Development of a multi-agent system based on Large Language Models for solar radiation prediction using meteorological data

Abstract

This short article presents a work in progress based on a hybrid approach for the estimation of hourly global solar radiation under variable sky conditions, combining supervised machine learning techniques and generative artificial intelligence tools. Specifically, a predictive model based on Gradient Boosting algorithms is developed, using as input variables meteorological data and the clarity index (Kt), a parameter derived from solar radiation, which allows classifying cloud cover and improving model accuracy. Hourly global solar radiation is the target variable of the system. The best performing model, evaluated using metrics such as the coefficient of determination (R²), mean absolute error (MAE) and root mean square error (RMSE), is integrated into an interactive chatbot designed to facilitate its use and promote access to exploratory analysis without technical requirements from the user. For example, a user can query: "What was the average solar radiation in April 2022?" and receive an answer accompanied by an automatic visualization. This system is implemented using Microsoft's Semantic Kernel technology, which enables the execution of programmed functions based on natural language interpretation, and is complemented by local language models managed through the Ollama platform, including instances of LLaMA and Qwen. The proposed solution improves the accuracy of solar radiation prediction and facilitates access to scientific models through intelligent and accessible interfaces.