Metodología para la elección de modelos de LLMS en aplicaciones de predicción metereológicas a través de algoritmos de ML sobre entornos de computación en la Nube y capturas de datos a través de IoT

Abstract

Meteorological prediction is one of the critical factors addressed from various approaches and is fundamental for a wide range of sectors, such as agriculture, renewable energy, disaster management, and urban planning. Recent advances in Large Language Models (LLMs), Internet of Things (IoT), and cloud computing have opened new opportunities to improve the accuracy and efficiency of predictions in these sectors. However, there are several challenges related to the constant variability of environmental conditions and the reliability of data obtained from sensors. This research proposes the development of a comprehensive methodology to evaluate the impact of integrating LLMs with IoT infrastructures and cloud computing, with the aim of determining precision and improving the accuracy of meteorological predictions. The methodology comprises five iterative phases: Identification, Development, Testing and Monitoring, Evaluation, and Analysis. This approach allows for the continuous evaluation of LLMs and the adaptation of the system based on the obtained results, addressing the changing needs of the IoT environment. The study focuses on designing specific metrics to evaluate the performance of LLMs compared to traditional models, deployed within a scalable cloud platform that facilitates the integration of data generated by IoT devices. The methodology incorporates the use of a ReAct (Reasoning and Acting) agent, which improves the system's precision and accuracy by detecting anomalies in the data and adjusting responses accordingly. This agent also demonstrated the ability to identify when the model's performance was insufficient, recommending the use of more reliable data sources as an alternative to ensure the quality of predictions. In the case study, it was evident that some models exhibited low performance, with metrics such as R² close to zero, indicating an inability to capture underlying patterns in the data. However, the inclusion of the ReAct agent mitigated these problems by making critical decisions to maintain the quality of predictions. The results demonstrated the system's ability to adjust and improve as new data is collected, making the process adaptive and more robust. It is expected that the results of this research will significantly contribute to the advancement of meteorological prediction, with direct benefits for critical sectors and various stakeholders. The developed methodology lays the foundation for future research and applications in this field, facilitating more accurate and reliable meteorological predictions. The combination of LLMs with IoT and reactive agents not only enhances predictive capability but also the system's adaptability in changing environments, which is essential for modern meteorological applications.