Short-Term Residential Load Forecasting with Baseline-Refinement Profiles and Bi-Attention Mechanism

• Published in: IEEE transactions on smart grid Vol. 15; no. 1; p. 1
• Main Authors: Xiao, Jiang-Wen, Liu, Peng, Fang, Hongliang, Liu, Xiao-Kang, Wang, Yan-Wu
• Format: Journal Article
• Language: English
• Published: Piscataway IEEE 01.01.2024; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: bi-attention mechanism; Convolutional neural networks; Cyclic loads; Electrical loads; Energy technology; Feature extraction; Forecasting; Load forecasting; Load modeling; Predictive models; Renewable energy; Renewable resources; Residential energy; Short term residential load forecasting; Smart grid; temporal convolutional network; Uncertainty
• ISSN: 1949-3053; 1949-3061
• DOI: 10.1109/TSG.2023.3290598

With the development of smart grid and renewable energy technologies, residential load forecasting has become an increasingly important task. Short-term residential load forecasting is not only conducive to power dispatching and peakshaving and valley filling of the grid, but also good for residents to obtain higher economic benefit from renewable energy. This paper proposes a new baseline-refinement forecasting framework consisting of two main steps, baseline profile construction and refinement predictions. Firstly, a baseline profile construction method is proposed to forecast the baseline load profile based on the similarity and cyclic patterns of daily load profiles. Secondly, for the refinement predictions, a bi-attention mechanism is proposed by combining self-attention mechanism and external attention mechanism to fulfill the feature transformation and is included in the temporal convolutional network to refine the baseline profile. The final load forecasting results are obtained by aggregating the baseline profile and the refinement predictions. The simulation results demonstrate that the proposed framework has smaller forecasting errors and higher forecasting stability than the commonly used models on two load forecasting metrics.