer (such as the information monitoring layer) to the upper layer (such as the collaborative analysis layer), the information gain is calculated to screen out high-value information (such as abnormal price data) that can reduce the uncertainty of information in the upper layer, and to eliminate redundant information, thus completing the initial filtering of effective information.

[0148] Secondly, objective function optimization is adopted. For the information after initial screening, the objective function is optimized by inter-layer information transfer. While maximizing information gain, the transfer cost is controlled, an efficient transfer path is determined, and effective information is upgraded to efficient information.

[0149] Finally, the quality indicators are reviewed. Before the information reaches the upper level, it is inspected using information quality assessment indicators (completeness, accuracy, timeliness). If it meets the standards, it enters the upper-level decision-making process (such as the regulatory implementation level formulating measures); if it does not meet the standards, it is returned to the lower level for reprocessing, and finally outputs high-quality information to support decision-making, completing the information progression closed loop.

[0150] The progressive information processing flow adopts the following mathematical model:

[0151] Interlayer information gain calculation: ;

[0152] in, For information entropy, For information collection, As an attribute, This represents the probability of the i-th type of information appearing in the information set S;

[0153] Inter-layer information transfer optimization objective function:

[0154] ;

[0155] in, For the number of floors, For the first Layer information, For information transmission costs, For regularization parameters;

[0156] Information quality assessment indicators:

[0157] ;

[0158] in, For the weight parameters, satisfying The information quality assessment index aims to evaluate the quality of information I transmitted to each layer from three dimensions: Completeness (I), Accuracy (I), and Timeliness (I). Completeness (I): The completeness of information I refers to whether the information set contains all the information relevant to the regulatory task and whether any key content is missing. For example, in the multi-source heterogeneous regulatory data of the electricity market, does it cover all necessary regulatory information such as electricity trading data, price fluctuation information, and market participant behavior data? Accuracy (I): The accuracy of information I refers to the degree to which the information conforms to the actual situation and whether there are any errors or deviations. For example, the consistency between monitored electricity trading volume, price, and other data and the actual market trading situation. Timeliness (I): The timeliness of information I refers to whether the speed of information acquisition and transmission can meet the time requirements of regulatory decision-making and whether there is any information lag. For example, in the supervision of abnormal events in the electricity market, can abnormal information be acquired and transmitted to the relevant processing levels in a timely manner?

[0159] In one feasible implementation, a multi-head attention mechanism can be used to fine-tune the domain-specific knowledge of each layer of the agent, and the attention calculation includes domain-specific weights. The specific process can be as follows:

[0160] Improvements to multi-head attention mechanisms:

[0161] ;

[0162] ;

[0163] Where Q represents the query matrix, which represents the task or information query requirement that the current agent needs to process; K represents the key matrix, which is used to match with the query matrix to measure the relevance of different information to the query requirement; V represents the value matrix, which contains the specific information content corresponding to the key matrix. When the query matrix matches the key matrix, the corresponding information is extracted from the value matrix according to the matching result as the output of attention calculation; Attention represents attention, Multihead represents multi-head attention value, and W represents different weight matrices.

[0164] In particular, attention calculation incorporates domain-specific weights:

[0165] ;

[0166] in, For professional field bias matrix; Describes the dimension of the key matrix K, for Divide the result by This is for scaling purposes, to avoid... An excessively large relevance score leads to an overly extreme output from the softmax function, affecting the reasonable allocation of attention weights. Softmax, or the softmax function, is used to convert relevance scores into attention weights, ensuring the sum of the weights equals 1.

[0167] Agent specialization loss function: ;

[0168] in, For mission losses, For domain adaptation loss, This results in a loss of consistency.

[0169] Model selection strategy: GPT-4o for multimodal analysis ( GPT-o1 is used for complex reasoning. GPT-4 is used for fast execution ( GPT-4o is used for equilibrium analysis. ).

