S-Bahn Delay Prediction (AI-CPS Project)

This repository contains our course project for predicting public transport delays (in minutes) for a given stop and timestamp.
We build two models: a baseline OLS regression and an improved ANN (neural network) with feature engineering.

We have trained the model with data we collected ourselves via the VBB API. However, it is also conceivable to collect (presumably less accurate) data via the gtfs standard.

What it does

• Predicts delay_minutes for each (stop_id, timestamp)
• Uses engineered features such as:
  • time features (hour/weekday + cyclical sin/cos)
  • rush-hour / weekend indicators
  • station one-hot encoding
  • lag features (previous delays per station)
  • station statistics (mean/std delay)

Project structure (high level)

• results/ – trained models, scalers/encoders, plots, and exported artifacts
• scripts/ or src/ – training, evaluation, and inference code (ANN + OLS)
• docker/ (or images/) – Dockerfiles for Subgoal 6/7 images and compose setups

Key outputs

• ANN model: ann_improved_model_new.keras
• OLS model: currentOlsSolution.pkl
• Activation data: activation_data.csv
• Plots: training history, predicted-vs-actual, feature importance

Docker (Subgoals 6 & 7)

We publish three Docker Hub images and provide docker-compose files for running:

1. knowledgeBase – model artifacts under /tmp/knowledgeBase/
2. activationBase – activation data under /tmp/activationBase/
3. codeBase – applies ANN or OLS to the activation data and writes prediction outputs into the shared /tmp volume

License

AGPL-3.0