Power transaction data interaction method and system based on large model and agent cooperation

Abstract

This invention discloses a method and system for power trading data interaction based on large-scale model and intelligent agent collaboration, belonging to the field of power system automation technology. This invention significantly improves the efficiency of power trading data interaction through the collaboration of large-scale models and intelligent agents. The semantic representation and multimodal feature fusion technology of the large-scale model, combined with structured data graphs, provides a precise foundation for data interaction; dynamic task allocation and hierarchical consensus verification among multiple intelligent agents improve task processing efficiency and accuracy, and reduce error rates. Simultaneously, an iterative feedback optimization mechanism ensures stable system operation, while blockchain encryption and cross-domain verification enhance data security. Furthermore, cross-domain data fusion and standardized model mapping break down format barriers, outputting results that conform to rules and include risk assessments, providing unified and reliable data support for power trading, promoting business standardization, and ensuring advantages in efficiency, security, and compatibility.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of power system automation, in particular to a power transaction data interaction method and system based on large model and agent cooperation. BACKGROUND

[0002] The existing data interaction method has problems such as non-uniform data format, complex cross-domain verification, and low processing efficiency, which is difficult to meet the real-time and security requirements. The Chinese patent application with publication number CN120125972A discloses a power image-text interaction method, system and related equipment based on a multi-modal large model. The power image-text interaction method includes collecting power pictures and general field pictures, training a pre-established power visual encoder; constructing a multi-modal large language model, and modifying the general visual encoder of the multi-modal large language model through the trained power visual encoder to obtain a power image-text large model; constructing a power image-text multi-task annotation data set, and fine-tuning the obtained power image-text large model; using the power image-text large model after fine-tuning, building a service, and answering the input pictures and questions. The patent application introduces a professional field visual encoder into the multi-modal large model, sends the output features of the power visual encoder into a new visual adapter, aligns and fuses the features with the general visual adapter, and then sends them into a decoder, thereby improving the analysis ability of the multi-modal large model for professional field images.

[0003] However, the above-mentioned prior art, although realizing the combination of professional field knowledge and general model, still has the following problems:

[0004] 1. It does not involve a multi-agent cooperation mechanism, and only relies on a single model to complete the image-text interaction task, which is difficult to cope with the multi-agent and cross-domain data interaction scene in power transaction, cannot realize the dynamic decomposition and distributed scheduling of tasks, and is limited in efficiency when processing massive concurrent transaction data;

[0005] 2. It lacks security transmission and cross-domain verification for power transaction data, which is difficult to meet the high requirements of power transaction data on security and non-tamperability, and cannot directly output standardized results that meet the power transaction rules;

[0006] 3. When an abnormal state occurs in the interaction process, it cannot real-time correct and optimize the interaction path, the system stability and fault tolerance are weak, and it is difficult to adapt to the real-time requirements of power transaction. SUMMARY

[0007] The purpose of the present application is to provide a power transaction data interaction method and system based on large model and agent collaboration, construct a data multi-modal feature space, and realize data semantic representation and fusion with a large model; simplify cross-domain verification and improve efficiency with the help of a multi-agent collaboration framework; and ensure the real-time and security of data interaction through iterative feedback and blockchain technology, to solve the problems raised in the above background technology.

[0008] To achieve the above purpose, the present application provides the following technical solutions:

[0009] The power transaction data interaction method based on large model and agent collaboration comprises:

[0010] Power transaction data multi-modal feature space construction: based on a large model, the original power transaction data is semantically represented, a structured data graph is constructed, and a multi-modal feature space of power transaction data is formed;

[0011] Multi-agent collaboration framework deployment and task scheduling: based on the structured data graph, a multi-agent collaboration framework is built, and based on the dynamic task allocation mechanism between agents, the obtained data interaction tasks are decomposed and scheduled;

[0012] Data interaction process optimization: based on the iterative feedback optimization mechanism of the agent framework and the large model, the abnormal state in the data interaction process is corrected in real time, and the optimal interaction path is generated;

[0013] Power transaction data interaction: based on the optimal interaction path, the safe transmission and cross-domain verification of power transaction data are carried out, and the standardized interaction result conforming to the power transaction rules is output.

[0014] Further, the multi-modal feature space of power transaction data is constructed, comprising:

[0015] Based on each data interface, the original power transaction data is obtained, and the original power transaction data is preprocessed, and the multi-modal feature of the preprocessed original power transaction data is extracted to generate a multi-modal feature vector;

[0016] Wherein, the original power transaction data includes structured data, unstructured text data and time series data, the numerical features of the structured data are extracted, the semantic features of the unstructured text data are extracted through a large model to generate a semantic feature vector; the trend features, periodic features and random fluctuation features of the time series data are extracted to form a time series feature sequence;

[0017] Quantify the correlation strength of each modal feature, map the modal features with correlation strength exceeding the preset correlation threshold, and weight and fuse according to the distribution weight value of each modal feature to generate a comprehensive feature vector;

[0018] Based on the fused comprehensive feature vector, a structured data graph is generated through a graph embedding algorithm. The structured data graph uses the transaction subject, transaction object, and transaction time as core nodes. By combining the correlation relationships of various modal features, connection edges are established between the core nodes, and attribute information of each core node and edge is added.

[0019] Furthermore, the deployment and task scheduling of the multi-agent collaborative framework include:

[0020] Based on the functional requirements of the power trading data interaction scenario, a data acquisition agent, a data verification agent, a path planning agent, a security monitoring agent, and a task coordination agent are initialized. Each agent communicates through a distributed message bus.

[0021] Construct an agent performance evaluation matrix, evaluate the multi-dimensional evaluation indicators of each agent when processing data interaction tasks based on the agent performance evaluation matrix, determine the load status of each agent, and dynamically update the agent performance evaluation matrix based on the preset number of data interaction tasks.

[0022] An initial task allocation strategy is generated, with task completion time, agent load balancing, and data interaction security as reward factors. A task allocation matrix is ​​generated based on task priority and agent load status.

[0023] Furthermore, the deployment and task scheduling of the multi-agent collaborative framework also includes building a consensus protocol among the agents:

[0024] In the same-domain agent verification, each agent performs the first verification of the task execution result based on the preset verification rule base and obtains the first-level verification result;

[0025] In cross-domain agent verification, agents from different domains perform secondary verification on the first-layer verification results based on cross-domain adaptability parameters to check the correlation and consistency of cross-domain data and obtain the second verification results.

