We investigate whether coupling a time‑series foundation model with hydraulic project knowledge improves surrogate forecasting of HEC‑RAS water‑surface elevation (WSE). The proposed KG‑Chronos‑2 integrates a frozen Chronos‑2 predictor, exact‑state residual decoding, graph‑conditioned historical retrieval, and input‑aligned correction. Task‑specific fine‑tuning uses the 2008 simulation, while evaluation spans 64 fixed 24‑hour windows from the 2011 and 2002 simulations, covering 4,675 cross‑sections in 71 reaches. KG‑Chronos‑2 is compared against persistence, a residual LSTM, project‑conditioned recurrent GeoFNO, a hydraulic DCRNN‑style model, and the frozen Chronos‑2 baseline. It achieves an event‑balanced RMSE of 0.246970 in native WSE units, reducing RMSE by 14.13% relative to frozen Chronos‑2, 29.38% relative to the DCRNN‑style model, and 39.54% relative to recurrent GeoFNO. The 95% hierarchical‑bootstrap interval for its RMSE difference from frozen Chronos‑2 is [-0.075177, -0.016317]. KG‑Chronos‑2 also records the lowest active‑window and final‑lead RMSE among the six systems. These results support that coupling a frozen temporal predictor with project knowledge yields superior warm‑start HEC‑RAS forecasting on the fixed benchmark.
Review