A Spatiotemporal Prediction Network for Vehicle Emissions on Large Road Network

Prediction of vehicle emissions on urban road networks can serve for intelligent vehicles to avoid high-emission driving modes and routes through high-emission roads. However, existing methods focus on the spatiotemporal features of emissions based on graph representation learning. Little attention has been given to the challenge of computational complexity when constructing deep networks for global feature aggregation in cases where the number of road nodes is large. In light of these limitations, a spatiotemporal transformer network for emission prediction is proposed. Specifically, the proposed network utilizes the spatiotemporal self-attention mechanism to aggregate embedded features, which is achieved through the dynamic attention weight to select crucial features. Furthermore, a graph reconstruction module is introduced to transform the original road network into a second-order connected graph, which ensures global feature propagation while reducing the complexity of secondary calculations for self-attention. The experimental results demonstrate that the proposed network achieves better prediction accuracy than existing methods when tested on the Xi’an vehicle emission dataset.