Avaliação de modelos de previsão de vazão baseados em aprendizagem de máquinas - Reservatório Fronteiras, Crateús-CE

Abstract

The semi-arid region of Northeastern Brazil is characterized by challenging climatic and hydrological conditions, such as low rainfall and irregular rain patterns. These factors, combined with the local geology, hinder the natural perennial flow of rivers and the recharge of aquifers. In response, the construction of artificial reservoirs, exemplified by the Fronteiras reservoir in Ceará, is crucial for the region's water sustainability. Consequently, predicting inflow to these reservoirs is a key element in the efficient management of water resources, where forecasts and uncertainties must be analyzed for system operation and planning. This study primarily aims to evaluate the predictability of the annual inflow regime to the Fronteiras Reservoir, based on the variability of Sea Surface Temperature (SST) indices using artificial neural networks and linear regression models. Predictor periods such as September- October-November (SON), June-July-August (JJA), March-April-May (MAM), and DecemberJanuary-February (DJF) were used, ranking different SST indices by relevance using machine learning techniques (Random Forest, Select K Best, and XGBoost). Prediction models were then defined using Multilayer Perceptron (MLP), Long Short-Term Memory (LSTM), Lasso and Ridge Regressions, and Multiple Linear Regression (MLR). In the results, the most significant factors identified were the Tropical Atlantic Dipole (TAD) and SST anomalies associated with it, specifically the North and South Tropical Atlantic. The most effective models for predicting flow belong to the SON predictor period, with the MLP-M7 model standing out, achieving a Nash-Sutcliffe Efficiency (NSE) of 0.329 and a correlation of 0.638 during the validation phase. Finally, despite satisfactory correlations, the NSE values were considered moderately low.