[0026] The verification results are compared with the corresponding preset verification pass rate. If any verification result is lower than the preset verification pass rate, an arbitration instruction is generated based on the historical interaction patterns in the structured data graph and the power trading rules to guide the agent to reconstruct the task execution plan.

[0027] Furthermore, the agent performance evaluation matrix also includes:

[0028] When intelligent agents perform data interaction tasks, they collect real-time operational data such as communication latency, task execution success rate, and consensus time.

[0029] The running state of the multi-agent collaborative framework is evaluated based on the agent performance evaluation matrix when a preset number of data interaction tasks are completed;

[0030] When the communication delay exceeds the preset continuous number of times and exceeds the preset delay threshold, the inter-agent communication link is adjusted by the path planning agent;

[0031] When the task execution success rate is lower than the preset success rate threshold, the target parameters of the corresponding agent are updated by the task coordination agent according to the performance evaluation matrix.

[0032] Further, the iterative feedback optimization mechanism comprises:

[0033] Based on the running state evaluation results of each agent, the communication delay data, data integrity indicators and abnormal event logs in the running data are extracted;

[0034] The data integrity indicators and abnormal event logs are input into a large model for anomaly detection, and the historical interaction patterns in the structured data graph are extracted and compared;

[0035] Based on the comparison result, the deviation degree of the current data integrity indicator and the historical data integrity indicator average value is obtained, and when the deviation degree exceeds the preset deviation threshold, a correction strategy is generated based on the abnormal type of the abnormal event log.

[0036] Further, the secure transmission and cross-domain verification of the power transaction data comprises:

[0037] The power transaction data is encrypted and stored as evidence, and a transaction voucher containing a timestamp and a digital signature is generated;

[0038] The data interface of different power transaction subjects is called, and multi-dimensional verification is performed based on the preset consensus protocol between agents;

[0039] The interaction result that passes the verification is mapped to a standard power transaction data model, and the standard model data is risk assessed by a large model, combined with the historical transaction risk records in the structured data graph, to calculate the price fluctuation risk value, the performance probability and the cross-domain transmission security level, and to generate a visual report containing transaction risk assessment.

[0040] Further, the task allocation matrix is generated according to the task priority and the agent load state, taking the task completion time, the agent load balancing degree and the data interaction security as reward factors, including:

[0041] Based on the power transaction data interaction function, the agent is divided into data collection agent, data verification agent, path planning agent, security monitoring agent and task coordination agent, and the scene instances are subdivided, and the performance index, current load rate and safety capability index of each type of agent are taken as the row dimension data of the task allocation matrix;

[0042] Based on the task type, the task is divided into transaction subject data collection task, transaction target parameter verification task, cross-domain data transmission path planning task, transaction data security monitoring task and abnormal data correction task, and a double attribute label is attached to each type of task, and the task type, task priority label and safety level label are taken as the column dimension data of the task allocation matrix; wherein the task priority label is set to three levels of high / medium / low according to the power transaction rules and quantified as a normalized weight value;

[0043] Based on the row dimension data and column dimension data of the task allocation matrix, the structural elements of the task allocation matrix are constructed;

[0044] Based on the security mandatory constraint condition extracted from the power transaction rules, the legal matching pairs are screened as safety layer data;

[0045] Based on the task dependency relationship extracted from the structured data graph generated by the large model, the graph layer data is obtained;

[0046] Based on the safety layer data and the graph layer data, the structural elements of the task allocation matrix are supplemented, and the task allocation matrix is generated through the following steps:

[0047] Apply the safety layer data to filter illegal matches and limit the feasible region of task allocation; apply the graph layer data to construct a task dependency graph and determine the task scheduling topological sequence; combine the agent performance constraints to generate an initial task allocation matrix;

[0048] Taking the task completion time, agent load balancing degree and data interaction security as reward factors, a normalized objective function is constructed to solve the task allocation matrix:

[0049] And based on the three-layer constraint rules, the objective function is constrained; the three-layer constraint rules include safety mandatory constraint, task sequence constraint and agent performance constraint.

[0050] Further, the power transaction data is encrypted and stored as evidence, and a transaction certificate containing a timestamp and a digital signature is generated, including:

[0051] Based on the large model, the semantic analysis of the power transaction data is performed, and the enterprise ID of the sender and the receiver is extracted;

[0052] The encryption agent generates a sender identity key and a receiver identity key based on an SM2 algorithm according to a sender enterprise ID, a receiver enterprise ID and a main key of the power transaction security center;

[0053] The sender agent generates a sender session random number, encrypts the random number by calling the identity key of the receiver, and sends the encrypted ciphertext to the receiver;

[0054] The receiver agent receives the encrypted ciphertext, decrypts it using its own identity key, and extracts the sender session random number;

[0055] The receiver agent generates a receiver session random number, encrypts the random number by calling the identity key of the sender, and returns the encrypted ciphertext to the sender;

[0056] After the sender and the receiver obtain each other's session random numbers, they are spliced and combined in a fixed order based on the sender session random number and the receiver session random number, and a dynamic session key is generated based on an SM3 hash algorithm;

[0057] The power transaction data is divided into a plurality of data blocks, and the time series generated based on the data blocks are sorted;

[0058] Each data block is encrypted based on the dynamic session key using an SM4 algorithm to obtain an encrypted data block;

[0059] The verification agent splices the encrypted data blocks in order, generates a hash value based on an SM3 algorithm, and encrypts the hash value using the sender private key to generate a digital signature;

[0060] The evidence storage agent obtains a time stamp from the time center, generates a time stamp record based on the sender enterprise ID, the receiver enterprise ID and the dynamic session key;

[0061] A transaction voucher is generated based on the time stamp, the digital signature, the encrypted data block and the sender and receiver enterprise IDs.

