A multi-model dynamic adaptive fusion model for irregular wave forecasting

Accurate forecasting of irregular ocean waves is important for offshore engineering safety and wave energy applications. Irregular wave sequences exhibit strong non-stationarity, multiscale dynamics, and sparse extreme events. To address this problem, this study proposes MM-DAF, a multi-model dynamic adaptive fusion forecasting model. It integrates two structurally complementary submodels in parallel: PatchTST (Patch Time Series Transformer) for long-sequence dependency modeling and N-HITS (Neural Hierarchical Interpolation for Time Series) for hierarchical interpolation-based multiscale decomposition and reconstruction. In addition, an LSTM-based gating network is constructed to encode the historical wave context, enabling adaptive dynamic weight fusion and bias compensation. Visualization of the learned coefficients further indicates that the gating network adapts across non-stationary windows. Bayesian optimization is employed to tune the hyperparameters of the submodels. Experiments are conducted on AQWA-generated numerical datasets and wave-tank experimental data over multiple forecasting horizons. The results show that MM-DAF outperforms representative baselines and static-fusion baselines in both conventional accuracy metrics and extreme-wave error metrics. On Dataset A, MM-DAF reduced RMSE by 44.8% and 28.3% relative to the best single model at the 32-step and 96-step forecasting horizons, respectively; on Dataset B, the corresponding RMSE reductions were 27.2% and 38.0%. Overall, MM-DAF provides an interpretable and practical framework for irregular wave forecasting, and offers a basis for future extensions to multivariable inputs and improved reliability under extreme events.