An improved TSLANet model for spatio-temporal forecasting correction of significant wave height

Accurate forecasting of significant wave height (SWH) is crucial for the safety of marine engineering, shipping, and fisheries. To address errors in numerical forecast results arising from the coexistence of multi-scale periods, strong temporal continuity, and spatiotemporal coupling in SWH time series, this study proposes an improved model based on the time series lightweight adaptive network (TSLANet), named the multi-scale feature-enhanced correct model (MS-FECM). This model incorporates a specially designed multi-scale adaptive spectral block (MS-ASB) and a convolutional neural network-bidirectional long short-term memory (CNN-BiLSTM) hybrid module, which replace the adaptive spectral block (ASB) and the interactive convolution block (ICB) in the TSLANet model, respectively. This achieves full-scale periodic coverage processing of SWH data during the correction process, as well as the synergistic modeling of local details and temporal correlations. The model was validated using observed SWH data from the Sea of Japan. The results indicate that, compared to the original numerical model forecast results, the corrected model significantly reduced prediction errors, with a MAE of 0.2827, a MAPE of 0.3149, and a RMSE of 0.3729, representing a 17.1% reduction in RMSE relative to the numerical model forecasts. When compared with neural network models such as FPPformer, PatchTST, and CNN-BiLSTM, the method proposed in this study demonstrated superior performance across metrics including MAE, MAPE, and RMSE. This validates the model’s effectiveness and superiority in correcting dynamic errors in numerical wave-height forecasts, demonstrating excellent correction performance under complex sea conditions.