Deep learning-based mesoscopic emission modeling

This study introduces DEP (Direct Emission Prediction), a deep learning-based model for estimating traffic emissions with enhanced accuracy and efficiency. Traditional models present limitations: macroscopic models are computationally efficient but lack accuracy, whereas microscopic models offer high accuracy but are computationally intensive for large-scale applications.DEP adopts a mesoscopic approach using a Gated Recurrent Unit (GRU) architecture. It processes sequential link-level trip data, vehicle characteristics, road attributes, and temperature to predict CO2 and NOx emissions. When integrated with MATSim, DEP enables high-resolution emission estimation over large areas.DEP achieves accuracy comparable to SPG-M, a hybrid mesoscopic–microscopic model, while significantly reducing computational time. Validation against on-road remote sensing emission measurements shows that DEP predictions align closely with real-world NOx emissions. In contrast, traditional macroscopic models such as COPERT and HBEFA underestimate NOx emissions. The model effectively captures cold-start emissions and their spatial distribution, which are particularly relevant in urban contexts.