Investigation of nearby monitoring station for hourly PM2.5 forecasting using parallel multi-input 1D-CNN-biLSTM

Abstract

Listen

Translate article

Air quality forecasting is a hot research topic that has been widely explored by the whole society. To better understand environmental quality, numerous methods have been proposed for investigating air pollutant data. Previous studies have used deep learning-based methods for hourly PM2.5 forecasting due to its success in various pattern recognition tasks. However, only using data from the target station for PM2.5 forecasting will limit its performance. In this study, we develop an automated hourly PM2.5 forecasting model using a parallel multi-input 1D-CNN-biLSTM model by combing data from both target and nearby monitoring stations. Specifically, six single-variable biLSTM-based models are first evaluated to select the best one for the subsequent analysis. Then, a parallel multi-input 1D-CNN-biLSTM architecture is constructed, where a nearby station is selected and combined with the target station for PM2.5 forecasting. Moreover, we forecast PM2.5 with an emphasis on seasonal variability. Experimental results show that combining the nearby station with the target station can improve the forecasting performance. The best average RMSE, MAE and R2 over nine sites using our proposed system are 3.88 and 2.52, and 0.94, respectively. Furthermore, the performance for Autumn is the best in terms of RMSE and MAE with emphasis on seasonality. Therefore, the proposed parallel multi-input 1D-CNN-biLSTM model can be adopted for PM2.5 forecasting.