CN-121996373-A - Multi-scene trend prediction system and method based on agent cooperation and small model scheduling

CN121996373ACN 121996373 ACN121996373 ACN 121996373ACN-121996373-A

Abstract

The embodiment of the specification provides a multi-scene trend prediction system and method based on agent cooperation and small model scheduling, wherein the system comprises a data perception module, a task planning module, a model scheduling module, a model execution module and a model execution module, wherein the data perception module is used for acquiring original data through a plurality of data agents, the task planning module is used for generating a task decomposition scheme through planning agents, the model scheduling module is used for selecting the most suitable small model combination through the models, determining execution sequence and distribution nodes of each small model through the scheduling agents to generate a scheduling plan, the model execution module is used for carrying out reasoning calculation according to the small models appointed by the execution agents according to scheduling plan loading, carrying out input standardization, model loading, prediction operation and result output, and the result fusion and interpretation module is used for fusing the output results of the small models through the result fusion agents and carrying out interpretation analysis on the comprehensive prediction result through the interpretation analysis agents.

Inventors

HU YAOYU
LIANG JINHUA
LIU KAI
DING BANGLIN
GONG YUQIAN

Assignees

北京京能信息技术有限公司

Dates

Publication Date: 20260508
Application Date: 20251226

Claims (10)

1. The utility model provides a multiscreen trend prediction system based on agent cooperation and small model dispatch which characterized in that, the system specifically includes: The data perception module is used for acquiring original data from an enterprise internal business system, an external interface or equipment through a plurality of data intelligent agents and recording the original data to a task blackboard; the task planning module is connected with the data perception module and is used for reading task blackboard information through a planning agent, analyzing a predicted task initiated by a user or a system and generating a task decomposition scheme; the model scheduling module is connected with the task planning module and is used for selecting the most suitable small model combination from the model registry according to a task decomposition scheme and input data characteristics by a model selection agent; The model execution module is connected with the model scheduling module and is used for performing inference calculation by executing the small model appointed by the agent according to the scheduling plan loading, and performing input standardization, model loading, prediction operation and result output; the result fusion and interpretation module is connected with the model execution module and used for fusing the output results of the small models through the result fusion agent according to the confidence coefficient and the historical error of each model to obtain a comprehensive prediction result, and outputting main characteristic contribution, abnormal reasons and confidence intervals through the interpretation analysis agent for performing interpretation analysis on the comprehensive prediction result.
2. The system of claim 1, wherein the system further comprises: The feedback and self-learning module is used for continuously monitoring the prediction precision and the system performance through the feedback agent, automatically triggering the model correction, parameter adjustment or retraining when the error is higher and/or the data distribution is drifted, and storing new data samples, model expression and optimization records into a knowledge base through the knowledge updating agent so as to realize the continuous learning and optimization of the system.
3. The system of claim 1, wherein the data perception module specifically comprises: the data acquisition unit is used for acquiring service indexes, log information and environment data in real time; The data cleaning unit is used for carrying out format conversion, anomaly detection and deletion repair on the acquired data; And the feature generation unit is used for automatically generating time features, period features and related statistical features according to the data types, outputting normalized feature data and registering the normalized feature data to the task blackboard.
4. The system according to claim 1, wherein the mission planning module is specifically configured to: According to task targets, prediction periods, service fields and precision requirements, a complex prediction task is split into a plurality of subtasks, and task structures and dependency relations are registered in a task blackboard.
5. The system of claim 1, wherein each agent of the system performs cooperative interaction in an asynchronous cooperative manner, and adopts a double-layer communication structure of a message queue and a task blackboard, wherein the message queue is used for transmitting task events and state signals to realize parallel processing, and the task blackboard is used for recording task states, model configuration and operation results to realize data sharing and state synchronization among agents.
6. A multi-scenario trend prediction method based on agent cooperation and small model scheduling, characterized in that it is used in the multi-scenario trend prediction system based on agent cooperation and small model scheduling according to any one of claims 1to 5, and the method specifically comprises: Collecting original data from an enterprise internal business system, an external interface or equipment through a plurality of data agents, and recording the original data to a task blackboard; Reading task blackboard information by planning an agent, analyzing a predicted task initiated by a user or a system, and generating a task decomposition scheme; Selecting the most suitable small model combination from a model registry by the model selection agent according to the task decomposition scheme and the input data characteristics; determining the execution sequence and the distribution node of each small model by a scheduling agent according to the system calculation force state, the task priority and the target constraint, and generating a scheduling plan; Performing inference calculation by executing the agent according to the scheduling plan loading specified small model, and executing input standardization, model loading, prediction operation and result output; The result fusion agent fuses the output results of the small models according to the confidence coefficient and the historical error of each model to obtain a comprehensive prediction result, and the interpretation analysis agent carries out interpretation analysis on the comprehensive prediction result to output main characteristic contribution, abnormal reasons and confidence intervals.
7. The method according to claim 6, wherein the method further comprises: The method comprises the steps of continuously monitoring prediction precision and system performance through a feedback intelligent agent, automatically triggering model correction, parameter adjustment or retraining when errors are higher and/or data distribution is drifted, and storing new data samples, model expression and optimization records into a knowledge base through a knowledge updating intelligent agent so as to realize continuous learning and optimization of the system; The intelligent agents perform cooperative interaction in an asynchronous cooperative mode, and a double-layer communication structure of a message queue and a task blackboard is adopted, wherein the message queue is used for transmitting task events and state signals to realize parallel processing, and the task blackboard is used for recording task states, model configuration and operation results to realize data sharing and state synchronization among the intelligent agents.
8. The method of claim 6, wherein collecting raw data from the enterprise internal business system, external interface or device by a plurality of data agents, recording the raw data to the task blackboard specifically comprises: The method comprises the steps of collecting service indexes, log information and environment data in real time, carrying out format conversion, anomaly detection and missing repair on the collected data, automatically generating time characteristics, period characteristics and related statistical characteristics according to data types, outputting standardized characteristic data, and registering the standardized characteristic data to a task blackboard; The task blackboard information is read by planning an agent, a predicted task initiated by a user or a system is analyzed, and the task decomposition scheme is generated specifically by the following steps: According to task targets, prediction periods, service fields and precision requirements, a complex prediction task is split into a plurality of subtasks, and task structures and dependency relations are registered in a task blackboard.
9. An electronic device comprising a memory, a processor and a computer program stored on the memory and executable on the processor, the computer program implementing the steps of the multi-scenario trend prediction method according to any one of claims 6 to 8 based on agent collaboration and small model scheduling when executed by the processor.
10. A computer-readable storage medium, wherein a program for implementing information transfer is stored on the computer-readable storage medium, and the program when executed by a processor implements the steps of the multi-scenario trend prediction method based on agent cooperation and small model scheduling according to any one of claims 6 to 8.