[0170] For step B3, the specific implementation method can be as follows:

[0171] Structured data transmission format design: mainly used for inter-layer information transmission, adopting a standardized JSON format to carry regulatory data. The data includes unified header information including data packet identifier, timestamp, data source, target recipient, data type, priority and other fields. The business subject information is designed with corresponding data structures according to different regulatory scenarios. It supports three transmission types: market monitoring data packets (high-frequency transmission), abnormal event report packets (real-time transmission), and regulatory decision result packets (access-controlled transmission), ensuring the efficiency and standardization of cross-layer data flow.

[0172] The natural language collaborative discussion mechanism is designed primarily for intra-layer agent interaction, fully leveraging the advantages of large-scale models in natural language understanding and generation. At the information monitoring layer, natural language is used to achieve semantic analysis and knowledge association of abnormal events. At the collaborative analysis layer, deep professional collaboration is achieved through four dialogue methods: regulatory logic discussion, evidence presentation and questioning, policy suggestion negotiation, and impact assessment exchange. At the regulatory execution layer, natural language is used to explain the decision-making reasoning process and optimize the execution plan, maintaining the transparency of the agent's reasoning process and the interpretability of the decision logic, and supporting multi-turn dialogue and context memory functions.

[0173] Hybrid Communication Protocol and Unified Message Routing System Construction: A hybrid communication mechanism combining inter-layer structured communication and intra-layer natural language collaboration is established. Inter-layer communication utilizes efficient structured data flow to ensure information transmission speed and accuracy, while intra-layer communication leverages natural language collaboration to fully utilize the reasoning and collaboration advantages of large models. A unified message routing system is constructed to support dynamic switching between synchronous collaboration and asynchronous communication. The optimal communication mode and data format are automatically selected based on communication type, urgency, complexity, and the number of agents involved. Message priority queues and load balancing mechanisms are designed to achieve seamless integration and complementary advantages between structured data transmission and natural language collaboration.

[0174] For step B4, a knowledge graph is constructed based on the entity set, relation set, and triple set. The knowledge graph includes a four-layer knowledge management strategy: real-time layer, short-term layer, long-term layer, and permanent layer. The specific process is as follows:

[0175] Knowledge embedding learning is performed based on the aforementioned knowledge graph;

[0176] The importance of knowledge in the knowledge graph is assessed.

[0177] Knowledge graph construction: ,in For a collection of entities, For a set of relations, It is a set of triples;

[0178] Knowledge Embedded Learning:

[0179] ;

[0180] in, For entities In relation The set of neighboring nodes; Representing entities In the The feature vector after layer knowledge embedding is used to represent the entity after the embedding operation of this layer. The feature representation can better reflect the relationship information of entities in the knowledge graph; This indicates the number of knowledge embedding layers and the number of iterations in the embedding process. Represents the correspondence of the l-th level The weight matrix; t represents the set of relations in the knowledge graph, containing all possible association types between entities; t represents an entity. In relation The neighboring entities, i.e., entities Through relationships Other connected entities; This represents the feature vector of the neighboring entity t at the l-th layer; Represents the correspondence of the l-th level The bias vector; This represents the activation function.

[0181] A four-tiered knowledge management strategy:

[0182] Real-time layer: Redis storage, storage duration 24 hours, access complexity ;

[0183] Short-term layer: Redis persistence, storage duration is 30 days, retrieval complexity ;

[0184] Long-term layer: PostgreSQL, storage duration of 3 years, query complexity ;

[0185] Permanent layer: Neo4j graph database, inference complexity ;

[0186] Knowledge Importance Assessment:

[0187] ;

[0188] in, For knowledge Frequency of use The time decay factor, To influence the decision-making process , , These are the corresponding coefficients, which can be set by those skilled in the art.

[0189] Based on step B4, we can also establish, for example Figure 3 The hybrid communication protocol shown supports structured data transmission and semantic-level collaborative discussion. Figure 3 This is a schematic diagram of hybrid communication in an agent cooperative network provided in an embodiment of this application.

