Graph-temporal convolutional network for steam heating network simulation considering dynamic characteristics

Heating systems play a crucial role in achieving carbon neutrality, with steam heating networks being a key component of heating system. Accurate and effective modeling of steam heating networks is essential for ensuring the safety and reliability of the system, as well as improving energy efficiency. However, steady-state modeling methods neglect the dynamic characteristics of steam heating networks, leading to the oversight of fluid network time delays and pipe storage effects. Dynamic modeling methods have the problems of long computation time and the inability to solve complex topological structures. To address these problems, this study proposes a dynamic modeling method based on neural networks for node state modeling in steam heating networks. This method learns and captures the operational patterns of steam heating networks from both temporal and spatial dimensions, achieving higher prediction accuracy compared to existing dynamic modeling approaches. Furthermore, interpretability analysis of the initial node state distribution and a comparative study of different edge weights are conducted. The results demonstrate that employing an imputation layer and selecting appropriate edge weights both contribute to enhancing model accuracy.