A hybrid machine-learning framework combining hydrodynamic simulation and tide-gauge observations for storm-surge nowcasting in the southern North Sea

Abstract. Storm surges are a major coastal hazard that can develop within a few hours, so accurate short-term predictions are essential for early warning and emergency response. Operational hydrodynamic models remain too computationally demanding to deliver on-demand, high-frequency forecasts at the required resolution, while purely data-driven models suffer from the limited duration of tide-gauge records. We propose a hybrid machine-learning framework that combines tide-gauge observations from 22 stations along the southern North Sea coast with surge simulations from the COupled Hydrodynamical–Ecological model for REgioNal and Shelf seas (COHERENS) and ERA5 atmospheric forcing. Two architectures — a Transformer and a parallel LSTM+Transformer — are trained on this combined input and compared against (i) the same architectures trained on gauge observations alone and (ii) the COHERENS operational baseline. For 2-hour-ahead nowcasting at 10-minute resolution, the hybrid LSTM+Transformer reaches an average root-mean-square error (RMSE) of 0.055 m (0.033–0.100 m across stations), substantially better than COHERENS (0.164 m; 0.137–0.185 m) and modestly better than the pure data-driven baseline (0.061 m; 0.043–0.100 m); the hybrid Transformer reaches 0.096 m (0.067–0.121 m). The hybrid models also reproduce the magnitude and timing of extreme events well. Applied recursively, the hybrid LSTM+Transformer degrades to an average RMSE of 0.146 m at the 12-hour horizon, remaining well below the COHERENS baseline. Inference takes less than one minute on a single GPU, making the framework directly suitable for operational nowcasting. The approach should generalize to other coastal regions with comparable observational coverage.