26 February 2025 in IEEE Transactions on Instrumentation and Measurement
How to achieve accurate wind power prediction results is an important task in modern new power systems. In order to enhance the accuracy of wind power forecasting, we propose a hybrid deep-broad ensemble learning architecture based on spatial-temporal feature selection. First, a novel pre-processing framework is constructed to improve the quality of the overall dataset. The outliers in external data (such as wind speed) are removed and corrected by the quartile algorithm and Hermite interpolation method, while for internal data (such as wind power, rotor speed, and pitch angle), a comprehensive outliers detection and correction algorithm based on operating mechanisms and dynamic operation curves is developed. Moreover, to select features that are most nonlinear relevant to the training targets, the partial autocorrelation function (PACF) and maximal information coefficient (MIC) are integrated to choose appropriate spatial and temporal information in a comprehensive manner. Subsequently, to enhance the strengths of both deep and broad learning, a novel deep-broad ensemble learning architecture, whose weight parameters are optimized by the snow ablation optimizer (SAO) algorithm, is constructed to extract spatial-temporal features thoroughly and broadly. By this means, the prediction accuracy can be improved. Finally, experimental results on real dataset demonstrate that our proposed framework achieves better prediction accuracy than traditional methods.
1 July 2026 in IEEE Transactions on Sustainable Energy