Deep spatial-temporal sequence modeling for multi-step passenger demand prediction

Abstract Supply–demand imbalance poses significant challenges to transportation systems such as taxis and shared vehicles (cars and bikes) and leads to excessive delays, income loss, and energy consumption. Accurate prediction of passenger demands is an essential step towards rescheduling resources to resolve the above challenges. However, existing work cannot fully capture and leverage the complex nonlinear spatial–temporal relationships within multi-modal data. They either include excessive data from weakly-correlated regions or oversight the correlations among those similar yet geographically distant regions. Moreover, these methods mainly focus on predicting the passenger demand for one future time step, whereas predictions over longer time scales are more valuable for developing efficient vehicle deployment strategies. We propose an end-to-end deep learning based framework to solve the above challenges. Our model comprises three parts: (1) a cascade graph convolutional recurrent neural network to extract spatial–temporal correlations within citywide historical vehicle demand data; (2) two multi-layer LSTM networks to represent the external meteorological data and time meta separately; (3) an encoder–decoder module to fuse the above two parts and decode the representation to achieve prediction over a longer time period into the future. We evaluate our framework on three real-world datasets and show that our model can better capture the spatial–temporal relationships and outperform the most discriminative state-of-the-art methods.