Neural Networks for Time Series Rainfall Forecasting: A Case Study in Manaus, Amazonas.


In this paper we aim at forecasting rainfall occurrence using time series and neural networks. In our approach, three meteorological variables (average high and low temperatures and relative humidity) are taken as input by neural networks with a single hidden layer in order to deliver one-step ahead rainfall occurrence predictions. We performed a case study to evaluate such approach considering 40 years of data from an automated weather station located in Manaus, Amazonas. From 38 neural networks suited to this scenario, we could identify, using Akaike’s Information Criterion, that the one with window size equal to 3 and 7 neurons in the hidden layer could forecast rainfall with an accuracy of 99.71% for a testing set. The results obtained indicate the feasibility and efficiency of time series neural networks for rainfall forecasting, suggesting a methodology that can be adopted by many other locations.

In: Encontro Nacional de Inteligência Artificial e Computacional, 2016, Recife, Pernambuco.