Time-weighted ensemble long–short-term memory for delay prediction in railway systems

Railway systems are complex and consist of many subsystems working together to ensure smooth and timely operation. Delays that cause deviations from the scheduled operations can lead to cascading issues throughout the railway network. These delays can be a result of a large number of causes, from adverse weather to random equipment failure. Delay prediction is crucial in the mitigation of delay effects. This paper proposes a data-driven approach to the prediction of railway delays. Applying Recursive Neural Networks in this data-driven approach allows it to exploit the sequential nature of train operations without having to make intermediate predictions common in event-driven approaches. Raw scheduling data are processed to compute features relevant to the analysis of delay and its propagation. In the proposed method, each row of data is weighted according to its temporal distance to the prediction horizon, assigning increased importance to more recent events. These time-weighted data are analysed at different scales via the component long–short-term memory (LSTM) models in the proposed ensemble model. Data of a 1-year time period from the Historical Service Records of the British Railway were used to train and test the proposed time-weighted ensemble LSTM architecture. The proposed architecture showed an improved performance compared to other data-driven models implemented as benchmarks, outperforming them by achieving a root mean squared error of 0.27 min, a mean absolute error of 0.17 min as well as a coefficient of determination (R2) value of 0.9875.