Non-Linear Spatio-Temporal Input Selection for Rainfall Forecasting Using Recurrent Neural Networks

Abstract

Rainfall is an important component of the hydrologic cycle and is used for planning in various fields. Based on the White test it is known that some weather variables correlate non-linearly to rainfall. Meanwhile, from correlation testing it is known that the observed weather data from weather stations in a region are mutually correlated. Therefore, statistical modeling using autocorrelation and cross correlation is less appropriate because the assumption of linear correlation is not fulfilled. In this paper, a new framework is proposed for non-linear feature extraction using detrended partial crosscorrelation analysis and predictor input selection using symmetrical uncertainty as a way to determine optimal nonlinear input features in rainfall forecasting. Forecasting was performed simultaneously for 3 weather station locations in addition to taking into account the dependencies of observation time. This is called a non-linear spatio-temporal recurrent neural network. The result of the forecasting method shows that the model performed better than univariate/multivariate time series forecasting and a recurrent neural network without input selection.