Predicting Water Levels Using Gradient Boosting Regressor and LSTM Models: A Case Study of Lago de Chapala Dam

Abstract

This paper explores the use of machine learning methodologies, specifically Gradient Boosting Regressor (GBR) and Long Short-Term Memory (LSTM) networks, to forecast water levels at the Lago de Chapala Dam. The research highlights the importance of integrating a 1-day lag feature to provide temporal context, significantly enhancing model performance. Analyzing daily records from 1991 to 2024, including variables like precipitation, evaporation, and temperature, the research shows that models with the 1-day lag feature achieved markedly better accuracy. The GBR model’s testing RMSE decreased from 886.37 to 32.49, and the LSTM model’s RMSE dropped from 957.689 to 67.474 with the lag feature, along with substantial improvements in $\hbox {R}^{2}$ values. These findings highlight the critical role of temporal dependencies in hydrological predictions, demonstrating the potential for integrating machine learning models into water resource management to improve prediction accuracy and ensure sustainable water supply and disaster readiness. The study aims to develop and validate predictive models, evaluate their accuracy and reliability, and provide insights for integrating these models into water management practices at Lago de Chapala.