A CNN-Bi_LSTM parallel network approach for train travel time prediction

Convolutional neural networks (CNNs) offer a broad technical framework to deal with spatial feature extraction and nonlinearity capture, whereas they cannot process sequence data and cannot capture the dependencies between the sequence information. Therefore, this paper proposes an improved deep learning model CNN-Bi_LSTM that combines the CNN, Bi_LSTM (i.e., bidirectional long short-term memory network), and fully connected neural network (FCNN) to process the complex dataset for the train travel time prediction. As a result, the presented deep learning framework can capture both the long-and short-term features of complex datasets and the characteristics of time series data. Besides, the multi-feature data fusion processing method is realized with the help of a parallel learning mechanism and the fully connected neural network. Based on a real-life case study of China Railway Express (Chengdu-Europe), the superiority of the CNN-Bi_LSTM model on the train travel time prediction is systemically evaluated and demonstrated, compared with the baseline models of Holt-Winters model, random forest (RF), support vector regression (SVR), LSTM, Bi_LSTM, LSTM with attention mechanism (LSTM_Attention), convolution-based LSTM (CLSTM), CNN_LSTM, hybrid deep learning model (CNN_GRU1), temporal convolutional network (TCN), and parallel deep learning model (CNN_GRU2). Moreover, the values of MSE, RMSE, MAPE, and MAE obtained from the CNN-Bi_LSTM model are equal to 4.647, 2.156, 2.643, and 1.769 respectively Consequently, it is concluded that our proposed CNN-Bi_LSTM model has good prediction results, and it is suitable for the train travel time prediction of China Railway Express.(c) 2022 Elsevier B.V. All rights reserved.

Automated summary

The paper proposes an improved deep learning model CNN-Bi_LSTM that combines CNN, Bi_LSTM, and FCNN to predict train travel time for complex datasets. The model can capture both long-term and short-term features of complex datasets and time series data, and incorporates a parallel learning mechanism and fully connected neural network for multi-feature data fusion. Through a case study comparing with baseline models, such as Holt-Winters, random forest, support vector regression, LSTM, Bi_LSTM, LSTM with attention mechanism, convolution-based LSTM, CNN_LSTM, hybrid deep learning model, temporal convolutional network, and parallel deep learning model, the superiority of the CNN-Bi_LSTM model in train travel time prediction is demonstrated. The evaluation metrics MSE, RMSE, MAPE, and MAE further support the accuracy of the model.