Hybrid Model for Prediction of Carbon Monoxide and Fine Particulate Matter Concentrations near a Road Intersection

Air quality time series near road intersections consist of complex linear and nonlinear patterns and are difficult to forecast. The backpropagation neural network (BPNN) has been applied for air quality forecasting in urban areas, but it has limited accuracy because of the inability to predict extreme events. This study proposed a novel hybrid model called GAWNN that combines a genetic algorithm and a wavelet neural network to improve forecast accuracy. The proposed model was examined through predicting the carbon monoxide (CO) and fine particulate matter (PM2.5) concentrations near a road intersection. Before the predictions, principal component analysis was adopted to generate principal components as input variables to reduce data complexity and collinearity. Then the GAWNN model and the BPNN model were implemented. The comparative results indicated that GAWNN provided more reliable and accurate predictions of CO and PM2.5 concentrations. The results also showed that GAWNN performed better than BPNN did in the capability of forecasting extreme concentrations. Furthermore, the spatial transferability of the GAWNN model was reasonably good despite a degenerated performance caused by the unavoidable difference between the training and test sites. These findings demonstrate the potential of the application of the proposed model to forecast the fine-scale trend of air pollution in the vicinity of a road intersection.