Hybrid machine learning for forecasting and monitoring air pollution in Surabaya

Abstract

This research aims to propose hybrid machine learnings for forecasting and monitoring air pollution in Surabaya. In particular, we introduce two hybrid machine learnings, i.e. hybrid Time Series Regression – Feedforward Neural Network (TSR-FFNN) and hybrid Time Series Regression – Long Short-Term Memory (TSR-LSTM). TSR is used to capture linear patterns from data, whereas FFNN or LSTM is used to capture non-linear patterns. Fifteen half-hourly series data, i.e. CO, NO2, O3, PM10, and SO2 in three SUF stations at Surabaya, are used as the case study. We compare the forecasting accuracy of these hybrid machine learnings with several individual methods (i.e. TSR, ARIMA, FFNN, and LSTM), and combined methods (i.e. TSR with AR error and TSR with ARMA error). The identification step showed that these air pollution data have double seasonal patterns, i.e. daily and weekly seasonality. The comparison results showed that no superior method that yields the most accurate forecast for all series data. Moreover, the results showed that hybrid methods gave more accurate forecast at 8 series data, whereas the individual methods yielded better results at 7 series data. It supported that methods that are more complex do not always produce better forecasts than simple methods, as shown by the first result of the M3 competition. Additionally, the results of the forecast of air pollution index for monitoring air pollution in Surabaya show that the air quality is in good and moderate air pollution levels for duration of 19.30 to 03.00 and 0.30 to 19.30, respectively.