Aprendizado de máquina com variáveis categóricas para previsão de vazão em bacias hidrográficas de Minas Gerais

Abstract

The sustainable management of water resources is essential in the face of climate change and its impacts on human and natural systems. In this context, computational models have supported decision-making processes; however, given the limitations of traditional physical models – which often require extensive data and complex calibration processes – it raises the question: to what extent can machine learning models provide robust and practical forecasts for water management in scenarios characterized by high variability and climate change? This study investigates the application of machine learning models, including Linear Regression, CatBoost, and Random Forest, for runoff forecasting in the hydrographic basins of Minas Gerais, Brazil. By analyzing the complex relationships between precipitation and river flow, this research aims to improve forecasting accuracy by incorporating specific categorical variables and addressing challenges such as data asymmetry and the “long tail” effect. The results highlight the potential of machine learning, particularly Linear Regression, as an efficient and cost-effective solution, offering valuable insights for water resource management and public policy planning. Furthermore, this study contributes to the scientific literature by filling a knowledge gap regarding the application of machine learning techniques in this regional context.