Elham Ahmadi: " Spatio-Temporal Transfer Learning with Station Embeddings for Train Delay Prediction"

Topic: A Unified Spatio-Temporal Transfer Learning Framework with Station Embeddings for Network-Wide Train Delay Prediction 

Precise predictions of train arrival delays are essential for the dependability of railway operations, particularly at interchange locations, where delays may propagate across travel segments. Most traditional models analyze each segment independently, which limits their ability to consider the sequential dynamics of delays in multi-leg rail itineraries. This study introduces a unified spatio-temporal transfer learning framework that employs Long Short-Term Memory (LSTM) networks to capture both spatial dependencies and temporal delay patterns among stations through station embeddings. This framework is designed to overcome this limitation. The model is first trained on upstream data and then refined using downstream data. This allows it to acquire temporal knowledge that can be applied to connected lines while simultaneously adapting to the unique characteristics of each segment. This design enables the model to be applied to the entire transportation chain rather than a single line at a time. The model learns to generalize across the entire transportation chain instead of treating each line as a separate system. In addition, the proposed multi-task framework jointly predicts both continuous arrival delays and the probability of exceeding operational thresholds.Evaluation on operational data from German regional railways shows that the LSTMbased transfer learning model achieves a mean absolute error of 0.33 minutes and R^2 above 0.94, with classification metrics exceeding 95%. A comparison with a hierarchical XGBoost meta-model reveals that although XGBoost attains a slightly superior R^2, the LSTM framework delivers consistent minute-level predictions and more effectively captures inter-leg relationships. It has been determined that both models have elevated MAPE values, which can be attributed to the predominance of near-zero delays. These findings underscore the importance of transfer-aware sequential modeling for precise delay prediction at critical transfer points, thereby supporting more resilient scheduling and real-time decision-making in rail operations.

Elham Ahmadi is a PhD candidate at Kühne Logistics University, where she has been working since May 2022 under the supervision of Prof. Dr. André Ludwig and Prof.Dr. Henrik Leopold. In the past years, she has been deeply involved in the CargoSurfer project, a federally funded initiative that aims to build a digital, machine-learning–driven platform for multimodal logistics. Her recent works combines advanced data science, causal ML, and real transportation data to understand and predict delays in regional rail systems—an area where she has become increasingly specialized.