A Forecasting Framework for Ultrashort-Term Wind Power Based on Transformer-BiLSTM
TL;DR
A novel forecast framework integrating improved complete ensemble empirical mode decomposition with adaptive noise, sample entropy reconstruction, a Transformer bidirectional long- short term memory (Transformer-BiLSTM) hybrid neural network, and the RIME optimization algorithm is proposed.
Abstract
In response to the challenges of ultrashort-term wind power forecasting, such as strong nonstationarity, multiscale dynamic features, and sensitivity of model hyperparameters, this paper proposes a novel forecast framework integrating improved complete ensemble empirical mode decomposition with adaptive noise (ICEEMDAN), sample entropy (SE) reconstruction, a Transformer bidirectional long- short term memory (Transformer-BiLSTM) hybrid neural network, and the RIME optimization algorithm. Firstly, high-impact factors are selected through dual filtering using Pearson correlation coefficients and the maximal information coefficient (MIC). Next, ICEEMDAN is employed for adaptive decomposition of the original power sequence, with SE used to assess the complexity of each intrinsic mode function (IMF), enabling the separate reconstruction of high-frequency and low-frequency components. Subsequently, a Transformer-BiLSTM hybrid model is constructed, with key hyperparameters optimized via the RIME algorithm. Finally, multiscale forecast results are fused through linear superposition. Multidimensional ablation experiments based on actual operational data from wind farms demonstrated that this forecasting approach outperforms existing methods in predictive accuracy.