An explainable multi-feature dimensionality reduction framework: Considering trend inconsistency in wind power sample

Offshore wind power prediction is significantly challenged by data quality issues arising from various factors such as environmental conditions and measurement errors, which severely compromise prediction accuracy and stability. This paper reveals a previously overlooked phenomenon in offshore wind power data, referred to as sample trend inconsistency, where dynamic offsets in meteorological features distort their expected physical relationship with power output. Such inconsistencies hinder the extraction of key features, disrupt feature–power coupling, and ultimately lead to degraded prediction performance. To address this issue, a Differential Compensation Dimensionality Reduction (DCDR) method is proposed to actively detect and mitigate trend inconsistencies during the dimensionality reduction process. Following normalized preprocessing of raw multi-feature meteorological data, the proposed DCDR method is employed to enhance sample trend consistency and perform dimensionality reduction by optimizing a selection coefficient to retain the most informative feature subset, which is then fed into deep learning models for training and accurate power forecasting. Experimental results demonstrate that DCDR achieves significant improvements over conventional dimensionality reduction methods, reducing RMSE and MAE by 21.1 % and 12.1 %, respectively. Furthermore, global feature importance analysis based on Shapley Additive Explanations (SHAP) confirms that the features retained by DCDR contribute more strongly to prediction accuracy and show improved consistency with the underlying physical relationships governing power output, thereby providing a more robust and interpretable framework that can enhance the operational reliability of offshore wind power forecasting models.