Abstract
PM2.5 concentrations prediction faces the challenge of insufficient ground stations and a lack of monitoring capability for future expectations. In this study, we performed time-series prediction of daily PM2.5 concentrations at the monitoring sites in the Pearl River Delta (PRD) region, China, based on Bidirectional Long Short-term Memory (BiLSTM) network, combining the 1-km resolution satellite aerosol optical depth (AOD) data as well as meteorological parameters, geographic features, socio-economic developments, and other influencing factors. The model results show good performance in model predictions (R2 of 0.58-0.78) and scrolling prediction (R2 of 0.42–0.83) at each PM2.5 monitoring site. Model performance varies by year and season, where the model performs best in 2015 (R2 of 0.75), better in summer and fall (R2 of 0.59–0.81), and worse in spring and winter (R2 of 0.49–0.72). The PM2.5 concentrations of January 2019 were predicted with evaluation indicators (R2 of 0.62, RMSE of 12.887), which was consistent with the expected performance of our model in the PRD region during winter. Accurate spatiotemporal predictions of future pollution distributions can provide decision support for PM2.5 management countermeasures. Further, more monitoring stations and the latest data are still urgently needed.
Published Version
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call