[0190] The above step S103 can specifically include the following steps:

[0191] C1: Configure a real-time data processing system and deploy a regulatory trigger condition identification algorithm. By setting dynamic trigger thresholds based on statistical learning and anomaly detection rules based on machine learning, it automatically sends standardized task instructions to the collaborative analysis layer when a regulatory event is detected.

[0192] C2: Construct an adversarial debate framework between an efficiency optimization agent and a risk control agent, design a multi-round debate process and a dynamic weight adjustment algorithm, and use a reinforcement learning mechanism to calculate and update the decision weights of the two dimensions in real time during the debate process;

[0193] C3: Deploy regulatory measures generation algorithms and multi-timescale execution scheduling systems to automatically convert the power market decision-making schemes output by collaborative analysis into specific regulatory instructions, and decompose tasks and allocate resources according to different time dimensions;

[0194] C4: Deploy an implementation effect monitoring system and an automatic model parameter update mechanism to collect feedback data on the implementation of regulatory measures and automatically optimize the debate weight strategy and agent model parameters.

[0195] In C1, the dynamic trigger threshold setting based on statistical learning can be achieved using the following steps:

[0196] The first step is to calculate the dynamic trigger threshold based on the historical mean, historical standard deviation, confidence coefficient, and time decay constant.

[0197] The second step is to calculate the isolated forest anomaly score of the original data and compare it with the dynamic trigger threshold; if it exceeds the dynamic trigger threshold, an anomaly event is confirmed.

[0198] The third step is to calculate the local density of the anomalous event in the neighboring data and determine whether the anomalous event is a local outlier.

[0199] Step 4: If it is a local outlier, confirm that the abnormal event is an isolated abnormal event; otherwise, it is a persistent abnormal event.

[0200] Step 5: Calculate the risk index of the abnormal event based on the risk type weight, severity, and corresponding threshold of the abnormal event; the risk index is used to characterize the impact of the abnormal event.

[0201] Dynamic threshold calculation:

[0202] ;

[0203] in, for Historical average at any given time For historical standard deviation, The confidence coefficient is... The time decay constant; This represents the time interval between the current time t and the historical reference time, that is, the length of time elapsed from the historical reference time to the current time.

[0204] Machine learning-based anomaly detection algorithms:

[0205] Isolated forest outlier score: ,in, For the sample Average path length across all isolated trees Standardization factor;

[0206] LOF local anomaly factors:

[0207] ;

[0208] in, For point Locally achievable density, Let B be the local reachability density. for of Nearest neighbor set;

[0209] Comprehensive risk assessment indicators:

[0210] ;

[0211] in, The weight of risk type i, As to the degree of severity, This corresponds to the threshold.

[0212] The specific implementation process of step C2 can be carried out in the following steps:

[0213] The first step is to initialize the adversarial debate agent and set the corresponding debate parameters;

[0214] The second step is to execute the structured multi-round debate process of the adversarial debate agent. The first round involves problem analysis and position statement, the second round involves evidence presentation and data support, the third round involves policy recommendations and solution comparison, the fourth round involves impact assessment and risk analysis, and the fifth round involves position adjustment and consensus seeking. Each round of debate generates a complete argumentation process through natural language.

[0215] The third step is to calculate the weight adjustment amount after each round of debate based on the strength of the argument and historical effect data, according to the reward function combined with the decision effect, market stability and execution cost, and set the weight adjustment range limit of the weight adjustment amount.

[0216] The fourth step involves applying the dynamically adjusted weights based on the aforementioned weight adjustment amount to perform mathematical weighted fusion of the analysis results of the intelligent agent, generating a power market decision-making scheme that includes specific regulatory measures, quantitative indicators of expected effects, key points of risk control, and an implementation timetable.

[0217] In practical applications, adversarial debate agents can be initialized and debate parameters can be set: the efficiency optimization agent is configured based on the GPT-o1 model with a preset efficiency priority weight of 0.6, the risk control agent is based on the GPT-4 model with a preset stability priority weight of 0.4, the maximum number of debate rounds is set to 5, and the time limit for each round is 3 minutes.

