A Hybrid Generalized Space-Time Autoregressive-Elman Recurrent Neural Network Model for Forecasting Space-Time Data with Exogenous Variables

Abstract

This research proposes a hybrid method by combining Generalized Space-Time Autoregressive with exogenous variables and Elman Recurrent Neural Network (GSTARX-Elman RNN) to forecast space-time data. GSTAR method is used for modeling and forecasting multivariate data which including time and location factors. The modeling GSTAR with exogenous variables is to capture time series factors, i.e., trend, seasonal, and calendar variation. This method combines with Elman RNN as a nonlinear forecasting method for the data that have a nonlinear pattern. Hybrid GSTARX-Elman RNN compares with time series regression and GSTARX methods based on RMSE criteria. This research focused on simulation data that consist of a trend, seasonal, and calendar variation patterns, and using two scenarios of noise, i.e., linear and nonlinear noise. The result of these simulations showed that time series regression and GSTARX method could capture well the exogenous variables, but hybrid GSTARX-Elman RNN is a more accurate method than others. Hybrid GSTARX-Elman RNN could capture nonlinearity data pattern from these simulations. In general, the hybrid models tend to provide more accurate forecast performance than individual forecast models that it is in line with the results of the M4 forecasting competition.