Efficient Deep Learning-driven Approach for PM2.5 Forecasting at Different Locations in Spain

Abstract

Forecasting dust pollution is necessary for achieving satisfactory air quality. This work proposes an improved deep learning-based forecasting approach for PM2.5 concentration forecasting. Importantly, this approach introduces an improved variational autoencoder (VAE) model by incorporating a bidirectional gated recurrent unit (BiGRU) at the encoder side of the VAE model. The forecasting quality of the coupled model is verified via comparisons with the traditional VAE model when forecasting PM2.5 concentration time-series data. The assessment is carried out using five statistical metrics. PM2.5 datasets from different stations in Spain are used in this study. Results reveal the accuracy of the improved VAE model for PM2.5 concentration forecasting over the traditional VAE, LSTM, GRU, biLSTM, and BiGRU.