[0062] The present application provides another technical solution, a power transaction data interaction system based on large model and agent cooperation, comprising:

[0063] A multi-modal feature processing module is configured to obtain and preprocess original power transaction data based on each data interface, extract and fuse multi-modal features from the preprocessed power transaction data, generate a comprehensive feature vector, construct a structured data graph, and form a multi-modal feature space of the power transaction data;

[0064] An agent cooperative management module is configured to initialize each agent, construct an agent performance evaluation matrix, and generate a task allocation matrix for decomposition and scheduling of data interaction tasks;

[0065] An interaction process optimization module is configured to collect running data in real time when the agent performs the data interaction task and perform anomaly detection, and generate a correction strategy;

[0066] A secure transmission and verification module is configured to encrypt and store the power transaction data based on blockchain technology, generate a transaction voucher, and perform multi-dimensional verification based on a consensus protocol to generate a visual report containing transaction risk assessment.

[0067] Compared with the prior art, the beneficial effects of the present application are:

[0068] Through the cooperation of large models and agents, the efficiency of power transaction data interaction is significantly improved. The semantic representation and multi-modal feature fusion technology of large models, combined with structured data graph, provide accurate basis for data interaction. Multi-agent dynamic task allocation and hierarchical consensus verification improve task processing efficiency and accuracy, and reduce error rate. At the same time, the iterative feedback optimization mechanism ensures stable operation of the system, and blockchain encryption and cross-domain verification enhance data security. In addition, cross-domain data fusion and standardized model mapping break down format barriers, output results that meet the rules and attach risk assessment, providing unified and reliable data support for power transactions, promoting business standardization, while ensuring efficiency, security and compatibility advantages. BRIEF DESCRIPTION OF DRAWINGS

[0069] Figure 1 The flowchart of the power transaction data interaction method based on the cooperation of large models and agents of the present application. DETAILED DESCRIPTION

[0070] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor fall within the scope of the present application.

[0071] To solve the technical problems that the prior art does not combine multi-agent cooperation mechanism, it is difficult to cope with power transaction multi-agent cross-domain scenarios and realize dynamic decomposition and scheduling of tasks, and lacks data security transmission and cross-domain verification, and cannot correct anomalies in real time to ensure system stability, please refer to Figure 1 The present embodiment provides the following technical solutions:

[0072] The power transaction data interaction method based on the cooperation of large models and agents comprises:

[0073] The power transaction data multi-modal feature space is constructed: based on the large model, the original power transaction data is semantically represented, a structured data graph is constructed, and a multi-modal feature space of the power transaction data is formed;

[0074] The multi-agent collaborative framework is deployed and the task is scheduled: based on the structured data graph, a multi-agent collaborative framework is built, and based on the dynamic task allocation mechanism between agents, the obtained data interaction tasks are decomposed and scheduled;

[0075] Optimization of data interaction process: based on the iterative feedback optimization mechanism of the agent framework and the large model, the abnormal state in the data interaction process is corrected in real time, and the optimal interaction path is generated;

[0076] In this embodiment, the iterative feedback optimization mechanism comprises:

[0077] Based on the running state evaluation result of each agent, the communication delay data, data integrity index and abnormal event log in the running data are extracted;

[0078] In this embodiment, the communication delay data records the data transmission time consumption in milliseconds, the data integrity index calculates the integrity rate by comparing the hash values of the data before and after transmission, and the abnormal event log includes event types such as data verification failure and timeout non-response and time stamps;

[0079] The data integrity index and the abnormal event log are input into the large model for anomaly detection, and the historical interaction mode in the structured data graph is extracted and compared;

[0080] Based on the comparison result, the deviation degree of the current data integrity index and the average value of the historical data integrity index is obtained, when the deviation degree exceeds the preset deviation threshold, based on the abnormal type of the abnormal event log, a correction strategy containing parameter adjustment suggestion is generated;

[0081] Power transaction data interaction: based on the optimal interaction path, the secure transmission and cross-domain verification of the power transaction data are performed, and the standardized interaction result conforming to the power transaction rules is output;

[0082] In this embodiment, the secure transmission and cross-domain verification of the power transaction data comprises:

[0083] The distributed ledger technology based on the block chain is adopted to encrypt and store the power transaction data, and the transaction certificate containing the time stamp and the digital signature is generated, wherein the encryption algorithm adopts the combination of symmetric encryption and asymmetric encryption, and the symmetric key is dynamically generated through the key negotiation mechanism between agents;

[0084] The cross-domain verification agent calls the data interfaces of different power transaction subjects, and performs multi-dimensional verification based on a preset consensus protocol between agents.

[0085] The interaction result that passes the verification is mapped to a standard power transaction data model, which includes core fields such as transaction identification, participant code, power quantity, power price, transaction status, etc. The conversion of non-standard fields to standard fields is realized through a field mapping table to ensure the consistency of different domain data formats. The standard model data is risk assessed by a large model, and the price fluctuation risk value, performance probability, and cross-domain transmission security level are calculated based on the historical transaction risk records in the structured data graph to generate a visual report containing transaction risk assessment.

[0086] In this embodiment, the semantic representation and multi-modal feature fusion of power transaction data are realized by a large model to construct a structured data graph. Combined with a multi-agent dynamic task allocation mechanism, efficient decomposition and scheduling of data interaction tasks are realized, which greatly reduces the task processing delay. At the same time, the data processing accuracy is ensured through hierarchical consensus verification, and the interaction error rate is reduced. Based on the agent performance evaluation matrix and iterative feedback optimization mechanism, abnormal states can be corrected and interaction paths can be optimized in real time to ensure stable operation of the multi-agent collaborative framework. The use of blockchain encryption storage and cross-domain multi-dimensional verification effectively prevents data tampering and unauthorized access, meeting the high security requirements of power transaction data. Through multi-modal feature space construction and standard power transaction data model mapping, the barriers of different domain data formats are broken down. Combined with the risk assessment function of the large model, standardized results conforming to industry rules are generated to provide unified and reliable data interaction support for power market participants and promote the standardized development of power transaction business.

[0087] In this embodiment, a multi-modal feature space of power transaction data is constructed, including:

[0088] Based on each data interface, raw power transaction data is obtained, and the raw power transaction data is preprocessed. Multi-modal feature extraction is performed on the preprocessed raw power transaction data to generate a multi-modal feature vector containing load curve, price fluctuation, and transaction volume dynamics.

