PM2.5 air pollution prediction through deep learning using meteorological, vehicular, and emission data: A case study of New Delhi, India

Abstract

Particulate matter (PM2.5) concentration is an air pollutant that can lead to serious health complications in humans. The detection of this air pollutant is essential so that government agencies can formulate policies to take effective measures. This study proposes and analyzes a Gated Recurrent Unit Based Encoder-Decoder (GRU-ED) method for predicting 1-hourly, 8-hourly, and 24-hourly PM2.5 concentrations in New Delhi, India, for three years (from 2008 to 2010). The study uses different input parameter combinations of meteorological (M), vehicle (V) population, and emissions (E) data. In all, the authors tested the proposed GRU-ED method with four models: Model 1: Vehicle population + Emission [no meteorological (VE)], Model 2: Meteorological + Emission [no vehicle population (ME)], Model 3: Meteorological + Vehicle population [no emission (MV)], and Model 4: Meteorological + Vehicle population + Emission (MVE). It is observed that the proposed GRU-ED method performed better than traditional machine learning predictive methods (Random Forest, Extreme Gradient Boosting, Artificial Neural Networks, and Long Short-Term Memory (LSTM)) in terms of forecast value accuracy. The GRU-ED method with Model 4 is found to be the most accurate forecasting model for 1-hourly PM2.5 concentration prediction (R2 = 0.959, NSE = 0.953, MAE = 1.770, RRMSE = 0.002, and MAPE = 0.190). It is also observed that among the meteorological, vehicle, and emission parameters, the presence of the meteorological parameter has a significant impact on the prediction accuracy. © 2023