A Hybrid Approach to Model Nonstationary Space-time Series

In recent years the Space-Time Autoregressive Moving-Average (STARMA) model family has been proven a useful tool in modelling multiple time series data that correspond to different spatial locations (which are called space-time series). The STARMA model family is a statistical inductive model that can be used to describe stationary (or weak stationary) space-time processes. However, in real applications STARMA model can not be applied directly because of the non-stationary nature of most space-time processes. To overcome this deficiency, a novel approach to model non-stationary space-time series is proposed in this study. It uses artificial neural network (ANN) to develop a non-parametric, robust model to extract the large-scale nonlinear space-time structures, then uses STARMA model to extract the small-scale stochastic space-time variations. The proposed approach has been applied to the forecasting of china annual average temperature at 137 international meteorological stations in China. The experimental results demonstrate that the forecasting using ANN+STARMA method obtains better forecasting accuracy than using conventional pure STARMA method. It proves the mixture of the data-driven ANN and the model-driven STARMA can become a very useful and efficient tool for space-time modelling and prediction of environment data with temporal and spatial dependence. * Corresponding author.