Wind power forecasting using multivariate signal decomposition and stacked GRU ensembles with error correction

Abstract

Accurate wind power forecasting is vital for enhancing power system reliability, security, and cost efficiency. Recent advancements have seen the rise of ensemble systems for short-term wind power prediction. However, traditional ensembles often use conventional pre-processing and fixed-weight sub-model integration, limiting their effectiveness. This study introduces a novel hybrid ensemble system for wind farm generation forecasting that integrates multivariate signal decomposition, deep learning, and prediction error correction. The proposed system utilizes an innovative data preprocessing technique that addresses wind series non-stationarity by decomposing the series into intrinsic mode functions and a residual component. Stacked Gated Recurrent Unit (GRU) networks are then utilized to make separate predictions for each decomposed series, with the GRU structures adjusted based on the decomposition levels to create diverse forecasters. The final predictions are refined with a Bagging-Boosting mechanism, improving accuracy and capturing trends effectively. Testing on real-world data from the Tuticorin wind farm in Tamil Nadu, India, included five comprehensive experiments to assess stability and forecasting ability. Results demonstrated the proposed system’s superior performance over single models and other hybrid ensembles, providing more precise and reliable wind power forecasts.