A novel method for multivariate short-term offshore wind forecasting via time–frequency clustering and inverted attention

Short-term offshore wind power forecasting is crucial for stable power system operations. However, accurate forecasting is hindered by multivariate interactions that generate multi-scale repetitive patterns and obscure cross-variate correlations. In this paper, we propose a hierarchical framework, the Time–Frequency Clustering Inverted Transformer, for multivariate offshore wind power forecasting. First, a Time–Frequency Clustering component applies Wavelet Packet Decomposition to each series and quantifies sub-series similarity by overall activity and evolutionary trend, grouping repetitive patterns into structured clusters. Second, an inverted Transformer captures multivariate correlations within clusters by treating time points of individual sub-series as variate tokens, enabling self-attention to focus on multivariate correlations rather than temporal dependencies. On two real-world offshore wind datasets (horizons 8–48 h), our proposed framework reduces MSE/MAE by 14.11% and outperforms 12 recognised baselines (e.g., PatchTST, TimesNet), with the advantage persisting even when the TFC component is applied to the baselines. Moreover, our method demonstrates remarkable generalisability on three public datasets (Solar-Energy, Traffic, and ECL), reducing MSE/MAE by 7.36%. These results indicate that associating repetitive patterns with attention to cross-variate structure materially improves multivariate offshore wind power forecasting. • Reveals potential risks from excessive variates, like multi-scale entanglement. • Reveals latent associations from entangled, multi-scale repetitive patterns. • Clusters series by behavioural similarity, not Euclidean distance. • Validates iTransformer as a superior base predictor for multivariate forecasting. • TFC-iTransformer performance by over 14% with Time–Frequency clustering.