Previsão climática sazonal para a região hidrográfica dos sertões de Crateús utilizando redes neurais artificiais

Abstract

The hydrographic region of the Backlands of Crateús has been in the last years (2011 to 2017) for a long period of drought, causing an emptying of the municipal supply system springs. Emergency plans such as water contingency and the transposition of waters from other watersheds were put in place in order to avoid the shortage of cities in the Backlands of Crateús hydrographic region. Thus, there is a need for short-term planning of the use of naturally limited water resources, highlighting the importance of climate accuracy, as it helps in decision making. The present work has as main objective to evaluate the predictability of the rainfall of the Backlands of Crateús hydrographic region from the Sea Surface Temperature (SST) variability using artificial neural networks techniques. For this, four predictive periods are analyzed: December to May (DJFMAM), February to May (FMAM), December to February (DJF) and March to May (MAM), using 12 SST indices with the quarters immediately preceding each one. predictive period, that is, analyzing the predictive period of DJFMAM, the predictor would be the average of the TSM indices of the quarters from September to November (SON), June to August (JJA), MAM and DJF, respectively. Where initially obtained through the Random Forest (RF) are the most important indices and then the seasonal climate prediction is performed using Artificial Neural Networks (ANN) with different amounts of input variables, ranging from 1 to 12 indices (which are organized in order of importance). The Tropical South Atlantic Index was classified more often as the most important variable, followed by the Tropical Atlantic Dipole. The models with the highest accuracy, that is, those that obtained the best correlation and NASH results were the predictive period of MAM where the DJF period was used as a predictor, having correlation and NASH values of 0.883 and 0.663, respectively. Finally, the ANN model obtained satisfactory correlation results, but the NASH values were low.