Modeling of Gas Turbines via a Knowledge-Embedded Condition-Aware Dynamic Spatiotemporal Graph Convolutional Network

High-precision gas turbine (GT) modeling is essential for the tasks of load control and health management. Low-calorific-value GTs exhibit strong multivariable nonlinear couplings and frequent operating condition shifts induced by variations in fuel calorific value, flow rate, and pressure, which make accurate modeling particularly challenging. This article proposes a GT modeling method via a knowledge-embedded condition-aware dynamic spatiotemporal graph convolutional network (KCA-DSTGCN). In this method, an adaptive graph learning (AGL) module combines a thermodynamics-guided static graph and an online dynamic graph. Subgraph extraction for the static graph handles heterogeneous sensor configurations, while the dynamic graph employs maximal information coefficient (MIC)-based correlation partitioning combined with change-point detection to capture regime transitions and update inter-variable dependencies in real time, enabling accurate tracking of evolving operating conditions. On these bases, a dual-attention spatiotemporal (DAST) module is developed to tackle the differing dependencies of node features on their historical and neighboring information. Furthermore, to account for the heterogeneous temporal responses of variables to load changes, multi-scale features are extracted with varied dilation factors and fused through a multi-scale feature fusion (MSFF) module with a linear layer. The performance of the proposed method is validated on real-world data from two representative types of GTs with high and low calorific values, and compared against state-of-the-art algorithms. Results demonstrate that it significantly outperforms existing methods in modeling accuracy.