Predicting rainfall using neural nets

Abstract

One of the most crucial factors that can help in making strategic decisions and planning in countries that rely on agriculture in some ways is successfully predicting rainfall. Despite its clear importance, forecasting rainfall up until now remains a big challenge owing to the highly dynamic nature of the climate process and its associated seemingly random fluctuations. A wide variety of models have been proposed in literature to predict rainfall among which statistical models have been one of the most relatively successful. In this paper, we propose a novel rainfall forecasting model using focused time-delay neural networks (FTDNNs). In addition, we also contribute in comparing rainfall forecasting performances, using FTDNNs, for different prediction time scales, namely: monthly, quarterly, bi-annually and yearly. We present the optimal neural network architecture parameters automatically found for each of the aforementioned time scales. Our models are trained to perform one-step-ahead predictions and we demonstrate and evaluate our results, measured by mean absolute percent error, on the rainfall dataset obtained from Malaysian Meteorological Department (MMD) for close to a thirty year period. For test data, we found that the most accurate result was obtained by our method on the yearly rainfall dataset (94.25%). For future work, dynamic meteorological parameters such as sunshine data, air pressure, cloudiness, relative humidity and wet bulb temperature can be integrated as additional features into the model for even higher prediction performance.