Dynamic Origin-Destination Demand Prediction with Improved LSTM Model

As the basis of dynamic traffic distribution, traffic demand prediction needs to address the complicated coupling correlations within the road network structure and the spatial and seasonal variation of traffic demand. Dynamic origin-destination (OD) demand prediction can present the correlation between nodes in the road network pattern and the time-varying and periodicity of traffic data. In this paper, we propose a dynamic OD prediction model considering the spatial-temporal and environmental aspects (STE) of the network (STE-CNN-LSTM) based on the Long-Short-Term-Memory (LSTM) algorithm. The proposed model consists of three modules including the spatial, temporal and external environment modules. The spatial module extracts the spatial correlation information of OD data. The temporal module extracts the time correlation information of OD data, and the external environment module deals with the impact of the external factors (e.g., weather, accidents) on OD demand. To verify the effectiveness of the proposed model, we conducted experiments using the OD dataset collected by the Caltrans Performance Measurement System (PeMS) of California in the United States, and the results show that the proposed STE-CNN-LSTM model performs better than the ARIMA, general LSTM and CNN-LSTM models in most respect in most respects.