A day-ahead prediction method for high-resolution electricity consumption in residential units. (15th February 2023)

• Record Type: Journal Article
• Authors: Liu, Che; Li, Fan; Zhang, Chenghui; Sun, Bo; Zhang, Guanguan
• Is Part Of: Energy. Volume 265(2023)
• Publication Date: 2023-02-15
• Subjects: Behavioral -- Convolution network -- Day-ahead household energy demand prediction -- Feature extraction -- High-resolution prediction -- Machine learning -- Temporal features; Power resources -- Periodicals; Power (Mechanics) -- Periodicals; Energy consumption -- Periodicals; 333.7905
• DOI: 10.1016/j.energy.2022.125999
• Languages: English
• ISSNs: 0360-5442

Abstract

Abstract: In order to accurately predict the day-ahead of household electric demand, a day-ahead high-resolution prediction model termed temporal-behavior coalescing forecast is proposed for the household energy demand. This model considers both behavioral and temporal dependencies. Feature extraction and data decomposition techniques are used to construct the behavioral and temporal inputs for the proposed model to reduce the negative impact of invalid data on its performance. In the proposed model the behavioral feature step and temporal feature step are constructed based on the convolutional network and long short-term memory network. Particularly, a coalescing step is designed in end the proposed model to strive for model convergence and enables the model to process two input matrices of different dimensions simultaneously. A 15-min resolution residential building energy demand dataset is used to validate the proposed model. The accuracy and generality of the proposed method are increased by 20.69% and 25.28%, respectively, compared with other related models. The validity of the proposed model is verified. In the robustness experimental, the proposed model can still maintain excellent prediction performance with the large noise introduced. A basis for its reproducibility and engineering application is provided. Highlights: A day-ahead prediction method is proposed for high-resolution household energy demand. The performance of prediction is improved by characterizing temporal and behavioral dependencies of the energy demand. A prediction model inputs construction approach is executed to make a positive effect on model convergence. The accuracy, generality, and robustness of the proposed method are validated and discussed.