A TCN‐Transformer Model With Similar Period Selection and Hierarchical Decomposition for Short‐Term Net Load Forecasting

ABSTRACT

The growing penetration of renewable energy in modern power systems has driven a paradigm shift from conventional load forecasting to net load forecasting (NLF). Improving short‐term NLF accuracy and robustness is essential for secure and economical system operation, efficient renewable energy accommodation and informed electricity market participation. This paper proposes an NLF method based on a temporal convolutional network (TCN)‐Transformer architecture, incorporating similar period selection and hierarchical decomposition to improve forecasting accuracy. First, a similar period selection strategy is employed to filter the historical dataset, thereby minimising the influence of irrelevant or redundant data on the forecasting model. Then, a hierarchical decomposition scheme is proposed to better capture and disentangle the complex temporal features embedded in the net load. The TCN‐Transformer model is employed to separately forecast the decomposed components and the results are aggregated to obtain the final NLF results. Case studies on real‐world regional net load data demonstrate the superior accuracy and robustness of the proposed method.