Ensemble Empirical Mode Decomposition-Based Gated Recurrent Unit Model for Short-Term Metro Passenger Flow Prediction

Abstract

An accurate short-term passenger flow prediction provides essential data support for Urban Rail Transit (URT) organizations. However, the short-term URT passenger flow data is neither linear nor stationary, which results in the inaccuracy of prediction using only single deep-learning models. In this paper, we proposed a hybrid model that integrates Ensemble Empirical Mode Decomposition (EEMD) and Gated Recurrent Unit (GRU) to achieve faster and better results for short-term passenger flow prediction. The approach contains four steps. First, the EEMD algorithm decomposes passenger flow data into several intrinsic mode functions (IMF) and a residual function. Second, the Spearman and Kendall correlation coefficients were used to examine the IMF components to select components that are highly related to the URT passenger flow. Third, the selected features and historical passenger flow are combined as inputs for the GRU model, respectively. Last, the outputs of each input feature and historical data were synthesized to obtain the result. We used the passenger flow of Xipu, a station of Chengdu Metro Line 2, to demonstrate the method. The experimental results indicate that the proposed EEMD-GRU architecture, compared with benchmark models, achieves superior performance. It can reduce the training time by 26.18% and improve the average accuracy by 29.58%, which shows great validity for large-scale short-term metro passenger flow forecasting.