Research

HYDRA: Hybrid Deep-learning for Residual Analysis

Post-processing streamflow forecasts with a hybrid GRU-Transformer to improve accuracy while respecting physical constraints. Seeking collaborations across hydrology, ML, and engineering to build actionable decision support.

Website GitHub

HYDRA Transformer-GRU Hybrid Architecture connecting sequence encoding, multi-scale temporal attention, and regime-conditioned residual correction.

Architecture

Inputs: Operational NWM forecasts, meteorological forcings (ERA5), static attributes.
Temporal Encoding: GRU for capturing sequential patterns.
Encoder: Transformer capturing multi-scale temporal context.
Pipeline: Automated feature normalization and leakage-safe splitting.

Evaluation

Metrics: RMSE, NSE, KGE (Hydrology-standard metrics) with bootstrap confidence intervals.
Validation: Spatial holdouts to test generalization across 3 Appalachian sites.
Diagnostics: Flow regime analysis across baseflow, rising limbs, and recessions.

Constraints

Must run within real-time operational latency limits.
Robustness to complex terrain and flashy watershed dynamics.

Reproducibility

Config-driven experiments (Hydra) with 19 unique configurations.
Strict data versioning and artifact tracking.
Full environment captures for every run.