Aiming at the shortcomings of the traditional prediction model for PM2.5 concentration prediction with low accuracy and poor interpretability,a spatio-temporal convolutional network prediction model integrating 2D convolution and attention(S-2D-TCNA)layers is proposed.Hourly air quality and meteorological data from 36 monitoring stations in Beijing from May 1,2014,to April 30,2015,are selected.Through the analysis of spatio-temporal correlations among multiple stations,the monitoring station data that meet the correlation threshold are inputted into a convolutional network that adopts a dimensionality expansion and reduction approach to obtain input features with spatio-temporal sequences.Attention is incorporated into the Temporal convolutional network model for predicting the PM2.5 concentrations at the central monitoring station for the next hour.In the process of optimizing the parameters for model training,the parameters of the deep learning model are trained by Adam,and then Bayesian optimization is used to adjust the hyper-parameters of the model,and this method finds the best parameters of the model,which reduces the root mean squared error and mean absolute error by 3.791%and 5.576%,respectively,and increases the goodness of fit by 0.67%.In terms of quality,the S-Conv2D-TCNA model has a mean root mean square error,mean absolute error and goodness of fit are 16.020 9,10.610 0 and 0.942 8,respectively,and this prediction model is better than the baseline model in terms of accuracy and stability.The results show that the forecasting system is promising in the early warning,regional prevention,and control of air pollution.
NETL
NSTL
万方数据
