Data augmentation strategy for small sample short‐term load forecasting of distribution transformer

Abstract

Short-term load forecasting at the distribution transformer level provides a basis for demand-side aggregators to take part in the power market. However, under the competitive market environment, certain parties might be discriminated and do not have access to enough data and thus the challenge of load forecasting under limited dataset arises. To tackle the small sample forecasting problem at the distribution transformer level, this paper first explores the load unbalance phenomenon and correlation among three-phase electrical loads. Then, the data augmentation strategy for learning models is proposed. The strategy includes three-phase datasets fusion and rolling forecast. The aim is to ameliorate overfitting and make sure the distribution of the training and test data as close as possible. The strategy is validated by the realistic data of the distribution transformers from South and East China. Comprehensive case studies demonstrate that, by utilizing the proposed strategy, learning models can achieve better forecast accuracy and the generalization ability is improved. Also, the proposed strategy is proved to boost robustness against the measurement uncertainty.