Research on future trends of electricity consumption based on conditional generative adversarial network considering dual‐carbon target

Abstract

The emergence of novel factors, such as the energy Internet and electricity supply‐side reform within the context of the dual‐carbon target (carbon peaking and carbon neutrality), has heightened the uncertainty surrounding electricity consumption (EC). This increased uncertainty poses challenges for accurate long‐term EC forecasting. Due to the complexities of feature extraction and the absence of labelled data, conventional supervised learning‐based forecasting methods, such as support vector machines (SVM) and long short‐term memory networks (LSTM), struggle to predict EC with precision in situations of heightened uncertainty resulting from the interplay of multiple factors. To address this issue, a novel method based on a conditional generative adversarial network (CGAN) is proposed. Initially, the dominant factors influencing future electricity consumption trends through grey correlation degree analysis and the K‐L information method are identified. Subsequently, an EC forecast model is introduced based on CGAN, adept at capturing essential factors and the non‐linear relationship between EC and exogenous factors. This approach effectively models the uncertainty of EC, accurately approximating the true distribution with only a small dataset. Finally, the proposed method by forecasting China's EC from 2015 to 2020 is validated. The results demonstrate that the authors’ method achieves lower root mean square error and mean absolute percentage error values, specifically 0.177% and 2.39%, respectively, outperforming established advanced methods such as SVM and LSTM.