Description

Multi-scene trend prediction system and method based on agent cooperation and small model scheduling Technical Field The present document relates to the technical field, and in particular, to a multi-scenario trend prediction system and method based on agent cooperation and small model scheduling. Background With the deep application of artificial intelligence technology in enterprise digitization, the dependence degree of enterprises on business trend prediction is continuously improved. Whether the energy power generation load, market demand prediction, supply chain inventory scheduling, financial income expenditure and public opinion trend analysis, the system is required to realize high-precision, low-delay and interpretable prediction results under different business scenes. However, the existing trend prediction system has technical bottlenecks and pain points commonly existing in multi-scene application, namely the single model generalization capability is insufficient, the traditional prediction method is mainly used for modeling by a single model or a similar algorithm, the data distribution, seasonality, emergency and external driving factors of different business scenes are obviously different, and the single model is difficult to consider the prediction precision and stability of all scenes, so that the generalization capability is weak. The model training and deployment cost is high, most prediction systems need repeated modeling and parameter adjustment in different service lines, the model mobility is poor, the calculation power waste and the development period are long, a unified model registration and scheduling mechanism is lacked, and the system multiplexing rate is low. The forecast resource utilization unevenness and response instability are that the traditional deployment mode is mostly static configuration, flexible scheduling cannot be carried out according to real-time task load, calculation resource state and Service Level Objective (SLO), so that the idle or bottleneck of calculation resource is caused, and the forecast response time and precision are difficult to guarantee under peak tasks. The multi-scene data has serious characteristic isomerization, namely, the data structures, calibers and time granularity of different service fields are inconsistent, a unified data management and characteristic engineering proxy mechanism is lacked, the characteristic extraction and cleaning processes are complex and difficult to multiplex, and the model effect and the online efficiency are affected. The prediction process lacks intelligent coordination and self-adaptive optimization, and most of the existing systems adopt a fixed model or manual configuration mode, so that intelligent Agent coordination decision based on task characteristics and real-time evaluation results cannot be realized. Model selection, parameter adjustment, integration and interpretation are still highly dependent on human experience, and it is difficult to support dynamic scene changes. The model prediction result has poor interpretability, and in key business scenes such as finance, energy scheduling and the like, the traditional model output lacks causal interpretation and confidence evaluation, cannot clearly influence key variables of trend, and is difficult to meet compliance, wind control and audit requirements of enterprises. The online learning and closed-loop optimization capability is weak, most systems only provide result display after prediction, and lack of online evaluation and self-learning mechanisms based on real-time feedback, so that the model cannot correct deviation in time after drifting along with time, and the prediction accuracy is reduced. With the rapid development of Artificial Intelligence (AI), data Intelligence (DI), cloud computing, edge computing, and the like, enterprise-level applications are moving from "process digitizing" to "intelligent decision making". Trend prediction has become a core technical means for supporting strategic decisions and business operations in a plurality of industrial scenes such as energy management, financial wind control, marketing, supply chain optimization, manufacturing scheduling and the like. Enterprises need to rely on multi-source heterogeneous data (including historical time sequence data, external event data, environment variables, policy information and the like) to model and predict future trends so as to realize capacity planning, resource allocation, cost optimization and risk early warning. However, the existing trend prediction system still mainly depends on a single algorithm or a fixed model, such as ARIMA, prophet, XGBoost, LSTM, and the like, and the models perform well in a single scene, but have obvious defects in multi-scene, multi-dimensional and multi-task environments, namely, the data characteristic difference is large, the scene complexity is high, the generalization capability of the single model is insufficient, the stable