New Hybrid Statistical Method and Machine Learning for PM10 Prediction

Abstract

The objective of this research is to propose new hybrid model by combining Time Series Regression (TSR) as statistical method and Feedforward Neural Network (FFNN) or Long Short-Term Memory (LSTM) as machine learning for PM10 prediction at three SUF stations in Surabaya City, Indonesia. TSR as an individual linear model is used to capture trend and seasonal pattern. Whereas, FFNN or LSTM is employed to handle nonlinear pattern. Thus, this research proposes two hybrid models, i.e. hybrid TSR-FFNN and hybrid TSR-LSTM. Data about PM10 level that be observed half hourly at three SUF stations in Surabaya are used as case study. The performance of these two hybrid models will be compared with several individual models such as ARIMA, FFNN, and LSTM by using sMAPEP. The results at identification step showed that the data has double seasonal patterns, i.e. daily and weekly seasonality. Moreover, the forecast accuracy comparison showed that hybrid TSR-FFNN produced more accurate PM10 forecast than other methods at SUF 7, whereas FFNN yielded more accurate forecast at SUF 1 and SUF 7. These results show that FFNN as an individual nonlinear model produce better forecast than TSR and ARIMA as an individual linear model. It indicates that the PM10 in Surabaya tend to have nonlinear pattern. Moreover, these results are also in line with the results of M3 competition, i.e. more complex method do not necessary produce better forecast than a simpler one.