Data-driven predictive modeling of PM2.5 concentrations using machine learning and deep learning techniques: a case study of Delhi, India

The present study intends to use machine learning (ML) and deep learning (DL) models to forecast PM2.5 concentration at a location in Delhi. For this purpose, multi-layer feed-forward neural network (MLFFNN), support vector machine (SVM), random forest (RF) and long short-term memory networks (LSTM) have been applied. The air pollutants, e.g., CO, Ozone, PM10, NO, NO2, NOx, NH3, SO2, benzene, toluene, as well as meteorological parameters (temperature, wind speed, wind direction, rainfall, evaporation, humidity, pressure, etc.), have been used as inputs in the present study. Moreover, this is one of the first papers that employ aerodynamic roughness coefficient as an input parameter for the prediction of PM2.5 concentration. The result of the study shows that the LSTM model with index of agreement (IA) 0.986, root mean square error (RMSE) 21.510, Nash-Sutcliffe efficiency index (NSE) 0.945, (coefficient of determination)R-2 0.945, and (correlation coefficient)R 0.972 is the best performing technique for the prediction of PM2.5 followed by MLFFNN, SVM, and RF models. The sensitivity analysis for the LSTM model reported that PM10, wind speed, NH3, and benzene are the key influencing parameters for the estimation of PM2.5. The findings in this work suggest that the LSTM could advance in PM2.5 forecasting and thus would be useful for developing fine-scale, state-of-the-art air pollution forecasting models.