Short-Term Building Electrical Load Prediction by Peak Data Clustering and Transfer Learning Strategy

With the gradual penetration of new energy generation and storage to the building side, the short-term prediction of building power demand plays an increasingly important role in peak demand response and energy supply/demand balance. The low occurring frequency of peak electrical loads in buildings leads to insufficient data sampling for model training, which is currently an important factor affecting the performance of short-term electrical load prediction. To address this issue, by using peak data clustering and knowledge transfer from similar buildings, a short-term electrical load forecasting method is proposed. First, a building’s electrical peak loads are clustered through peak/valley data analysis and K-nearest neighbors categorization method, thereby addressing the challenge of data clustering in data-sparse scenarios. Second, for peak/valley data clusters, an instance-based transfer learning (IBTL) strategy is used to transfer similar data from multi-source domains to enhance the target prediction’s accuracy. During the process, a two-stage similar data selection strategy is applied based on Wasserstein distance and locality sensitive hashing. An IBTL strategy, iTrAdaboost-Elman, is designed to construct the predictive model. The performance of proposed method is validated on a public dataset. Results show that the data clustering and transfer learning method reduces the error by 49.22% (MAE) compared to the Elman model. Compared to the same transfer learning model without data clustering, the proposed approach also achieves higher prediction accuracy (1.96% vs. 2.63%, MAPE). The proposed method is also applied to forecast hourly/daily power demands of two real campus buildings in the USA and China, respectively. The effects of data clustering and knowledge transfer are both analyzed and compared in detail.