Daily Electric Load Forecasting with Transformer via Multi-Source Attention and Trainable Weight Mechanisms
Year: 2025
Type: conference-paper
Abstract: Load forecasting is a key component in power grid operation and energy management. To improve prediction accuracy, it is essential to fully leverage multi-source information beyond historical load data. Addressing the limitations of existing methods, which often overlook the correlations and dynamic importance among different sources during multi-source feature fusion, this paper proposes a multi-branch Transformer forecasting model that integrates multi-source attention and a trainable weighting mechanism. Specifically, independent Transformer encoders are constructed for historical load and each auxiliary feature to extract temporal characteristics. Then, using the load features as queries, the multi-source attention mechanism selectively captures relevant information from other features. In the fusion stage, trainable feature weights are introduced to dynamically allocate the contribution of each feature according to the input, thereby achieving effective information integration. Finally, the fused representation is combined with the load features and fed into a multilayer perceptron to generate the prediction results. Experiments based on real-world data demonstrate that the proposed model outperforms various benchmark methods in both prediction accuracy and robustness, while also providing interpretable insights into the role of each feature in the forecasting process.
Authors: Fangmei Bie, Zhigang Zhu, Haoyu Wang, Huihui Li, Heng Ding, Runhua Wang, Qing Liu, Liang Qin
Subfield: Electrical and Electronic Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W4417494680