@article{Hu2026, 
author = {Xiao Hu and Jinduo Yang and Yang Gao and Yiyan Li and Qingyuan Zhang and Jin Zhao},
title = {Model-free Multiple Energy Flow Calculation Using Edge-enhanced Heterogeneous Graph Attention Mechanism},
year = {2026},
journal = {CSEE Journal of Power and Energy Systems},
volume = {12},
number = {5},
pages = {2543-2554},
keywords = {Edge-enhanced heterogeneous graph attention, heterogeneous graph neural network, integrated energy systems, multi-task learning, multiple energy flow calculation, visualized interpretability},
url = {https://www.sciopen.com/article/10.17775/CSEEJPES.2025.07460},
doi = {10.17775/CSEEJPES.2025.07460},
abstract = {Multiple energy flow calculation (MEFC) serves as a fundamental technology within Integrated Energy Systems (IESs). Nevertheless, MEFC has faced increasing challenges due to the complex coupling among heterogeneous energy sources, information constraints, and extreme scenarios. This study proposes a data-driven approach for the MEFC task based on an edge-enhanced heterogeneous graph attention mechanism (EEHGAT). A heterogeneous graph structure is first employed to model the electricity-gas-heat coupled IES with three types of nodes and four types of edges. Then, a dual-model architecture comprising a node model and an edge model is constructed. This architecture aims to fully map MEFC information while simultaneously preventing error accumulation. Subsequently, a multi-task learning framework incorporating a variable-weight loss function is integrated to enhance the flexibility of the training process. Moreover, visualized interpretability is developed to demystify the “black–box” nature of the neural network. Numerical results demonstrate that the Mean Absolute Percentage Error (MAPE) for the proposed node and edge models is 0.1012% and 0.2060%, respectively. Meanwhile, under conditions such as measurement errors, information limitations, and extreme scenarios, the proposed method exhibits greater adaptability and robustness than the Newton-Raphson method and traditional data-driven methods.}
}