A dynamic weighted deep-broad learning architecture for wind power interval prediction

Predicting wind power is essential for the stable and economical operation of power grids. Traditional point forecasts often struggle to capture volatility, while wind power interval prediction (WPIP) effectively quantifies associated uncertainties, providing comprehensive information for power dispatch. This paper proposes a novel deep-broad WPIP ensemble framework. First, considering the relationships among variables collected by the wind turbine’s supervisory control and data acquisition system, an operational mechanism-based method is designed to identify the outliers, while a novel multiple imputation approach is proposed to fill in the missing values and rectify the detected outliers to guarantee good data quality. Then, based on fuzzy and rough set principles, rolling fuzzy information granulation is used to capture the wind power fluctuation and construct high-quality training labels. Subsequently, the graph attention network with dynamic attention and gated recurrent unit, and the stacked broad learning system are developed separately as wind power interval predictors. These two models can extract and fuse spatial–temporal features selected from the training data thoroughly and efficiently to enhance the WPIP accuracy. Moreover, the human evolutionary optimization algorithm is used to dynamically optimize the combined weights online and integrate the prediction results of these two predictors to improve the interval prediction performance effectively. The comparative and ablation experiments demonstrate that our model outperforms the benchmark methods.