CO concentration forecasting with attention-based LSTM and EMD for urban areas in Selangor

Abstract

Deterioration of air quality levels due to the high concentration of air pollutants in ambient air, especially within urban areas, has severely affected human health. Constant exposure to a dangerous air pollutant of carbon monoxide (CO) may lead to serious health problems such as heart diseases, lung damage and respiratory system failure, which may increase mortality risk. Therefore, this study aims to forecast CO concentration in urban areas in Malaysia using a hybrid deep learning model. Empirical mode decomposition (EMD) is used to decompose CO concentration data into multiple components, namely intrinsic mode functions (IMFs) and a residual. Attention-based long short-term memory (ALSTM), the combination of multiple LSTM layers and an attention layer, is used to forecast the decomposed components individually. Then, the forecasted sub-sequences are accumulated to obtain the final forecasting of CO concentration. In this study, forecasting CO concentration is based on hourly historical time series data considering the effect of meteorological parameters. EMD-ALSTM outperforms individual LSTM and ALSTM models in terms of statistical evaluation analysis. The results indicate that the hybrid forecasting model has successfully forecasted CO concentration with reliable accuracy.