[0218] The algorithm executes a structured multi-round debate process: the first round involves problem analysis and position statement; the second round involves evidence presentation and data support; the third round involves policy recommendations and solution comparison; the fourth round involves impact assessment and risk analysis; and the fifth round involves position adjustment and consensus seeking. Each round of debate generates a complete argumentation process using natural language.

[0219] Deploy a dynamic weight adjustment system based on the PPO (Proximal Policy Optimization) algorithm: Design a reward function that combines decision effect (weight 0.4), market stability (weight 0.3) and execution cost (weight 0.3). After each round of debate, dynamically calculate the weight adjustment amount based on the strength of the argument and historical effect data. The weight adjustment range is limited to ±0.2 to ensure system stability.

[0220] Generate a weighted and integrated balanced electricity market decision-making scheme: Apply dynamically adjusted weights to mathematically weight and integrate the analysis results of the two agents to generate a complete decision-making scheme that includes specific regulatory measures, quantitative indicators of expected effects, key points of risk control, and implementation timetable. The feasibility of the scheme is ensured through a logical consistency check algorithm.

[0221] The specific implementation process of step C3 can be carried out in the following steps:

[0222] Step 1: Deploy regulatory decision analysis and automatic task decomposition algorithms: Configure a decision analysis module based on the GPT-o1 model to automatically decompose abstract regulatory solutions into an executable task list according to four time dimensions: immediate response (0-30 minutes), short-term adjustment (1-24 hours), medium-term optimization (1-30 days), and long-term construction (1-12 months);

[0223] The second step is to establish a multi-timescale task scheduling and resource allocation system: design a task priority scoring algorithm (urgency × impact scope × execution difficulty), configure a resource pool management system including human resources, technical systems and monitoring tools, and achieve optimal resource allocation and task scheduling through an integer programming algorithm;

[0224] The third step is to deploy an intelligent execution scheduling engine and progress tracking system: Based on the GPT-4o model and a rule engine, establish a mechanism for real-time monitoring of task execution status, automatic identification of execution deviations, and dynamic adjustment of scheduling schemes, and maintain real-time statistics of key indicators such as task execution rate, resource utilization rate, and target achievement rate.

[0225] Step 4: Configure the execution effect monitoring and emergency adjustment response system: Deploy market parameter monitoring sensors, set execution effect evaluation index thresholds, and automatically trigger the re-analysis process and generate adjustment plans when execution deviations exceed the preset range or new risk situations occur.

[0226] The specific implementation process of step C4 can be carried out in the following steps:

[0227] Step 1: Deploy a multi-dimensional execution performance data collection and evaluation system: Configure a KPI monitoring dashboard to collect data such as regulatory target achievement rate, market stability indicators, participant satisfaction and execution costs, establish an execution performance scoring algorithm and set the data collection frequency to once per hour, with evaluation cycles at three levels: daily, weekly and monthly.

[0228] The second step is to configure the reinforcement learning model training and weight strategy automatic optimization system: the performance score is used as a reward signal to input into the PPO algorithm, the learning rate is set to 0.001, the batch size is 64, the experience replay buffer size is 10000, the model is updated once every 100 samples are collected, and the debate weight adjustment strategy is continuously optimized.

[0229] The third step is to deploy an automatic tuning and knowledge update mechanism for the intelligent agent model parameters: Based on the execution feedback data, the model parameters of each professional analysis intelligent agent are automatically adjusted weekly using meta-learning algorithms. The reasoning patterns of successful cases are integrated into the intelligent agent knowledge base through knowledge distillation technology, and the performance threshold and safety boundary of model updates are set.

[0230] The fourth step is to establish an automated assessment and optimization process for continuous improvement of regulatory decision-making capabilities: Configure an A / B testing framework to compare the differences in the effectiveness of new and old decision-making strategies, set up a quarterly overall system performance assessment, and generate system output reports on decision quality improvement and system optimization suggestions through automated reporting to achieve a spiral evolution of regulatory capabilities.

