HYDRA

Problem

National water models often struggle with local precision. During events like Hurricane Helene, NWM forecasts significantly underestimated peak flows in complex terrain, leaving communities with uncertain flood warnings.

Impact

Achieved up to 48% improvement in predicting streamflow across unregulated Appalachian watersheds. Live at hydramodel.ai. Manuscript in preparation for Water Resources Research (AGU).

Technical Approach

• Developed an advanced 1-million parameter hybrid GRU-Transformer model to catch residuals that physical models miss.
• Integrated multi-source data (NWM v3.0+, ERA5, USGS) for a comprehensive view.
• Designed leakage-safe evaluation to ensure performance holds up in real-time constraints.
• Preparing manuscript for submission to Water Resources Research (AGU).

Architecture Decisions

• Residual-correction pipeline that ingests NOAA NWM forecasts, forcing signals, and basin context.
• Transformer encoder for multi-scale temporal context, paired with a recurrent residual head.
• Config-driven training and evaluation with strict train/validation/test time boundaries.

Reliability and Evaluation

• Leakage-safe splitting by basin and time horizon to match operational inference constraints.
• Reproducible runs through fixed seeds, immutable data artifacts, and tracked configuration snapshots.
• Stress-tested against seasonal drift and missing-sensor windows for production realism.

Delivery and Operations

• Packaged experiments with Hydra + MLflow for repeatable model iteration and comparison.
• Built deployment-ready outputs for downstream operational routing and dashboard consumption.
• Defined clear handoff surfaces so model updates can be integrated without pipeline rewrites.

Results

• Up to 48% improvement in predictions across unregulated Appalachian watersheds.
• Consistent improvement across varying forecast horizons and extreme storm events.
• Demonstrated stability under seasonal distribution shifts.

What I Learned

• Hybrid architectures can correct physical model biases without overfitting.
• Rigorous evaluation design is the difference between a demo and a product.
• Balancing research goals with operational constraints mirrors how ML systems are built in production teams.