EEMD-PE-LSTM Based Traffic State Prediction Method for Freeway Section

The rapid development of the highway network and the diversification of traffic demand have made traffic congestion, road carrying capacity bottlenecks, road design optimization and other issues more and more prominent. This seriously restricts the travelling experience of travelers and the service effectiveness of traffic management departments. To accurately quantify the gain of traffic conditions and weather conditions on the prediction performance of traffic flow parameters, we a combined prediction model of traffic flow parameters of highway sections construct based on the ensemble empirical modal decomposition (EEMD), permutation entropy (PE) and long short-term memory (LSTM). The study applies the EEMD algorithm to decompose the average travelling speed sequence, screens and integrates the decomposed components through the PE algorithm and proposes to perform spatio-temporal matching and feature grouping of the traffic and weather data to identify the most influential factors and their interaction modes; combined with the sliding time window strategy, the input configurations are dynamically adjusted. With the LSTM network as the core, the optimal history sequence length and feature combination are determined by iterative optimizations, and then the optimal value of the average driving speed of the target road section is obtained. At the same time, the regional characteristic-oriented traffic state determination mechanism is proposed, i.e., the 85% quartile of the average driving speed of the road section is adopted as the normal speed benchmark. Taking a highway in Hubei Province as a case study, the empirical results show that: in terms of prediction accuracy, compared with a single LSTM model, the average absolute error of the combined prediction model is significantly reduced by 73.4%; in terms of computational efficiency, it is improved by 67% compared with the EEMD-LSTM model. Especially when the length of the sliding time window is 40 min, the combined model maintains the lowest prediction error under various types of travelling scenarios and diversified feature inputs, showing good stability and robustness. In addition, the model incorporating traffic conditions reduces the prediction error range by about 60% compared to the model relying only on historical speed sequences, highlighting the key role of traffic factors in speed prediction. This study can provide scientific management decision support for traffic management departments during peak traffic periods, special events, and traffic accident emergencies.