A Precipitation forecasting in central ecotone region in Brazil using artificial neural networks and public climatic data

Abstract

Precipitation forecasting may be of great value for farming, helping to reduce crop losses and irrigation costs, besides leveraging crop yield estimates. In Brazil, the state of Tocantins has a great part of its economy based on agriculture, where precipitation forecasting may help improve local production. Besides, rainfall forecasting can also contribute to urban planning through the management of water resources, among other applications Artificial Neural Networks (ANNs) have been used with relative success in precipitation forecasting for different locations and climates. With that, this work presents a method for weekly precipitation forecasting in six locations in the Brazilian state of Tocantins using ANNs and public climatic data. For that, MultiLayer Perceptron (MLP) networks were trained with data from local weather stations and El Niño Southern Oscillation (ENSO) related indices. First, input variables were selected using the forward selection algorithm. After that, ANN hyperparameters and input variables lag were optimized. The average Root Means Square Error (RMSE) of the final models was of 31.35 mm/week for the training dataset and 33.38 mm/week for the test dataset. Respectively, these values represent 9,83% and 10,46% of the maximum weekly precipitation found in the work dataset, which was of 319.1 mm. The results suggest that the created models are capable of reasonably good weekly precipitation forecasts, providing valuable information for farming, water resources management, urban planning and other related activies. Although there is possibly room for model improvement.