[0089] In this embodiment, the raw power transaction data mainly comes from the transaction behavior records of power market participants, the data collection of power grid operation monitoring system, and the business systems of power transaction centers. Through API interface, transaction declaration data and transaction data are obtained from the business systems of each participant in real time. The data collection terminal extracts metering and dispatching data related to power transaction from the power grid operation system. At the same time, historical transaction data is batch extracted from the historical database of the power transaction center by means of database query technology.

[0090] The original power transaction data includes structured data, unstructured text data and time series data. Numerical features of the structured data are extracted by a statistical analysis method, including mean, variance, extreme value, etc. A large model is used to extract semantic features of the unstructured text data, generating a semantic feature vector containing information such as transaction subject intent and clause constraints. A time series decomposition algorithm is used to extract trend features, periodic features and random fluctuation features of the time series data, forming a time series feature sequence.

[0091] The correlation strength of each modality feature is quantified, and the modality features with a correlation strength exceeding a preset correlation threshold are correlated and mapped. According to the importance of each modality feature in the power transaction decision, a weight value is assigned, and a comprehensive feature vector containing multi-modal information is generated by weighted fusion. The semantic correlation network of the time series feature sequence and the semantic feature vector is fused to realize efficient integration of information.

[0092] Based on the fused comprehensive feature vector, a structured data graph containing spatio-temporal attributes and business logic is generated by a graph embedding algorithm. The structured data graph takes transaction subjects, transaction targets and transaction times as core nodes, establishes connection edges between the core nodes based on the correlation of each modality feature, and adds attribute information of each core node and edge. The attribute information includes feature values, correlation strengths, timestamps, etc.

[0093] In the present embodiment, the multi-agent collaborative framework deployment and task scheduling include:

[0094] According to the functional requirements of the power transaction data interaction scene, data collection agents, data verification agents, path planning agents, security monitoring agents and task coordination agents are initialized. Each agent communicates through a distributed message bus to ensure the confidentiality of instructions and data between agents.

[0095] The parameters of the data collection agent include data collection frequency and target data source identifier. The parameters of the data verification agent include verification rule library version and verification response timeliness threshold. The parameters of the path planning agent include path calculation accuracy and network topology update period.

[0096] An agent performance evaluation matrix is constructed. Based on the agent performance evaluation matrix, multi-dimensional evaluation indicators of each agent when processing data interaction tasks are evaluated, including task processing efficiency, data processing accuracy, cross-domain adaptability, etc. The load state of each agent is determined, and the agent performance evaluation matrix is dynamically updated every 10 interaction tasks based on a preset number of data interaction tasks. When a certain dimension indicator is lower than a preset threshold, the agent capability compensation mechanism is triggered, and other agents are dispatched by the task coordination agent to share the load.

[0097] An initial task allocation strategy is generated, taking task completion time, agent load balancing degree, and data interaction security as reward factors, wherein the weight of task completion time is 0.4, the weight of load balancing degree is 0.3, and the weight of data interaction security is 0.3; a task allocation matrix is generated according to the task priority (the priority of an emergency task is 1, and the priority of a regular task is 0.5) and the load state of the agent;

[0098] A consensus protocol between agents based on a layered practical Byzantine fault tolerance algorithm is constructed:

[0099] In the same-domain agent verification, each agent performs a first verification on the task execution result based on a preset verification rule library, and the verification content includes data format specification, feature value matching degree, timestamp validity, etc., to obtain a first layer verification result, and the verification pass rate needs to reach 80%;

[0100] In cross-domain agent verification, different domain agents perform a second verification on the first layer verification result based on cross-domain adaptability parameters, check the relevance and consistency of cross-domain data, obtain a second verification result, and the verification pass rate needs to exceed 2 / 3;

[0101] In this embodiment, the cross-domain data includes structured data such as user information tables and order record tables from different regional databases, unstructured data such as text files, image files, and audio / video files generated across regions, semi-structured data such as JSON format configuration information and XML format log data exchanged between cross-domain systems, business data such as user transaction data, device operation data, and market promotion data generated in different regions, and access permission information and data descriptive metadata attached during cross-domain transmission;

[0102] In this embodiment, during the verification process, the agent packages the task execution result into a proposal message containing a digital signature, a timestamp, and a data hash value, and broadcasts it to the corresponding level of verification nodes through a distributed message bus.

[0103] The verification result is compared with the corresponding preset verification pass rate value, and if any verification result is lower than the preset verification pass rate value, an arbitration instruction is generated based on the historical interaction mode in the structured data graph and the power trading rules to guide the agent to reconstruct the task execution scheme.

[0104] In the embodiment, by constructing a multi-modal feature space of power transaction data, efficient integration of different sources and different types of data is realized, and a comprehensive feature vector is generated by combining correlation mapping and weighted fusion to construct a structured data atlas, solving the problem of inconsistent data formats and providing consistent and rich information basis for subsequent interaction, improving data utilization efficiency. Based on the agent performance evaluation matrix, the load state is dynamically mastered, and the task allocation matrix is generated by combining reinforcement learning, realizing reasonable decomposition and scheduling of tasks, and improving task processing efficiency. The consensus protocol constructed by the layered practical Byzantine fault tolerance algorithm simplifies the cross-domain verification process and ensures the verification accuracy through the same domain and cross-domain double-layer verification. When the verification fails, the arbitration instruction is generated by means of historical interaction mode and transaction rules, enhancing the fault tolerance and reliability of the system.

[0105] In the embodiment, the agent performance evaluation matrix further comprises:

[0106] When the agent performs the data interaction task, the running data of communication delay, task execution success rate, and consensus time consumption of each agent are collected in real time;

[0107] When a preset number of data interaction tasks are completed, the running state of the multi-agent collaborative framework is evaluated based on the agent performance evaluation matrix;

[0108] When the communication delay exceeds the preset continuous number of times and exceeds the preset delay threshold, the network topology reconstruction algorithm is started, and the communication link between agents is adjusted by the path planning agent;

[0109] When the task execution success rate is lower than the preset success rate threshold, the agent function parameter reconfiguration process is triggered, and the target parameters of the corresponding agent are updated by the task coordination agent according to the performance evaluation matrix, including the number of agents, communication protocol, task allocation algorithm, decision weight, resource quota, and update frequency, etc., to ensure that the running efficiency and stability of the multi-agent collaborative framework meet the real-time requirements of power transaction data interaction.