[0231] See Figure 4 , Figure 4 The diagram below illustrates a multi-agent collaborative power market regulatory system architecture. This application also provides a multi-agent collaborative power market regulatory decision-making system, comprising:

[0232] The data acquisition module is used to acquire multi-source heterogeneous regulatory data and preprocess it to generate a standardized regulatory dataset containing market status characteristics. The multi-source heterogeneous regulatory data includes at least one of electricity trading data, price fluctuation information, and market participant behavior data. The market status characteristics include real-time trading prices, trading volumes, market participant identity information, historical violation records, electricity supply and demand balance indicators, market risk assessment indicators, and market relationship characteristics.

[0233] The collaborative system construction module is used to build a three-layer progressive multi-agent collaborative network consisting of an information monitoring layer, a collaborative analysis layer, and a regulatory execution layer. Based on a pre-trained language model, the agents at each layer are fine-tuned with professional domain knowledge, a hybrid communication protocol supporting structured data transmission and semantic-level collaborative discussion is established, and a four-layer progressive knowledge management architecture is constructed through a graph neural network to obtain the multi-agent collaborative system.

[0234] The decision generation module is used to call the multi-agent collaborative system to perform multi-layer collaborative analysis on real-time data. An adversarial debate mechanism is introduced in the collaborative analysis layer to enable the efficiency optimization agent and the risk control agent to conduct multiple rounds of debate. Based on the reinforcement learning algorithm, the debate weights are dynamically adjusted to generate a balanced power market decision scheme.

[0235] This application also provides an embodiment corresponding to a computer-readable storage medium. The computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of the method described in the above method embodiments.

[0236] It is understood that if the methods in the above embodiments are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and executes all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0237] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. As the system provided in the embodiments corresponds to the method provided in the embodiments, the description is relatively simple; relevant parts can be found in the method section.

[0238] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. It should be noted that those skilled in the art can make several improvements and modifications to this application without departing from the principles of this application, and these improvements and modifications also fall within the protection scope of this application.

[0239] It should also be noted that, in this specification, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

Claims

1. A multi-agent based electricity market risk monitoring method, characterized in that, The method comprises the following steps: acquiring multi-source heterogeneous regulatory data and preprocessing to generate a standardized regulatory data set containing market state features; The multi-source heterogeneous regulatory data includes at least one of power transaction data, price fluctuation information and market participant behavior data; the market state features include real-time transaction price, transaction volume, market participant identity information, historical violation record, power supply and demand balance index, market risk assessment index and market relationship feature; A three-layer progressive multi-agent collaboration network composed of information monitoring layer, collaborative analysis layer and regulatory execution layer is constructed, each layer agent is fine-tuned based on pre-trained language model in professional field knowledge, a hybrid communication protocol supporting structured data transmission and semantic level collaborative discussion is established, a four-layer progressive knowledge management architecture is constructed through graph neural network, and a multi-agent collaboration system is obtained; The multi-agent collaboration system is called to perform multi-layer collaborative analysis on real-time data, an adversarial debate mechanism is introduced in the collaborative analysis layer to enable efficiency optimization agent and risk control agent to perform multiple rounds of debate, and a balanced power market decision scheme is generated based on reinforcement learning algorithm dynamic adjustment of debate weight. 2.The multi-agent based electricity market risk monitoring method of claim 1, wherein, The method comprises the following steps: A unified data acquisition interface is established, and the unified data acquisition interface is called to acquire original data from multiple data sources, a multi-source data fusion algorithm based on timestamp synchronization is used to convert the original data into enhanced JSON format and then store them in a standardized manner; The missing values of time series type data in the original data are repaired by using linear interpolation method, the missing sections of continuous data are processed by using forward filling algorithm and backward filling algorithm based on sliding window, and the real-time data containing noise are smoothed by using Kalman filtering method; The numerical transaction data are subjected to format standardization and context enhancement processing, the original numerical values are converted into structured description containing unit, timestamp, same period change rate and other context information, the text type regulation provisions are subjected to semantic vectorization processing based on BERT, the time series type data are segmented according to business cycle and trend index is calculated, and multi-modal data convenient for large language model understanding and analysis are formed; A market participant relationship graph and a transaction behavior relationship graph are constructed by using a market relationship feature extraction algorithm based on graph neural network, the network topology features, centrality indexes and community structure features of the market participant relationship graph and the transaction behavior relationship graph are extracted as multi-modal data, the multi-modal data are classified and integrated according to time sequence and business type, and a standardized input data set is generated; the standardized input data set contains data content, processing state, timestamp, hierarchical identifier and relationship feature of the multi-modal data.

