An electricity load forecasting model for Integrated Energy System based on BiGAN and transfer learning. (November 2020)

• Record Type: Journal Article
• Authors: Zhou, Dengji; Ma, Shixi; Hao, Jiarui; Han, Dong; Huang, Dawen; Yan, Siyun; Li, Taotao
• Is Part Of: Energy reports. Volume 6(2020)
• Page Start: 3446
• Page End: 3461
• Publication Date: 2020-11
• Subjects: Integrated energy system -- Deep learning -- Load forecast -- Bidirectional Generative Adversarial Networks -- Transfer learning; Power resources -- Periodicals; Energy industries -- Periodicals; Power resources; Periodicals; Electronic journals; 621.04205
• DOI: 10.1016/j.egyr.2020.12.010
• Languages: English
• ISSNs: 2352-4847

Abstract

Abstract: Integrated Energy System (IES) is able to collaborate various energy systems and boost energy supply efficiency. To further facilitate the energy scheduling in IES, load forecasting model of the system is required to describe the conditions continuously on a future time span. While the IES is a service model with frequent in-and-out users which are always dynamically changed, thus the dataset for some new users is always not enough sufficient to build the predicting model. Most of present researches focus on model refinement and accuracy boosting but rarely consider such data lack problem in IES. To tackle this issue, an integrated load forecasting model based on Bidirectional Generative Adversarial Networks (BiGAN) data augmentation and transfer learning techniques is proposed in this paper. Ten different types of data-driven models including the proposed model have been compared on two cases, resident and commercial users, in order to carry out the ablation and contrast experiment. Accuracy with the presented model is 2.08% and 1.50% higher than the original model averagely on resident and commercial users respectively, proving the effectiveness of the new model. And impact of sample size is analyzed and disclosed the effect patterns of the two modules. Result shows that the two modules can flexibly couple with different predictive models and boost their efficiency on both resident and commercial cases on data missing problem. And load forecasting becomes feasible for users with fewer samples or even zero samples when adopting the proposed framework. Highlights: An integrated forecasting model via BiGAN and transfer learning is proposed. This model is suitable for IES with in-and-out users, leading to insufficient historical data. 10 models coupled with proposed model are evaluated proving the effectiveness. Two different cases are carried out to prove the robustness of the model. Impact of sample size is analyzed on two modules to explore their features.