[0110] In the embodiment, the agent performance evaluation matrix collects running data such as communication delay, task execution success rate, and consensus time consumption in real time, constructs a multi-dimensional evaluation system, and forms a dynamic feedback mechanism, breaking through the static management mode, realizing the adaptive optimization of the agent collaborative framework, and continuously adapting the real-time requirements of the running efficiency and stability, significantly improving the fault tolerance and response speed in complex scenarios compared with the fixed configuration mode in the prior art.

[0111] In the embodiment, the task completion time, agent load balancing degree, and data interaction security are taken as reward factors, and the task allocation matrix is generated according to the task priority and agent load state, including:

[0112] Based on the power transaction data interaction function, the intelligent agent is divided into data collection intelligent agent, data verification intelligent agent, path planning intelligent agent, security monitoring intelligent agent and task coordination intelligent agent, and the scene instances are subdivided, and the performance index, current load rate and safety capability index of each type of intelligent agent are taken as the row dimension data of the task allocation matrix;

[0113] Based on the task type, the task is divided into transaction subject data collection task, transaction target parameter verification task, cross-domain data transmission path planning task, transaction data security monitoring task and abnormal data correction task, and a double attribute label is attached to each type of task, and the task type, task priority label and safety level label are taken as the column dimension data of the task allocation matrix; wherein the task priority label is set to three levels of high / medium / low according to the power transaction rules and quantified as a normalized weight value;

[0114] Based on the row dimension data and column dimension data of the task allocation matrix, the structural elements of the task allocation matrix are constructed;

[0115] Based on the security mandatory constraint condition extracted from the power transaction rules, the legal matching pairs are screened as safety layer data;

[0116] Based on the task dependency relationship extracted from the structured data graph generated by the large model, the graph layer data is obtained;

[0117] Based on the safety layer data and the graph layer data, the structural elements of the task allocation matrix are supplemented, and the task allocation matrix is generated through the following steps:

[0118] Apply the safety layer data to filter illegal matches and limit the feasible region of task allocation; apply the graph layer data to construct a task dependency graph and determine the task scheduling topological sequence; combine the intelligent agent performance constraints to generate an initial task allocation matrix;

[0119] Taking the task completion time, intelligent agent load balancing degree and data interaction security as reward factors, a normalized objective function is constructed to solve the task allocation matrix:

[0120] ;

[0121] Wherein, represents the optimal solution of the task allocation matrix; 、 、 is a weight coefficient, and the sum of and is 1; represents the reference time, which is taken from the minimum total completion time of historical power transactions, and is used for normalizing the time item; represents the total completion time of all power transaction tasks; represents the predicted load rate of the a-th intelligent agent; ; represents the current load rate of the a-th agent, represents the maximum load capacity of the a-th agent, represents the workload of the t-th task, represents the processing speed of the a-th agent; represents the predicted average load rate of all agents; represents the total number of agents; represents a binary label, whether the a-th agent is assigned to the t-th task, that is, the a-th agent is selected to perform the t-th task; represents that the agent a is not selected to perform the t-th task; represents the security adaptation degree of the a-th agent to the t-th task; m represents the total number of tasks; represents the priority weight of the t-th task;

[0122] and the objective function is constrained based on three-layer constraint rules; the three-layer constraint rules include security mandatory constraints, task sequence constraints, and agent performance constraints.

[0123] In this embodiment, the scene instance is specifically illustrated as follows: the data collection agent includes a thermal power collection agent and a wind power collection agent; and the security monitoring agent includes an encryption monitoring agent and a cross-domain authority monitoring agent.

[0124] In this embodiment, the performance indicators are specifically illustrated as collection rate and verification accuracy.

[0125] In this embodiment, the real-time load rate is the ratio of the current number of tasks to the maximum number of tasks.

[0126] In this embodiment, the security capability indicators include encryption algorithm level and cross-domain authority, and are obtained based on an agent performance evaluation matrix.

[0127] In this embodiment, the priority label is specifically illustrated as including emergency (such as 5-minute spot trading, with a value of 0.9) and regular (such as historical transaction account synchronization, with a value of 0.3).

[0128] In this embodiment, the security level label is specifically illustrated as including high security (such as user-side load curve, with a value of 0.8).

[0129] In this embodiment, the structural elements of the task allocation matrix, that is, the initial task allocation matrix is generated according to the task priority and the agent load state;

[0130] ;

[0131] wherein, represents a matrix element; denotes whether the agent a is assigned to the task t, denotes that the a-th agent is selected to perform the task t; denotes that the a-th agent is not selected to perform the task t; denotes the priority of the task, which is the fusion of the urgency of the task and the load of the agent, and the calculation formula is denotes the real-time load rate of the a-th agent; denotes the value of the priority label of the t-th task; denotes the security adaptation degree, that is, the matching degree of the security capability of the agent and the security demand of the task, denotes the security level of the a-th agent; denotes the security demand threshold of the t-th task.

[0132] In this embodiment, based on the extraction of security mandatory constraint conditions in the power transaction rules, the legal matching pairs are screened, and specific examples are illustrated as follows: if the cross-domain task must be assigned a security monitoring agent, then the cross-domain task and the security monitoring agent are legal matching pairs.

[0133] In this embodiment, based on the extraction of task dependency relationship in the structured data graph generated by the large model, specific examples are illustrated as follows: the execution order of the collection task→ verification task is used to support the order constraint of the matrix.

[0134] In this embodiment, the agent performance constraint, that is, the rule that the number of tasks of the agent ≤ the upper limit of the ability of the agent to process tasks, the data layer provides the capacity upper limit of the rule.

[0135] In this embodiment, the security mandatory constraint: through the data of the security layer, the feasible domain is limited to ensure that all allocation schemes meet the security rules of the power transaction.

[0136] In this embodiment, the task order constraint: based on the data of the graph layer, it is forced to require that the dependent tasks meet the topological order.

[0137] In this embodiment, the three-layer constraint rules include security mandatory constraint, task order constraint and agent performance constraint.

[0138] In this embodiment, the agent performance constraint includes agent type matching constraint (such as data collection task is only assigned to data collection agent), load capacity constraint and task workload compatibility constraint.

[0139] In this embodiment, the solution of the task allocation matrix adopts a mixed integer programming algorithm, and the three-layer constraint rules are integrated into the solving process as a set of feasibility conditions.

