Short‐term traffic travel time forecasting using ensemble approach based on long short‐term memory networks

To enhance the control effect of intelligent transportation system (ITS) of expressway, avoid blind travel causing traffic congestion, improve road capacity, and achieve the purpose of smooth and efficient operation of the network, 3σ criterion and 2σ cycle removal criterion were used to eliminate the abnormal data of travel time. Thereafter, a combined prediction model of Long Short‐Term Memory (LSTM) based on the complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN) algorithm and attention mechanism is proposed. Travel time data, collected from the real world, is decomposed into several Intrinsic Mode Functions (IMFs) and a residue with the CEEMDAN. Then, the decomposed sub‐sequence is predicted and the final prediction result is obtained by comprehensive superposition. The results indicate that: the CEEMDAN algorithm has significantly improved the prediction performance. Compared with the AT‐LSTM model, mean absolute percentage error (MAPE) is reduced by 10.75%, 24.84%, 31.80%, 24.98%, 14.29%, and 31.42% when the time window was 3, 6, 9, 12, 15, and 18, respectively, in the dataset of Yaxi expressway. In the dataset of Chengnan expressway, the number of error points with MAPE greater than 40% is significantly reduced after the adoption of the LSTM model. After the introduction of CEEMDAN algorithm, the error at each time point is controlled within 30%. The proposed model controls the error at each time point within 25%. In different datasets, the proposed model has better prediction ability for highly non‐linear travel time, and can better capture the trend abrupt characteristics of rush hour traffic flow. The combined model of CAT‐LSTM proposed in this paper has high accuracy for travel time prediction with different time windows. The proposed model has a better effect on short‐term prediction and the degree of ‘delay situation’ is relatively light.