Impact of climate dataset variability on spatiotemporal modelling of PM2.5 concentration using deep neural network models

Abstract

Meteorological parameters have been used as features in spatiotemporal PM_2.5 models owing to their influence on the formation, concentration, and dispersion of PM_2.5 particles. This study investigated the effectiveness of various meteorological datasets, resampled using different methods, in predicting PM_2.5 concentrations with optimized deep neural network (DNN) models. The results showed that a cubic-spline-resampled European Centre for Medium-Range Weather Forecasts (ECMWF) dataset, as a meteorological feature in the DNN models, produced the best results in predicting PM_2.5 values with MAE and RMSE values of 3.597 and 5.363 μg/m³, respectively. Day of the year, the aerosol optical depth, and temperature were the features with the most significant influence in predicting PM_2.5 concentrations based on their SHAP values for each DNN models. The distribution and correlations between resampled values of each resampling method and dataset combination were discussed. This study also discusses future exploration for improving the spatiotemporal PM_2.5-concentration modelling process.