Multi-horizon forecasting of coastal high water temperature events for operational marine-heatwave early warning in South Korean aquaculture: a walk-forward benchmark of eleven statistical and deep-learning models

IntroductionHigh water temperature (HWT) marine-heatwave events along the southern coast of South Korea cause repeated mass mortalities in cage aquaculture of olive flounder (Paralichthys olivaceus), rock bream (Oplegnathus fasciatus), red sea bream (Pagrus major) and Korean rockfish (Sebastes schlegelii). Existing operational forecasts rely on coarse numerical ocean models that lack the temporal granularity and site specificity required by farm managers, while the recent wave of deep-learning sea surface temperature (SST) models has been tuned almost exclusively on smoother open-ocean grids.MethodsWe present the Hybrid GNN-BiLSTM, a 3.1 M-parameter graph–temporal architecture combining a weighted spatial GNN with a BiLSTM temporal core, reversible instance normalization (RevIN) and a persistence skip connection, and benchmark it against ten alternatives – three reference baselines (Persistence, Seasonal-naïve, Climatology), the ARIMAX statistical model, four lightweight or linear deep models (DLinear, PatchTST, N-HiTS, CNN-LSTM) and two attention-based models (Informer, Temporal Fusion Transformer) – on a five-year hourly insitu panel from 30 monitoring stations along the Tongyeong–Geoje–Yeosu coast (2020–2025) under a unified rolling-origin walk-forward protocol (720 h lookback, 168 h horizon, one forecast per station every 24 h across a 12-month validation window). Marine-heatwave detection is scored at tiered 26/28/30 °C operational thresholds.ResultsThe Hybrid GNN-BiLSTM achieves the lowest 1 h RMSE (0.057 °C), 0.85 °C at h = 72 and 1.24 °C at h = 168, with the highest long-horizon residual skill of any deep model on this panel (SS_res = 0.71 at h = 72). CNNLSTM attains the lowest mean horizon root-mean-square error (RMSE = 0.636 °C) thanks to a nearly flat error curve, and PatchTST is competitive at short-to-medium horizons (0.694 °C mean). The Hybrid GNN-BiLSTM reaches F1 = 0.82 at 26 °C and F1 = 0.65 at 28 °C, and a probabilistic Gaussian negative-log-likelihood head supplies calibrated 90 % prediction intervals without post-hoc recalibration.DiscussionA 38 h mean advisory lead time was demonstrated in the 2024 NIFS regional HWT alert period, indicating that a site-specific graph–temporal model can deliver operationally actionable early warning for coastal aquaculture.