A Dual-Tower Global–Local Framework for Short-Term Multi-Buoy Significant-Wave-Height Forecasting

Abstract

Accurate short-term significant-wave-height forecasting is important for the safe and efficient operation of coastal and offshore activities. However, forecasting across sparse buoy networks remains challenging because wave conditions exhibit complex temporal variability and spatial dependence. This study proposes an Enhanced Crossformer, a dual-tower spatiotemporal framework that combines a Crossformer branch for modelling long-range inter-station dependencies with a Cheb–LSTM branch for capturing localised graph-based interactions. The framework was evaluated using 3-hourly significant-wave-height observations from 52 National Data Buoy Centre stations along the U.S. West Coast over the period 2020–2024. Missing observations were first estimated using a graph neural network imputation model, which achieved test-set R2 values ranging from 0.85 to 0.95 on artificially masked observations. Forecasting performance was then assessed for lead times of 3–24 h and compared with LSTM, Transformer, Cheb–LSTM, and Crossformer baselines. The Enhanced Crossformer achieved the lowest MAE, MSE, and MAPE values and the highest R2 and correlation values at all forecast steps. Relative to the strongest baseline, it reduced MSE by up to approximately 8.7% at short lead times and maintained positive improvements through the 24 h horizon. Forecast skill was strongest under low-to-moderate sea states but decreased for high-energy wave conditions and several sheltered stations in the Southern California Bight. These results demonstrate the potential of combining global attention with graph-based local recurrence for short-term multi-buoy wave forecasting.