Seasonal forecasting of inflows to the Carnaubal Reservoir in the Brazilian semi-arid region using artificial neural networks

Abstract

Given the challenges posed by the climatic conditions of semi-arid regions that rely on reservoir-based water management policies to ensure water availability, machine learning methods have been increasingly used as tools for efficient water resources management. This study aimed to calibrate a seasonal streamflow forecasting model for the Carnaubal Reservoir, located in the semi-arid region of Ceará, Brazil, using artificial neural networks (ANN), specifically the Multilayer Perceptron (MLP) architecture. Two models were developed using different sets of sea surface temperature (SST) indices representing oceanic conditions in the tropical Atlantic and equatorial Pacific, including Niño 3, Niño 1+2, TAN, and TAS, for the period from 1912 to 2015. The forecasts were evaluated using deterministic, probabilistic, and classification-based performance metrics. The model based on Atlantic predictors showed superior performance, with a Nash–Sutcliffe efficiency (NSE) of 0.58 and a correlation coefficient (R) of 0.78 in the test dataset, and it also outperformed climatology (Performance index of 1.16). In classifying dry, normal, and wet years, the model achieved 57% accuracy and performed better at predicting wet years (F1-score of 0.74). The results highlight the relevance of tropical Atlantic variability for streamflow predictability in the region and demonstrate the potential of neural networks as decision-support tools for water resources management in semi-arid environments. Future studies may explore multimodel approaches to further improve forecasting accuracy.