Comparison ARIMA and GSTAR model to forecasting air pollution in Surabaya

The Air Quality Index (AQI) measures the concentration of air pollutants in ambient air pollution based on air pollutants, which are then divided into categories and number ranges with health risks. Given the time dimension and space dimension, this study proposes a Generalized Space Time Autoregressive (GSTAR) model to capture the space dependence and time pattern in the data and an ARIMA model built for each measurement location. This study uses two weights, normalized cross-correlation and inverse weight to distance for the GSTAR model. The data used consists of four measurement locations of air pollution concentrations in Surabaya City. Model goodness criteria are measured using RMSE and MAPE values because the variables used have various values. The results show that the ARIMA model for training data location Benowo, Wiyung, Kertajaya and Keputih obtained MAPE values in order of 15.48%, 17.14%, 11.20%, and 15.57%, while the testing data is 18.53%, 25.25%, 18%, and 17.67%. Based on the weights, it can be concluded that the weights with normalized cross-correlation are better because they have a smaller AIC value, even though they are not significantly different. The MAPE value using normalized cross-correlation weighting on the GSTAR (21) model is 15.67%, 18.66%, 11.97%, and 16.19%, respectively, while for testing data it is 22.32%, 33.32%, 25.81%, and 12.43%. Although the GSTAR model with normalized cross-correlation weights gives higher MAPE results on training data, it has an advantage in performance on testing data in some locations, suggesting that it is more robust for prediction on new data.