[0140] ​​​In this embodiment, the "cross-provincial thermal power trading data interaction" scenario is taken as an example to specify the specific instances of rows (agents) and columns (tasks):

[0141] 1. Row dimension (agent dimension): 5 types of agents + scene subdivision, as shown in Table 1:

[0142] Table 1

[0143] ;

[0144] 2. Column dimension (task dimension): 5 types of tasks + double attribute labels

[0145] Taking "cross-provincial thermal power trading" as the core, the task instances and labels are defined as shown in Table 2:

[0146] Table 2

[0147] ;

[0148] Example 1: Thermal power collection agent A → Task 1

[0149] The matrix element is , taking "thermal power collection agent A → Task 1 (thermal power enterprise A data collection)" and "cross-domain verification agent D → Task 4 (cross-provincial transaction target verification)" as examples, the calculation logic is:

[0150] Value: Task 1 is "thermal power enterprise data collection", agent A is "thermal power collection agent", function matching, and no conflict constraint → = 1.

[0151] Calculation (formula: , assuming , : .

[0152] Calculation (formula: ): The security level of agent A = 1 (AES-256), the → (truncated to 1, representing complete adaptation)

[0153] Example 2: Cross-domain verification agent D → Task 4

[0154] Value: Task 4 is "cross-provincial transaction target verification", agent D is "cross-domain verification agent", function matching; and the security mandatory constraint requires that cross-domain tasks must be assigned verification agents → .

[0155] Compute

[0156]

[0157] , Task 4 of (truncated to 1, representing a perfect fit)

[0158] Example 3: Constraint-driven Flag: Cross-domain Permission Monitoring Agent H→Task 4

[0159] Task 4 is a cross-domain task, and the security enforcement constraint requires that "a cross-domain task must be assigned a cross-domain permission monitoring agent"→ =1 (must be assigned even if H's load is low).

[0160] Compute ( =0.9, =0.1): =0.6x0.9+0.4x(1-0.1)=0.54+0.36=0.90.

[0161] Compute (H's security level =1, =0.8): =1 / 0.8=1.25 (truncated to 1).

[0162] The working principle and beneficial effects of the technical solution are as follows: taking the task completion time as the reward factor, the intelligent agent can be reasonably arranged to perform the task by constructing a task allocation matrix and solving the optimal solution, and unnecessary time waste is reduced; it helps to prevent some intelligent agents from performance degradation due to task overload, while avoiding the situation that other intelligent agents are idle, and improves the utilization rate of intelligent agent resources in the whole system; the data interaction security is taken as the reward factor, and the security adaptation degree of the intelligent agent to the task is fully considered in the construction and solution process of the task allocation matrix; based on the security constraint condition in the power transaction rule, the legal matching pair is screened as the security layer data, and the task dependency relationship is extracted from the structured data graph generated by the large model as the graph layer data, which supplements the task allocation matrix; this makes the task allocation not only consider the attributes of the intelligent agent and the task itself, but also meet the safety requirements in the power transaction rule and the logical relationship between tasks, improving the rationality and accuracy of task allocation; through detailed classification of intelligent agents (such as data collection intelligent agents, data verification intelligent agents, etc.) and detailed division of tasks (such as transaction subject data collection tasks, transaction target parameter verification tasks, etc.), and considering the double attribute labels of the task (task priority label and security level label), various complex task types and scene requirements in the power transaction data interaction can be flexibly adapted to realize more accurate task allocation; when constructing the objective function, the task completion time, the intelligent agent load balancing degree and the data interaction security are comprehensively considered, and the weight coefficient is adjusted to improve the overall performance of the whole power transaction data interaction system.

[0163] In the embodiment, the power transaction data is encrypted and stored, and a transaction voucher containing a timestamp and a digital signature is generated, including:

[0164] Based on the large model, the semantic analysis of the power transaction data is performed, and the enterprise IDs of the sender and the receiver are extracted;

[0165] The encryption intelligent agent generates the sender identity key and the receiver identity key based on the SM2 algorithm according to the sender enterprise ID, the receiver enterprise ID and the master key of the power transaction security center;

[0166] The sender intelligent agent generates a sender session random number, encrypts the random number by calling the identity key of the receiver, and sends the encrypted ciphertext to the receiver;

[0167] The receiver intelligent agent receives the encrypted ciphertext, decrypts it using its own identity key, and extracts the sender session random number;

[0168] The receiver intelligent agent generates a receiver session random number, encrypts the random number by calling the identity key of the sender, and returns the encrypted ciphertext to the sender;

[0169] The sender and the receiver obtain the session random number of the other party respectively, and then splice and combine the sender session random number and the receiver session random number in a fixed order to generate a dynamic session key based on an SM3 hash algorithm;

[0170] The power transaction data is divided into several data blocks, and the time sequence generated based on the data blocks is sorted;

[0171] Each data block is encrypted based on the dynamic session key by using an SM4 algorithm to obtain an encrypted data block;

[0172] The agent verifies that the encrypted data block is spliced in the order according to the sorting, generates a hash value based on the SM3 algorithm, and then encrypts the hash value by using the private key of the sender to generate a digital signature;

[0173] The evidence storage agent obtains the time stamp of the time center, and generates a time stamp record based on the sender enterprise ID, the receiver enterprise ID and the dynamic session key.

[0174] The transaction voucher is generated based on the time stamp, the digital signature, the encrypted data block and the enterprise IDs of the sender and the receiver.

[0175] In this embodiment, SM2 (elliptic curve public key algorithm) is formulated by the China National Cryptography Administration.

[0176] In this embodiment, the SM3 hash algorithm is a commercial cryptographic hash algorithm standard published by the China National Cryptography Administration in 2010, and belongs to the national cryptographic algorithm system, which, together with SM2 (elliptic curve public key algorithm) and SM4 (symmetric block encryption algorithm), constitutes a cryptographic system with independent intellectual property rights of China.

