Multi-energy flow calculation in integrated energy system via topological graph attention convolutional network with transfer learning

Abstract

Multi-energy flow (MEF) calculation plays a key role in integrated energy system (IES) analysis. However, its practical implementation is challenged by the high computational burden and frequently changing topology. To mitigate these issues, the topological graph attention convolutional network with transfer learning (TGACN-TL) is proposed for MEF calculation in IES with electrical, natural gas, and heating networks. Specifically, the topological physics information is embedded in TGACN to improve the MEF calculation accuracy. Besides, the attention mechanism is leveraged by TGACN to capture and represent intricate graphical patterns inherent in the MEF data, thereby preserving essential graphical structure features. Furthermore, transfer learning is applied to utilize previously learned topological knowledge, facilitating adaptation to new electricity network topologies through updating partial model's parameters. The simulations demonstrate that the proposed TGACN with transfer learning achieves superior MEF calculation accuracy, maintaining robust performance across diverse conditions of uncertainty and topological variations.