Exploring a deep LSTM neural network to forecast daily PM2.5 concentration using meteorological parameters in Kathmandu Valley, Nepal

Abstract

Fine particulate matter (PM2.5) is a complex air pollutant with numerous gaseous and solid constituents. PM2.5 possesses a significant hazard due to its ability to penetrate deep into the lungs, corrode the alveolar wall, and impair lung functions. Modeling the non-linear and dynamic time series of daily PM2.5 concentration remains a challenge. This study proposes a deep LSTM neural network to forecast accurate PM2.5 concentration in the Kathmandu valley. Correlation analysis illustrates that dew, minimum ambient temperature, maximum ambient temperature, and pressure are strongly correlated with PM2.5 concentration. Hence, five models are developed based on different input parameter combinations and are eventually evaluated to determine the best performing model. Model 2 with single-step prediction is the best performing deep LSTM model with RMSE of 13.04 μg/m3 and MAE of 10.81 μg/m3. The SARIMA model applied to the univariate PM2.5 data series illustrates the RMSE of 19.54 μg/m3 and MAE of 15.21 μg/m3 for the test data. Hence, the deep LSTM model with past PM2.5 data and dew as inputs is recommended to predict future PM2.5 concentration in the Kathmandu valley. The negative impact of PM2.5 concentration on public health can be minimized with efficient forecasting.