Efficient wind power prediction via hybrid preprocessing and enhanced Reformer

Accurate wind power forecasting is critical for renewable energy integration but remains challenging due to high-dimensional, noisy, and long-sequence data. This study proposes an enhanced Reformer-based framework that combines hybrid preprocessing with a lightweight architecture to address these issues. The method integrates LSTM-based anomaly detection, variational autoencoder (VAE) compression for dimensionality reduction, and multi-scale feature extraction to capture both local fluctuations and global trends. By optimizing the locality-sensitive hashing (LSH) attention mechanism and introducing segmented feed-forward networks, the proposed model significantly reduces computational complexity and memory usage without sacrificing accuracy. Experiments on real-world datasets show that the framework outperforms six benchmark models, achieving a 4.5% improvement in prediction accuracy and a 23% reduction in computational cost. These results demonstrate the model’s effectiveness in delivering robust and efficient wind power forecasting under complex meteorological conditions.