[0177] The working principle and beneficial effects of the above technical solution are: by analyzing the power transaction data based on a large model to obtain the enterprise ID, and then generating the sender and receiver identity key using the SM2 algorithm; this process ensures the authenticity and uniqueness of the identities of both parties to the transaction, preventing illegal users from participating in power transactions; the generation of the identity key is based on the enterprise ID and the power transaction security center master key, increasing the security and reliability of the key and laying the foundation for subsequent encrypted communication; the sender and receiver exchange session random numbers and concatenate them in a fixed order based on the SM3 hash algorithm to generate a dynamic session key; this dynamic session key generation method has high security because the session key is randomly generated based on the interaction of both parties, and the session key for each transaction is different, greatly reducing the risk of key cracking and effectively protecting the security of power transaction data during transmission; using the SM4 algorithm to encrypt the segmented power transaction data blocks based on the dynamic session key ensures the confidentiality of the data content. At the same time, the verification agent generates a hash value and encrypts it with the sender's private key to obtain a digital signature, which ensures the integrity and non-repudiation of the data. Once the data is tampered with, the digital signature will not pass the verification, thereby ensuring the authenticity and integrity of the power transaction data; based on the timestamp, digital signature, encrypted data block, and enterprise ID, a transaction voucher is generated, which contains key information about the power transaction. It completely records the identities of both parties to the transaction, the encryption of the transaction data, the integrity verification of the data (digital signature), and the time information of the transaction, making it easy to provide comprehensive and reliable evidence when there is a dispute or the need for auditing in a power transaction, enhancing the traceability and credibility of the power transaction.

[0178] The power transaction data interaction system based on large model and agent cooperation includes:

[0179] The multi-modal feature processing module is configured to obtain and preprocess the original power transaction data based on each data interface, extract and fuse the multi-modal features of the preprocessed power transaction data, generate a comprehensive feature vector, construct a structured data graph, and form a multi-modal feature space of the power transaction data;

[0180] The agent cooperation management module is configured to initialize each agent, construct an agent performance evaluation matrix, and generate a task allocation matrix for the decomposition and scheduling of data interaction tasks;

[0181] The interaction process optimization module is configured to collect real-time running data of the agents performing data interaction tasks and perform anomaly detection to generate a correction strategy;

[0182] The secure transmission and verification module is configured to encrypt and store the power transaction data based on blockchain technology, generate a transaction voucher, and perform multi-dimensional verification based on a consensus protocol to generate a visual report containing transaction risk assessment.

[0183] In the embodiment, the multi-modal feature processing module serves as a data foundation layer, obtains raw data from each interface and pre-processes, generates a comprehensive feature vector through multi-modal feature extraction and fusion, constructs a structured data atlas, and provides a unified and rich multi-modal feature space; the intelligent agent collaborative management module is responsible for intelligent agent initialization and task scheduling, masters the intelligent agent state through the construction of a performance evaluation matrix, generates a task allocation matrix to realize reasonable task decomposition and scheduling, and plays a collaborative advantage; the interaction process optimization module collects running data in real time, detects abnormalities and generates a correction strategy, and dynamically optimizes the interaction process in combination with a large model; the secure transmission and verification module encrypts and stores data based on a blockchain, generates a certificate, verifies and generates a report containing risk assessment according to a consensus protocol, and ensures safety and compliance.

[0184] The above merely describes a preferred embodiment of the present application, but the protection scope of the present application is not limited thereto, and any person skilled in the art can make equivalent replacements or changes to the technical solution and the inventive concept of the present application within the technical scope disclosed by the present application, which should be covered within the protection scope of the present application.

Claims

1. A power transaction data interaction method based on large model and agent cooperation, characterized in that, Comprise: Electric power transaction data multi-modal feature space construction: based on the original electric power transaction data semantic representation of large model, structured data atlas is constructed, and the multi-modal feature space of electric power transaction data is formed; Multi-agent collaborative framework deployment and task scheduling: based on the structured data atlas, multi-agent collaborative framework is built, and based on the dynamic task allocation mechanism between agents, the obtained data interaction task is decomposed and scheduled; Data interaction process optimization: based on the iterative feedback optimization mechanism of agent framework and large model, the abnormal state in data interaction process is corrected in real time, and the optimal interaction path is generated; Electric power transaction data interaction: based on the optimal interaction path, the safe transmission and cross-domain verification of electric power transaction data are carried out, and the standardized interaction result meeting the electric power transaction rules is outputted; Constructing the multi-modal feature space of electric power transaction data, comprising: Based on each data interface, the original electric power transaction data is obtained, and the original electric power transaction data is preprocessed, and the multi-modal feature extraction is carried out on the preprocessed original electric power transaction data to generate multi-modal feature vector; Quantify the correlation strength of each modal feature, correlate and map the modal features with correlation strength exceeding the preset correlation threshold, and weight and fuse according to the distribution weight value of each modal feature to generate a comprehensive feature vector; Based on the fused comprehensive feature vector, a structured data graph is generated through graph embedding algorithm, wherein the structured data graph takes transaction subject, transaction target and transaction time as core nodes, establishes connection edges between each core node according to the correlation of each modal feature, and adds attribute information of each core node and edge; The multi-agent collaborative framework deployment and task scheduling, comprising: According to the functional requirements of electric power transaction data interaction scene, initialize data collection agent, data verification agent, path planning agent, security monitoring agent and task coordination agent, and each agent communicates through distributed message bus; Construct an agent performance evaluation matrix, evaluate the multi-dimensional evaluation indexes of each agent when processing data interaction task based on the agent performance evaluation matrix, determine the load state of each agent, and dynamically update the agent performance evaluation matrix based on the preset data interaction task quantity; Generate an initial task allocation strategy, take task completion time, agent load balancing degree and data interaction security as reward factors, and generate a task allocation matrix according to task priority and agent load state.

2. The method of claim 1, wherein the method further comprises: The multi-agent collaborative framework deployment and task scheduling further comprise constructing consensus protocol between agents: In the same domain agent verification, each agent performs first verification on the task execution result based on the preset verification rule library to obtain the first layer verification result; In cross-domain agent verification, agents in different domains perform second verification on the first layer verification result based on cross-domain adaptability parameters to check the correlation and consistency of cross-domain data, and obtain the second verification result; The verification result is compared with the corresponding preset verification passing rate value, and if any verification result is lower than the preset verification passing rate value, an arbitration instruction is generated based on the historical interaction mode in the structured data graph and the power transaction rule to guide the agent to reconstruct the task execution scheme.

