Spatiotemporal networks for multi-city and multi-task air pollutant prediction ——Beijing, Shanghai and Shenzhen as examples

Predicting air pollutant concentrations accurately is crucial for directing the development of policies and pollution mitigation measures. This paper proposes ST-RTNet, a unique deep learning model, to achieve high-precision prediction of air pollutant concentrations in multiple cities by integrating spatiotemporal attention mechanisms. Three components make up the model: reconstructed residual network with spatial attention module, temporal attention module based on multi-head self-attention mechanism, and temporal convolutional network (TCN). Convolutional neural network (CNN) and spatial attention module work together to create the reconstructed unit, which is the reconstructed residual network used to thoroughly extract the spatial characteristics of meteorological and air pollutant concentrations at monitoring stations. To concentrate on the impact of temporal features of past air pollution data on the forecast at each time step, the temporal attention module uses a multi-head self-attention mechanism. Subsequently, the TCN effectively captures the multi-step spatial and temporal dependencies, further extracts the temporal features, and finally completes the air pollutant prediction task through the fully connected layer. On real datasets from Beijing, Shanghai and Shenzhen, ST-RTNet exhibits higher accuracy and lower RMSE in both single-step and multi-step forecasting. The outcomes of the experiment demonstrate that ST-RTNet exhibits a larger absolute error but better trend prediction ability in the high concentration region, and this property is most significant on the Beijing dataset (PM_2.5 mean 34.285 μg/m³), which presents highest RMSE (8.237) and Corr (0.981). In addition, ST-RTNet demonstrated good robustness in the prediction of several cities and different air pollutant concentrations, verifying its wide applicability.