A multi-agent runoff prediction method and system based on a large language model drive

Abstract

This invention discloses a multi-agent runoff prediction method and system based on a large language model, relating to the fields of artificial intelligence and hydrological prediction technology. The method includes: receiving natural language requests and parsing them into standard tasks; searching a scheme library; if no match is found, entering a cold start phase, scheduling data processing, forecast factor construction, and model computation agents to collaboratively construct a new scheme, employing a structure combining outer-layer process orchestration and inner-layer intelligent decision-making; if a match is found, entering a hot run phase, intercepting large model requests and directly driving the execution of underlying functions; after prediction, extracting physical features such as rainfall lag time, and performing multiple verifications based on causality and response logic. This invention, through a dual-track cold / hot start mechanism and physical consistency constraints, solves the problems of lacking physical common sense and high inference latency in general large models, achieving adaptive construction and highly reliable application of runoff prediction.

Images