W. Meng, K. Zhang, G. Xi, C. Ma, X. Huang, X. Wu, Y. Lai

Pages: 195-208

Abstract
The expressway service area's short-term traffic flow forecast is helpful for adjusting expressway traffic and formulating time-sensitive traffic flow adjustment strategies. To solve the problems of strong stochastic fluctuation of traffic flow, insufficient degree of traffic flow data mining and low prediction accuracy in common deep learning prediction models, a hybrid model combining ensemble empirical mode decomposition and bidirectional gated recurrent unit (EEMD+BI_GRU) is proposed for expressway intelligent service area traffic flow prediction. The model decomposes the traffic time series using the EEMD algorithm, with the aim of reducing the non-stationarity of the initial traffic series and providing a potential feature set for prediction. The BI_GRU structure as a prediction model enables the model to process traffic data more deeply and efficiently to improve prediction accuracy. Based on the real traffic flow data of Taishi service area in Gansu Province, several reference models were developed and analyzed in comparison with the suggested model. The results demonstrate that, as compared to traditional mainstream models, the method of applying signal decomposition technique-provided sequences as input characteristics may significantly enhance prediction accuracy.
Keywords: traffic flow prediction; time series data; EEMD; intelligent transportation; BI_GRU