Coastal Flooding Forecasting System
14-day coastal flooding risk forecasts from seven days of hourly sea-level measurements.
Overview
A forecasting model must learn from recent sea-level patterns and remain useful at tide-gauge stations it has not seen during training. Probability estimates also need to fit inference runtime constraints.
A reproducible Python training and inference pipeline converts rolling seven-day windows of hourly sea-level data into fourteen-day flooding-risk predictions.
How it works
- Hourly sea-level observations are cleaned and organized into rolling input windows.
- Time-series features feed models built with XGBoost and LightGBM.
- Probability calibration and runtime-constrained inference support risk predictions.
Technical implementation
- Pandas and NumPy support preprocessing and time-series feature engineering.
- Seven-day rolling windows standardize the recent history available to the model.
- Training and inference are organized as a reproducible pipeline.
- Probability calibration addresses the interpretation of model outputs as risk.
Cross-station validation addresses one important generalization question; it does not by itself establish performance during rare extremes or changing coastal conditions. No measured accuracy, runtime, or station coverage is claimed here.
Results
1st Place in the Scientific Mood Fair Challenge at HackAI 2026. Implemented the training and inference pipeline with station-held-out evaluation.
Leave-one-station-out cross-validation evaluates generalization to unseen tide-gauge stations. The supplied resume also documents probability calibration and runtime-constrained inference. Evaluation scores and calibration plots were not supplied.