Temporal-Spatial Prediction of Trip Demand Using Neural Networks and Points of Interest

One of the major challenges in intelligent transportation and smart city is efficient assignment of vehicles to satisfy trip demands, and thus to achieve road utilization and energy efficiency. Traditional order-driven dispatches often result in instability of resource utilization toward unsatisfactory user experience and waste in scheduling, due to lacking of knowledge in future trip demands. Accurate and immediate prediction of trip demand of each region is important for coordinating regional autonomous fleet allocation and improving service efficiency. This paper proposes a temporal-spatial prediction model for trip demand based on points of interest (POI) and multivariate long short term memory (MLSTM) model. Parameters of models are trained using heterogeneous data sets. Influence of meteorology information on the reliability of historical trip demand data is considered in temporal prediction, while the regional imbalance caused by POI is considered in spatial prediction. The prediction accuracy of MLSTM is improved when compared with the univariate LSTM. A Beijing taxi trajectory dataset is used to verify the temporal-spatial prediction method.