LAUSR

It is important to assess environmental impact of intelligent transportation systems, and hence developing a vehicle emission model with high accuracy has been a long-standing topic in transportation research. However, current vehicle emission models are either overly simple using average speed, resulting in low estimation accuracy, or they are too complicated requiring excessive inputs, relying on too much prior knowledge. In this study, we develop and evaluate a deep learning-based vehicle emission model (DL-VEM) to estimate the instantaneous CO2 emissions of taxicabs. First, we examine the correlation between observed emissions and vehicle driving condition data collected in a PEMS experiment. Then, an end-to-end deep learning structure is developed to model patterns of vehicle emissions. Specifically, LSTM networks are used to learn temporal dependencies of historical driving patterns, and fully connected networks are employed to extract deep features of current driving behaviors and external environment. Our model aggregates the outputs of these networks using different learnable weights. Experiments were conducted in Wuhan, China, where our model was trained and validated using observed datasets. Compared with the state-of-the-art models, our model achieved higher accuracy in estimating CO2 emissions. Thereafter, it was applied to a taxicab trajectory dataset in one day, and spatiotemporal patterns of CO2 emissions were presented using different fuel types. Importantly, we find that an increment of 24.94% emissions can be expected if petrol instead of compressed natural gas was used by each taxicab in Wuhan.