Rainfall prediction using backpropagation algorithm optimized by Broyden-Fletcher-Goldfarb-Shanno algorithm

Abstract

An extreme climate change results in a long dry season and an extreme rainfall results the losses in various areas of life. Rainfall prediction becomes an important thing for planning in many life sectors. Many prediction methods have been proposed, such as Autoregressive Integrated Moving Average (ARIMA) and Artificial Neural Networks (ANN). ANN has some advantages compared with the ARIMA model. Backpropagation algorithm is one of the ANN which has been successfully used in various fields. However, the performance of the backpropagation algorithm depends on the architecture and the optimization method used. The standard backpropagation algorithm optimized by gradient descent method works slowly to get a small error. The Broyden-Fletcher-Goldfarb-Shanno (BFGS) algorithm works faster than gradient descent method. For this reason, this paper proposes the rainfall prediction using the backpropagation algorithm optimized by the BFGS algorithm. From the experiment results, it can be shown that the backpropagation algorithm optimized by the BFGS algorithm gives better result compared with the standard backpropagation algorithm for rainfall prediction. The big number of neuron hidden causes overfitting and the small number of neuron hidden make the worst accuracy. Choosing the right learning rate will produce better accuracy.