Previsão climática sazonal para a Região Hidrográfica dos Sertões de Crateús utilizando redes neurais artificiais recorrentes

Abstract

One of the great problems that the Sertão of the Brazilian Northeast faces is water scarcity. This is due to the low rainfall index that this region has and the concentration of rainfall in a few months of the year. Thus, a short-term planning of the use of water resources is necessary, highlighting the importance of climate accuracy for decision making. Thus, this work aims to generate precipitation forecast models for the Hydrographic Region of the Sertões de Crateús, using a Recurrent Artificial Neural Network called Long Short Term Memory. For this, the rainy quarter (FMAM) was analyzed, using as predictor four quarters, SON, DJF, MAM and JJA. Each predictor has 11 TSM indices ranked in order of relevance and 11 forecast models were generated for each predictor using the Python language. The predictor that presented the highest accuracy, i.e., with higher correlation and NASH values was DJF, reaching 0.808 in correlation and 0.530 in NASH, both in the validation phase. Nevertheless, in comparison with similar work also using neural networks, it was observed that the NASH and correlation values in the validation phase of this work were better than the comparison phase, in other words, the network developed in this work has a greater power of generalization.