3. The multi-agent based electricity market risk monitoring method of claim 2, wherein, The missing values of time series type data in the original data are repaired by using linear interpolation method, which comprises the following steps: The missing values of time series type data in the original data are repaired according to the linear interpolation formula; the linear interpolation formula is as follows: ； wherein, is a reference time, is a reference day, is an interpolation interval length; is represented as day is a continuous time series value sampled at a time, is represented as day is a continuous time series value sampled at a time, is represented as day is a continuous time series value sampled at a time; is an adaptive weight factor; wherein is an adaptive parameter, is a data standard deviation.

4. The multi-agent based electricity market risk monitoring method of claim 2, wherein, The real-time data containing noise are smoothed by using Kalman filtering method, which comprises the following steps: The state equation and observation equation corresponding to the Kalman filtering method are obtained; generating an adaptive noise covariance matrix according to the state equation and the observation equation; iterative prediction and update steps are performed on real-time data containing noise according to the adaptive noise covariance matrix to obtain an optimal estimation value; wherein the state equation is: ； the observation equation is: ； the adaptive noise covariance matrix is: ， ； wherein, is the system state vector at the kth time instant; A is the state transition matrix, which is used to represent the transition rule of the system state from the (k-1)th time instant to the kth time instant; is the system state vector at the (k-1)th time instant; B is the control input matrix; is the control input vector at the (k-1)th time instant; is the process noise vector at the (k-1)th time instant; is the observation vector at the kth time instant; is the observation matrix; is the noise matrix; is the process noise covariance matrix at the kth time instant; I is the identity matrix; is the predicted value of the system state at the kth time instant based on the state at the (k-1)th time instant; is the observation noise covariance matrix at the kth time instant; is the state prediction error covariance matrix is the trace of the state prediction error covariance matrix is the process noise, is the observation noise, and are adaptive adjustment parameters.

5. The multi-agent based electricity market risk monitoring method of claim 2, wherein, classifying and integrating the multi-modal data according to time series and business types includes: statistical feature enhancement is performed on numerical data in the multi-modal data: ； wherein, is the mean, is the standard deviation, is the skewness, is the kurtosis, is the trend coefficient; semantic vectorization is performed on text data in the multi-modal data: ； wherein represents a feature stitching operation; different modal data are fused based on a multi-modal feature fusion formula; the multi-modal feature fusion formula is: ; is a time-series data in the multi-modal data; wherein, are weight matrices, is a bias vector.

6. The multi-agent based electricity market risk monitoring method of claim 2, wherein, the construction process of the market participant relationship graph includes: a market participant relationship graph is constructed according to a participant node set, a transaction relationship edge set and an adjacency matrix; Market participant relationship graph construction: wherein, is a set of participant nodes, is a set of transaction relationship edges, is an adjacency matrix; Graph convolutional neural network feature extraction: ; wherein, is an adjacency matrix, is a degree matrix, is an l-th layer weight matrix, is an l-th layer weight matrix, is an output feature matrix of an l+1-th layer graph convolutional neural network, is an activation function; a centrality index is calculated: Betweenness centrality: ; PageRank centrality: ; wherein, is a damping coefficient, is the total number of nodes, is a set of nodes pointing to the node .

