|
Existing data-driven wind power prediction models inevitably suffer from the learning offset problem in the modeling process, which leads to these models being biased to learn only part of the samples in the distribution set, resulting in poor generalization ability in the practical applications. A short-term wind power prediction method based on the correction of model learning offset is proposed to address the aforementioned issue. Firstly, by exploring the principle of differential samples causing model prediction performance bias, temporal samples are classified and characterized. Subsequently, for extreme weather samples, abnormal samples, and similar imbalanced samples that are difficult to predict in historical data, scene generation, progressive mask detection, and sample feature enhancement strategies are used to jointly correct the bias learned by the model. Finally, the Shapley value method is used to evaluate the importance of various samples, in order to verify the necessity and rationality of the offset correction strategy. Practical examples show that the proposed method can significantly improve the short-term wind power prediction accuracy of various models, and has good generalization in multi-scenario modes.
|