3. The method of claim 2, wherein the method further comprises: The agent performance evaluation matrix further includes: When the agents perform data interaction tasks, the communication delay, task execution success rate, and consensus achievement time of each agent are collected in real time; The running state of the multi-agent collaborative framework is evaluated based on the agent performance evaluation matrix every time a preset number of data interaction tasks is completed; When the communication delay exceeds the preset continuous number of times and exceeds the preset delay threshold, the path planning agent adjusts the communication link between the agents; When the task execution success rate is lower than the preset success rate threshold, the task coordination agent updates the target parameters of the corresponding agent according to the performance evaluation matrix.

4. The method of claim 3, wherein the method further comprises: The iterative feedback optimization mechanism includes: Based on the running state evaluation results of each agent, the communication delay data, data integrity indicators, and abnormal event logs in the running data are extracted; The data integrity indicators and abnormal event logs are input into the large model for anomaly detection, and the historical interaction mode in the structured data graph is extracted and compared; Based on the comparison result, the deviation degree of the current data integrity indicator and the historical data integrity indicator average value is obtained, and when the deviation degree exceeds the preset deviation threshold, a correction strategy is generated based on the abnormal type of the abnormal event log.

5. The method of claim 4, wherein the method further comprises: The secure transmission and cross-domain verification of the power transaction data includes: The power transaction data is encrypted and stored as evidence to generate a transaction voucher containing a timestamp and a digital signature; Different data interfaces of different power transaction subjects are called, and multi-dimensional verification is performed based on the preset consensus protocol between agents; The interaction result that passes the verification is mapped to a standard power transaction data model, and the standard model data is risk assessed by the large model, combined with the historical transaction risk records in the structured data graph, to calculate the price fluctuation risk value, the performance probability, and the cross-domain transmission security level, and generate a visual report containing transaction risk assessment.

6. The method of claim 5, wherein the method further comprises: The task allocation matrix is generated based on the task priority and the agent load state, taking the task completion time, the agent load balancing degree, and the data interaction security as reward factors, including: Based on the power transaction data interaction function, the agents are divided into data collection agents, data verification agents, path planning agents, security monitoring agents, and task coordination agents, and are further divided into scene instances, and the performance indicators, current load rate, and security capability indicators of each type of agent are taken as the row dimension data of the task allocation matrix; Based on the task type, the tasks are divided into transaction subject data collection tasks, transaction target parameter verification tasks, cross-domain data transmission path planning tasks, transaction data security monitoring tasks, and abnormal data correction tasks, and each type of task is attached with a double attribute label, and the task type, task priority label, and security level label are taken as the column dimension data of the task allocation matrix; wherein the task priority label is set to three levels of high / medium / low according to the power transaction rules and is quantified as a normalized weight value; The structural elements of the task allocation matrix are constructed based on the row dimension data and the column dimension data of the task allocation matrix; Based on the extraction of security mandatory constraints in the power transaction rules, the legal matching pairs are screened as security layer data; Based on the extraction of task dependency relationships in the structured data graph generated by the large model, the graph layer data is obtained; Based on the security layer data and the graph layer data, the structural elements of the task allocation matrix are supplemented, and the task allocation matrix is generated through the following steps: Apply the security layer data to filter illegal matches and limit the feasible region of task allocation; apply the graph layer data to construct a task dependency graph and determine the task scheduling topological sequence; combine the agent performance constraints to generate an initial task allocation matrix; Taking the task completion time, agent load balancing degree and data interaction security as reward factors, a normalized objective function is constructed to solve the task allocation matrix: And based on the three-layer constraint rules, the objective function is constrained; the three-layer constraint rules include security mandatory constraints, task sequence constraints and agent performance constraints.

7. The method of claim 6, wherein the method further comprises: The power transaction data is encrypted and stored as evidence to generate a transaction certificate containing a timestamp and a digital signature, including: Based on the large model, the semantic analysis of the power transaction data is performed to extract the enterprise IDs of the sender and the receiver; The encryption agent generates a sender identity key and a receiver identity key based on the SM2 algorithm according to the sender enterprise ID, the receiver enterprise ID and the master key of the power transaction security center; The sender agent generates a sender session random number and encrypts it using the receiver's identity key, then sends the encrypted ciphertext to the receiver; The receiver agent receives the encrypted ciphertext, decrypts it using its own identity key, and extracts the sender session random number; The receiver agent generates a receiver session random number and encrypts it using the sender's identity key, then sends the encrypted ciphertext back to the sender; After the sender and the receiver obtain each other's session random numbers, they are spliced and merged in a fixed order based on the sender session random number and the receiver session random number, and a dynamic session key is generated based on the SM3 hash algorithm; The power transaction data is divided into several data blocks, and the time series generated by the data blocks are sorted; Each data block is encrypted using the SM4 algorithm based on the dynamic session key to obtain the encrypted data block; The verification agent splices the encrypted data blocks in the sorted order, generates a hash value based on the SM3 algorithm, and then encrypts the hash value using the sender's private key to generate a digital signature; The storage agent obtains the time stamp from the time center, generates a time stamp record based on the sender enterprise ID, the receiver enterprise ID and the dynamic session key; Based on the timestamp, digital signature, encrypted data block and sender and receiver enterprise ID, a transaction certificate is generated.

8. The power transaction data interaction system based on the cooperation of large models and agents, applied in the power transaction data interaction method based on the cooperation of large models and agents as claimed in claim 7, characterized in that, It includes: A multi-modal feature processing module is configured to obtain and preprocess original power transaction data based on each data interface, extract and fuse multi-modal features from the preprocessed power transaction data, generate a comprehensive feature vector, construct a structured data graph, and form a multi-modal feature space of the power transaction data; The intelligent agent cooperative management module is configured to initialize each intelligent agent, construct an intelligent agent performance evaluation matrix, and generate a task allocation matrix for decomposition and scheduling of data interaction tasks. The interaction process optimization module is configured to collect running data in real time when the intelligent agent executes the data interaction task and perform anomaly detection, and generate a correction strategy. The secure transmission and verification module is configured to encrypt and store evidence of power transaction data based on blockchain technology, generate transaction credentials, and perform multi-dimensional verification based on a consensus protocol to generate a visual report containing transaction risk assessment.

Images