Analysis of heterogeneous traffic flow characteristics based on LSTM

To better understand the characteristics of heterogeneous traffic flows involving Connected and Autonomous Vehicles (CAVs) and Human-driven Vehicles (HVs), an accurate description of driver behaviour within the model is essential. Since vehicle acceleration decisions are influenced not only by the vehicle’s state at a specific moment, as in traditional Cellular Automaton (CA) models, but also by historical driving data over a period, the Long Short-Term Memory (LSTM) model was introduced. This model is particularly effective at managing long-term sequences, complex nonlinear features, and multivariate dynamic influence data. By training the LSTM model on vehicle driving data from the NGSIM database, it uncovers the underlying driving habits of human drivers and predicts the next acceleration decision. The trained model was then combined with cellular automata to create the Long Short-Term Memory-Cellular Automaton (LSTM-C) model for simulating traffic flow. Comparisons with other models show that the model can replicate the phase transition process in three-phase traffic flow while partly reducing issues caused by sudden vehicle acceleration and abrupt braking. Simulations were performed for heterogeneous traffic flows at different CAV penetration rates, and the results were analysed. The analysis reveals that as CAV penetration increases, the maximum traffic flow also rises. For example, at an 80% penetration rate, the maximum flow increases by 42.5%. Further analysis indicates that higher CAV penetration provides benefits such as faster vehicle speeds, fewer lane changes, and reductions in congestion coefficient and congestion duration.