Hybrid ANN and physically-based models for regional PM2.5 forecasts

Air pollution has affected people's health and lowered our living quality. Among all pollutants, PM_2.5, which is smaller than 2.5 microns, can easily penetrate human lungs and seriously affect human health. Therefore, PM_2.5 control is a very crucial action. Air pollution modelling can roughly categorize into two types, stochastic model (Artificial neural network (ANN) model) and deterministic model (physically-based model). Since the variation of PM_2.5 concentrations is dynamic, the physically-based model struggles to handle the uncertainty from its complex interaction. With the aid of the nonlinearity of ANNs, we can overcome these uncertainties. We proposed a hybrid convolutional (CNN)-based ANN to extract features from the dataset to provide three days ahead PM_2.5 forecast. The physically-based model first generates the simulated dataset. Over 40 thousand historical and simulated hourly datasets are collected to construct the deep learning model. This hybrid model that learns historical information and future trends performs better in terms of R² (0.58-0.72) than the baseline model (0.40-0.44). Besides that, its forecast time horizon is relatively long (<72 hours) if we compare it with the pure ANN model (<12 hours). As a result, the proposed hybrid model can provide accurate regional air pollution forecasts by inheriting the characteristics of physically-based model and ANN.

Keywords: Artificial Neural Network; Deep learning; Convolutional neural network (CNN); Regional air quality forecast