PM_2.5 Prediction with a Novel Multi-Step-Ahead Forecasting Model Based on Dynamic Wind Field Distance

Particulate matter with a diameter of less than 2.5 μ m (PM_2.5) has damaged public health globally for a decade. Accurate forecasts of PM_2.5 concentration can provide early warnings to prevent the public from hazard exposure. However, existing methods have not considered the available spatiotemporal data sufficiently due to their architecture or inadequate input, and most neglected wind impact on spatiotemporal correlation when selecting related sites. To fill this gap, we proposed a long short-term memory-convolutional neural network based on dynamic wind field distance (LSTM-CNN-DWFD) to predict PM_2.5 concentration of a specific site for the next 24 h. A KNN method based on dynamic wind field distance was developed and applied to select highly related sites considering wind impact. A local stateful LSTM model was employed to capture temporal correlations in historical air quality and meteorological data for each related site. Then, these temporal features were integrated as a spatiotemporal matrix, and input into CNN for extracting spatiotemporal correlation features. Weather forecasts were also integrated into the model to promote accuracy. Hourly PM_2.5 data from 36 monitoring sites in Beijing, China collected from 1 May 2014 to 30 April 2015 were used as experimental dataset. Six-fold rolling origin method was employed to conduct experiments on each site, and the results of 216 experiments validated the performance of the proposed LSTM-CNN-DWFD model. The mean R 2 values of the next 1-6 h prediction were 0.85, 0.81, 0.76, 0.70, 0.64, and 0.59, respectively, showing a decrease trend, indicating that the prediction accuracy decreases as the prediction time increases. Comparisons of LSTM-CNN-DWFD results to results from six other methods show that it delivered higher accuracy PM_2.5 predictions, with the mean RMSE (MAE) of 1-6, 7-12, and 13-24 h prediction were 43.90 (29.17), 57.89 (42.16), and 63.14 (47.64), respectively. The results also demonstrate that the sites selected based on dynamic wind field distance are more related to the central site than that based on geographical distance, also contributing to prediction accuracy.