A distributed EMDN-GRU model on Spark for passenger waiting time forecasting

It is hard to forecast waiting time from mobile trajectory big data on the traditional centralized mining platform, and especially the taxi driving direction cannot be clearly distinguished by the GPS trajectories of taxicabs in intelligent transportation systems. To this end, we propose a direction identification method (named CA-D) combined with C oordinate A xis and GPS D irection to distinguish taxi driving direction and then establish a distributed model (named EMDN-GRU) on Spark to forecast passenger waiting time based on an E mpirical M ode D ecomposition (EMD) algorithm with N ormalization and a G ated R ecurrent U nit (GRU) model. Specifically, in the process of waiting time forecasting, the CA-D method is used to differentiate directions to reduce the data interference caused by different taxi driving directions. Furthermore, the EMD algorithm with normalization is utilized to process large-scale GPS trajectory data. Finally, the GRU model with adjusted parameters is employed to forecast the time-series data obtained in the previous step, and the forecasting results are denormalized and superimposed to produce waiting time for passengers. Compared with LSTM, GRU, EMD-LSTM, EMD-GRU, CNN, and BP, the experimental results from a case study indicate that EMDN-GRU is significantly superior to others. In particular, from three data sets of weekday, weekend, and one week, the MAPE values of EMDN-GRU are reduced by 92.48%, 95.01%, and 90.47% at most.