Building’s electricity load prediction using representative data selection and stacking ensemble strategy

Building’s short-term electricity load forecasting is closely related to building energy management, power grid safety, and power plant scheduling. This paper proposes a lightweight prediction framework based on the representative data selection method and stacking ensemble strategy. By using kernel principal component analysis (KPCA), the dimension of input features is reduced, and the first principal component is extracted.to for K-means data clustering. The selected data samples closest to the cluster centers are treated as representative data for model training. An improved stacking type ensemble strategy is proposed. Three basic models and three deep learning models are serially integrated together, and their outputs are used for the training of a secondary learner (Ridge regression based on gradient boosting). From two cases studies, the proposed prediction framework could not only effectively improve the model training speed (saving about 49.62% of time), but also has better prediction accuracy and generalization performance.