Deep learning-based sea level anomaly forecasting around Taiwan Island integrating ConvLSTM and attention mechanism

Sea surface height around Taiwan Island shows complex multi-scale variability, posing a major forecasting challenge. To address this, the Taiwan Island Adjacent Seas Sea Level AI Forecaster (TAS-SLAF), a deep learning framework integrating convolutional layers, ConvLSTM, and the Convolutional Block Attention Module (CBAM) to capture spatiotemporal features, is applied for 15-day sea level anomaly (SLA) prediction in this sea area. Fed with 15-day combined SLA, sea surface wind speed and ocean current data, the TAS-SLAF achieves root mean square errors (RMSEs) of 2.38 cm, 4.83 cm and 5.95 cm on the 5th, 10th and 15th forecast days, respectively, with higher RMSEs in shelf/slope regions than in the open ocean. Over a 3–10 day horizon, it reduces RMSE by 11.46%–50.76% compared to leading numerical models (GOPAF and ESPC-D-V02), with greater improvement in the open ocean. The TAS-SLAF also performs robustly during large Kuroshio intrusion events northeast of Taiwan Island, showing better agreement with observations than reanalysis data. This study highlights the TAS-SLAF’s application value in improving sea level predictability around Taiwan Island with markedly error reduction.