An Ultra-Short-Term Multi-Step Prediction Model for Wind Power Based on Sparrow Search Algorithm, Variational Mode Decomposition, Gated Recurrent Unit, and Support Vector Regression

Abstract

Accurate ultra-short-term wind power prediction techniques are crucial for ensuring the efficient and safe operation of wind farms and power systems. Combined models based on data decomposition-prediction techniques have shown excellent performance in ultra-short-term wind power forecasting. This study introduces a novel ultra-short-term multi-step prediction model for wind power, which integrates the sparrow search algorithm (SSA), variational mode decomposition (VMD), gated recurrent unit (GRU), and support vector regression (SVR). An optimization variational mode decomposition technique is developed by adaptively determining VMD hyperparameters using SSA. The optimization VMD decomposes the original wind power sequence into sub-modes, and the resulting sequence of decomposed sub-modes calculates permutation entropy (PE) values. Sub-modes with similar PE values are combined, reorganized, and categorized into high-frequency and low-frequency. High-frequency sub-modes data with high complexity and non-stationarity are predicted by the GRU neural network. Low-frequency sub-modes data with low complexity and strong nonlinearity are predicted with SVR. The proposed model was evaluated against seven others using three error metrics: MAE, RMSE, and R², along with their corresponding enhancement percentages. Experimental results indicate that the proposed model extracts detailed and trend information from the wind power series more effectively and stably than the comparison models. It also demonstrates superior multi-step prediction performance, offering significant value for practical engineering applications.