EMD‐EN‐AE‐BLS:  A Lightweight Prediction Model for Ultra‐Short‐Term Wind Power Based on Signal Decomposition and Dual‐Stage Feature Selection

ABSTRACT

To address the problem of low accuracy in wind power prediction, this study proposes a four‐progressive‐stage lightweight hybrid model. The model integrates empirical mode decomposition (EMD), ElasticNet (EN), autoencoder (AE), and broad learning system (BLS), termed EMD‐EN‐AE‐BLS. Firstly, the original power series is decomposed into multiple intrinsic mode functions and a residual component via EMD. Subsequently, an EN‐based feature contribution assessment is employed to mitigate multicollinearity while preserving critical predictors. Next, the high‐contribution parts are processed through an AE to extract latent features. Finally, these refined features are fed into the BLS for predictive modelling, where the system's streamlined network structure ensures both forecasting accuracy and computational efficiency. Experimental results show that EMD‐EN‐AE‐BLS achieves an R ² improvement of 3.36% over the BLS and 1.77% over the EMD‐EN‐BLS on Dataset 1. Except for the EMD‐EN‐AE‐extreme gradient boosting model, the EMD‐EN‐AE‐BLS achieves an average reduction of approximately 97.6% in training time compared with the other five four‐progressive‐stage models.