Abstract: Background, aim, and scope As known, both PM2.5 and PM10 (denoted as PM2.5/10) are typical fine particles matters, which have serious threats to air quality and human health. To better understand the temporal variations of air quality, this research is dedicated to the prediction of PM2.5/10 concentrations. Materials and methods In this research, the RNN (Recurrent Neural Network) model based on the NARX (Nonlinear Autoregressive with External Input) method has been proposed by using the hourly monitoring data of pollutants (incl. PM2.5, PM10, NO2、NO, NOX and CO) and meteorology (incl. wind direction, wind speed, temperature, humidity, pressure, etc.) at Liangsidu Monitoring Station located in Xianyang, Shaanxi, China. Result Six optimal structures of the neurons in hidden layer have been determined to predict the PM2.5 and PM10 concentrations in the following 6 hours, 12 hours and 24 hours, respectively. The conducted experiments shown that (a) for the PM2.5 prediction of 6 hours, the performance becomes the best when the neuron number of hidden layers has been settled as 8, while for the PM2.5 prediction of 12 hours and 24 hours, the neuron number of hidden layers should be turned to 12 and 7 for the best prediction accuracy; and (b) for the prediction PM10 prediction of 6, 12 and 24 hours, the optimal settlements of the neurons number in hidden layers are 12, 10 and 13 respectively. Discussion In general, the proposed model has revealed satisfactory performance for the predictions of PM2.5/10 concentrations and the prediction accuracy for the next 6 hours is slightly better than that for 12 and 24 hours. Some uncertain predictions, however, still exist especially when unusual meteorological situation occurs. In addition, the data used for the neural-network training are not quite enough. Conclusion It has been demonstrated that the established RNN model based on the NARX network can be implemented to effectively predict the concentration of PM2.5/10: the R values for the PM2.5 predictions of the following 6, 12 and 24 hours reaches 0.929281, 0.906767 and 0.889691, respectively; the corresponding RMSE values are 0.0008, 0.0010 and 0.0012; for the prediction of PM10, the R reaches 0.929867, 0.921972 and 0.917757, and the corresponding RMSE are 0.0013, 0.0014 and 0.0017 for the following 6, 12 and 24 hours respectively. Recommendations and perspectives In order to further improve the prediction performance of the PM2.5/10 concentrations, the effect of unusual methodological conditions should be considered by the proposed RNN model; moreover, the sensitivity analysis of the different input parameters need to be further investigated.
Keywords: PM10; PM2.5; air quality; NARX; recurrent neural network; air pollution prediction