Long Short-Term Memory Deep Neural Network Model for PM2.5 Forecasting in the Bangkok Urban Area

Abstract

Accurately forecasting fine particulate matter of less than a 2.5 micrometer diameter (PM2.5) concentration levels is important to better manage the air pollution situation and to give advance warnings to residents and officials. In this paper, a Long Short-Term Memory (LSTM) deep neural network model and a Seasonal AutoRegressive Integrated Moving Average with eXogenous regressor (SARIMAX) were trained on air quality and meteorological time series data at the Chokchai metropolitan police station area in Bangkok from 2017 to 2018. After figuring out the best configuration of both algorithms, the performance of the LSTM model to predict PM2.5 concentrations for 24 hours was evaluated and compared against the SARIMAX model. Our experiments indicated that LSTM had a better prediction accuracy as indicated by the RMSE and MAE values for each of the time steps. LSTM could forecast one hour ahead at a very low RMSE of 3.11 micrograms per cubic meter on average, and a MAE of 2.36 micrograms per cubic meter on average, while SARIMAX errors were more than doubled. When the time steps were farther apart, the number of errors were higher for both models.