Development of the GSTARIMA(1,1,1) model order for climate data forecasting

Abstract

The space-time model combines spatial and temporal elements. One example is the Generalized Space-Time Autoregressive (GSTAR) Model, which improves the Space-Time Autoregressive (STAR) model. The GSTAR model assumes that each location has heterogeneity characteristics, and that the data is stationary. In this research, the moving average component is calculated by involving the relationship between variable values at a certain time and residual values at a previous time, and it is assumed that the data is not stationary, so the model used is the Generalized Space-Time Autoregressive Integrated Moving Average (GSTARIMA) Model. The model order for GSTARIMA is determined through the Space-Time Autocorrelation Function (STACF) and Space-Time Partial Autocorrelation Function (STPACF) to ensure accurate forecasting. Previous research only discussed the GSTARIMA(1,1,1) model, so in this research, the GSTARIMA(3,1,1) model will be addressed as a form of development of the GSTARIMA(1,1,1) model and applied to climate data. The climate data used in this research is sourced from NASA POWER and consists of rainfall variables with large data sizes, requiring the use of the data analytics lifecycle method to analyse Big Data. The lifecycle includes six phases: discovery, data preparation, model planning, model building, communicating results, and operationalization. Based on the data processing results with Python software, the GSTARIMA(3,1,1) model has a MAPE value of 9% for out-sample data and 11% for in-sample data. In contrast, the GSTARIMA(1,1,1) model has a MAPE value of 11% for out-sample data and 12% for in-sample data. So the GSTARIMA(3,1,1) model provides more accurate forecasting results. Therefore, selecting the correct model order is crucial for accurate forecasting.