7. The multi-agent based electricity market risk monitoring method of claim 1, wherein, a three-layer progressive multi-agent collaboration network composed of an information monitoring layer, a collaborative analysis layer and a supervision execution layer is constructed, pre-trained language models are used to fine-tune the professional field knowledge of agents at each layer, a hybrid communication protocol supporting structured data transmission and semantic-level collaborative discussion is established, a four-layer progressive knowledge management architecture is constructed through a graph neural network, and a multi-agent collaboration system is obtained, including: a three-layer progressive multi-agent collaboration network architecture including an information monitoring layer, a collaborative analysis layer and a supervision execution layer is constructed; agents corresponding to each layer of the information monitoring layer, the collaborative analysis layer and the supervision execution layer are constructed and functionally configured; the information monitoring layer deploys a full-market information fusion agent, a compliance review agent and a market memory and knowledge management agent, the collaborative analysis layer deploys a market supervision analysis module and a supervision strategy research module, and the supervision execution layer deploys a multi-level supervision execution agent, a full-market risk monitoring agent and a dynamic supervision optimization agent; structured data transmission formats are used to process quantitative supervision information, natural language collaborative discussion mechanisms are designed to realize the exchange of reasoning processes between agents, and a unified message routing system is constructed to support the dynamic switching of synchronous collaboration and asynchronous communication; a four-layer progressive knowledge management architecture is constructed based on a graph neural network, pre-trained language models based on attention mechanisms are used to fine-tune the professional field knowledge of agents at each layer, and a hybrid training strategy combining supervised learning, reinforcement learning and contrastive learning is used to realize automatic abstraction of knowledge and static construction of the multi-agent collaboration system.

8. The multi-agent based electricity market risk monitoring method of claim 2, wherein, a three-layer progressive multi-agent collaboration network architecture including an information monitoring layer, a collaborative analysis layer and a supervision execution layer is constructed, including: the information monitoring layer as the bottom layer of the system is used to perform real-time collection, preliminary processing and abnormality identification of multi-source heterogeneous supervision data; the collaborative analysis layer as the core layer of the system is used to perform supervision analysis and generate supervision strategy suggestions, including a market supervision analysis module and a supervision strategy research module used to generate supervision strategies through an adversarial debate mechanism; the supervision execution layer as the top layer of the system is used to formulate and execute supervision decisions, has multi-time scale supervision response capabilities, and establishes a full-coverage supervision execution system from real-time emergency response to long-term system construction; The three-layer progressive multi-agent collaborative network architecture includes a bottom-up information enhancement transmission mechanism and a top-down instruction feedback mechanism. 9.The multi-agent based power market risk monitoring method of claim 7, wherein, When constructing the three-layer progressive multi-agent collaborative network architecture including the information monitoring layer, the collaborative analysis layer, and the regulatory execution layer, the following is also included: Adding a progressive information processing flow to the three-layer progressive multi-agent collaborative network architecture; The progressive information processing flow adopts the following mathematical model: Inter-layer information gain calculation: ; wherein, is the information entropy, is the information set, is the attribute; Optimization objective function for inter-layer information transmission: ； wherein, is the number of layers, is the first layer information, is the information transfer cost, is the regularization parameter; Information quality evaluation index: ； wherein is a weight parameter satisfying .

10. The multi-agent based electricity market risk monitoring method of claim 7, wherein, The professional field knowledge fine-tuning of each layer agent using the pre-trained language model based on the attention mechanism includes: Fine-tuning the professional field knowledge of each layer agent based on the multi-head attention mechanism; wherein the attention calculation includes professional field weights. 11.The multi-agent based electricity market risk monitoring method of claim 7, wherein, The four-layer progressive knowledge management architecture based on the graph neural network includes: Building a knowledge graph based on the entity set, the relationship set, and the triple set; the knowledge graph includes four-layer knowledge management strategies of the real-time layer, the short-term layer, the long-term layer, and the permanent layer; Knowledge embedding learning based on the knowledge graph; Knowledge importance evaluation of the knowledge in the knowledge graph.

