A Novel Hybrid GSTARX-RNN Model for Forecasting Space-Time Data with Calendar Variation Effect

Abstract

Recent development in space-time data forecasting includes a hybrid model. In this study, we propose a hybrid spatio-temporal model by combining Generalized Space-Time Autoregressive with exogenous variable and Recurrent Neural Network (GSTARX-RNN) for space-time data forecasting with calendar variation effect. The GSTARX model as a linear model is used to modeling space-time data with exogenous variables while the RNN model as nonlinear model is used to modeling the nonlinear patterns of the data. In particular, we employ two variants of RNNs, i.e. Elman RNN and Jordan RNN. We apply our methods on the simulation study. The results show that the proposed methods yielded more accurate forecast especially in the simulated data containing nonlinear patterns. Moreover, the GSTARX-Elman RNN as a more complex model tends to give more accurate forecast than the GSTARX-Jordan RNN.