Train delay prediction in Guangzhou-Shenzhen-Hong Kong High-speed Railway with improved CNN-LSTM Model

Abstract To address the issue of enhancing the accuracy of delay predictions for high-speed trains, a novel model based on CNN and LSTM architectures has been proposed. This model incorporates the KAN model to supplant the traditional fully connected layer, thereby augmenting the model’s predictive accuracy. The proposed architecture comprises three primary layers: the CNN layer, which is tasked with processing spatial features; the LSTM layer, which serves as the core of the model and is dedicated to the analysis of time series data; and the KAN layer, which replaces the fully connected layer to generate the output results. Additionally, an integrated feature selection algorithm has been developed to identify relevant features, with the selected feature variables subsequently input into the CNN-LSTM-KAN model. Experiments conducted on the Guangzhou-Shenzhen-Hong Kong (GSH) high-speed Railway dataset from early 2020 demonstrate that our CNN-LSTM-KAN model surpasses the performance of XGBoost, SVR, BiLSTM, and CNN-LSTM models, as indicated by the metrics of MAE, RMSE, and R-squared. The model’s prediction accuracy within a three-minute error margin reached as high as 96.11%, which is an improvement of 5.8% compared to the highest accuracy without the KAN layer.