12. The multi-agent based electricity market risk monitoring method of claim 1, wherein, Calling the multi-agent collaborative system for multi-layer collaborative analysis of real-time data, introducing an adversarial debate mechanism in the collaborative analysis layer to enable the efficiency optimization agent and the risk control agent to conduct multiple rounds of debate, and dynamically adjusting the debate weights based on the reinforcement learning algorithm to generate a balanced power market decision scheme includes: By setting a dynamic trigger threshold based on statistical learning and an abnormality detection rule based on machine learning, sending a standardized task instruction to the collaborative analysis layer when a regulatory event is detected; Designing a multi-round debate process and weight dynamic adjustment algorithm, updating decision weights in real time during the debate process through a reinforcement learning mechanism, and outputting a power market decision scheme; Converting the power market decision scheme into control instructions, and decomposing the control instructions into tasks and allocating resources according to the time dimension; Collecting feedback data after the execution of the power market decision scheme, and adjusting the debate weight strategy and the agent model parameters according to the feedback data.

13. The multi-agent based electricity market risk monitoring method of claim 12, wherein, The dynamic trigger threshold based on statistical learning adopts the following mathematical model: Calculating the dynamic trigger threshold based on the historical mean, the historical standard deviation, the confidence coefficient, and the time decay constant; Calculating the isolation forest anomaly score of the original data and comparing it with the dynamic trigger threshold; If it exceeds the dynamic trigger threshold, it is confirmed that there is an abnormal event; Calculating the local density of the abnormal event in the adjacent data to determine whether the abnormal event is a local outlier; If it is a local outlier, it is confirmed that the abnormal event is an individual abnormal event; Otherwise, it is a persistent abnormal event; Calculating the risk index of the abnormal event according to the risk type weight, the severity, and the corresponding threshold of the abnormal event; the risk index is used to represent the impact of the abnormal event.

14. The multi-agent based electricity market risk monitoring method of claim 12, wherein, The design of the multi-round debate process and the weight dynamic adjustment algorithm, which updates the decision weights in real time during the debate process through a reinforcement learning mechanism, and outputs a power market decision scheme includes: Initializing the adversarial debate agent and the corresponding debate parameters; The structured multi-round argumentation process of the adversarial argumentation agent is performed; wherein the first round carries out question analysis and position exposition, the second round carries out evidence display and data support, the third round carries out policy suggestion and scheme comparison, the fourth round carries out impact evaluation and risk analysis, the fifth round carries out position adjustment and consensus seeking, and each round of argumentation generates a complete argumentation process through natural language generation; According to the reward function, the decision effect, the market stability and the execution cost, the weight adjustment amount is calculated according to the argumentation strength and the historical effect data after each round of argumentation, and the weight adjustment range limit of the weight adjustment amount is set; The analysis results of the agent are mathematically weighted and fused by applying the weight adjusted based on the weight adjustment amount, to generate a power market decision scheme including specific regulatory measures, expected effect quantitative indicators, risk control points and implementation schedule.

15. A multi-agent cooperation based power market supervision decision system, characterized in that, It includes: A data acquisition module for acquiring multi-source heterogeneous regulatory data and preprocessing to generate a standardized regulatory data set containing market state characteristics; The multi-source heterogeneous regulatory data includes at least one of power transaction data, price fluctuation information and market participant behavior data; the market state characteristics include real-time transaction price, transaction volume, market participant identity information, historical violation record, power supply and demand balance index, market risk assessment index and market relationship characteristics; A collaboration system construction module for constructing a three-layer progressive multi-agent collaboration network composed of an information monitoring layer, a collaborative analysis layer and a regulatory execution layer, fine-tuning each layer agent based on a pre-trained language model in a professional field, establishing a hybrid communication protocol supporting structured data transmission and semantic-level collaborative discussion, and constructing a four-layer progressive knowledge management architecture through a graph neural network to obtain a multi-agent collaboration system; A decision generation module for calling the multi-agent collaboration system to perform multi-layer collaborative analysis on real-time data, introducing an adversarial argumentation mechanism in the collaborative analysis layer to enable the efficiency optimization agent and the risk control agent to carry out multi-round argumentation, and dynamically adjusting the argumentation weight based on a reinforcement learning algorithm to generate a balanced power market decision scheme.

Images