Forecast of Hourly Train Counts on Rail Routes Affected by Construction Work
summary
DB Systel GmbH builds a model pipeline on AWS that converts legacy CSV to Parquet, engineers features with Pandas/Polars, trains LightGBM and PyTorch-Geometric GNNs, and persists pipelines with joblib for inference. The network covers 33,350 km and ~40,000 daily connections; construction can cut nominal capacity from 22 to ~6.6 trains/hour on a section. Models attain up to 15% more accurate real-time capacity estimates than humans, Polars solved memory limits in the cloud, and GraphSAGE wins hyperparameter search but is not always superior.
AI-written summary · about
chapters (20)
- 0:00Project team introduction
- 1:12Deutsche Bahn capacity challenge
- 3:47Rail capacity calculation
- 6:39Manual planning burden
- 7:20Data pipeline to cloud
- 8:26Train forecasting data sources
- 9:31Feature engineering for ML
- 10:56Model deployment pipeline
- 12:46Machine learning models intro
- 13:51LightGBM limitations
- 16:00Histogram-based gradient boosting
- 17:03Graph neural networks testing
- 18:49Model performance by train type
- 20:13Impact on operations
- 20:56Lessons from data modeling
- 22:22Polars solves memory issues
- 23:04Graph neural network challenges
- 23:49Security risks in persistence
- 24:29Conclusion
- 25:33Answering audience questions
related talks
21m"Capacity Prediction" instead of "Capacity Planning":
18mExtending Mobility - How AI Contributes towards the Future of Railway Transportation
48mUsing Open Source Tech to Swap Out Components of Your Data Pipeline
22mMicrodosing the Economy
17mA Brief History of Data Traffic at DB - about the Big Bang and Black Holes
48mSwiss Transport in Real Time: Tribulations in the Big Data Stack
28mContainer Tracking for Bosch Tools Supply Chain Operations
16mTidy Transit: Real Life Data Modeling for Public Transportation