Selective Heterogeneous Ensemble Method Based on Local Learning and Evolutionary Multi-objective Optimization for Wind Power Forecasting

To solve the problems of strong randomness and fluctuation in wind power forecasting, a novel wind power forecasting model using selective the heterogeneous ensemble (SHeE) method based on local learning and evolutionary multi-objective optimization (EMO) is proposed. First, Lasso regression is used to feature selection to remove irrelevant variables and redundancy. Then, a set of sample sets is constructed via a novel K-nearest neighbor (KNN)-based clustering method to enrich ensemble diversity from the perspective of data. To further inspire the diversity, three modeling methods, support vector regression (SVR), Gaussian process regression (GPR), and partial least squares (PLS) are applied to build the heterogeneous model library. Subsequently, ensemble pruning is performed by EMO. When a new sample arrives, the ensemble prediction result can be obtained by simple averaging. Finally, the effectiveness and superiority of the proposed method are demonstrated through a real wind farm dataset.