A Physics-Informed Graph Attention-based Approach for Power Flow Analysis

Risk-informed management of power grids requires accurate and computationally efficient power flow analysis. However, existing methods for solving power flow problems are computationally inefficient considering the many simulations needed to quantify uncertainties in system performance. This work presents a novel physics-informed graph attention-based method for power flow analysis in power transmission systems. We employ a graph attention network (GAT) based architecture which leverages the self-attention mechanism. As a result, structural information of a graph is learned and utilized to implicitly consider the importance of nodes in the graph. Through the integration of the GAT model, the power flow analysis is efficient and applicable to inductive learning problems where the model has to generalize to a priori unseen power grids. Furthermore, the physics-based knowledge of the power flow analysis is directly implemented by enforcing minimization of the violation of Kirchhoff’s law at each bus during training. To illustrate the performance of the proposed model, well-known IEEE power grid testbeds, namely, case9, case14, case30, and case118 are selected and the graph attention-based model is tested and compared with state-of-the-art methods. The result of these analyses indicates the efficacy of the physics-informed graph attention-based approach in achieving a superior accuracy and less computational cost. Furthermore, the robustness of the proposed model to the variations in power grid topology is demonstrated. Therefore, it shows a reliable performance in inductive learning problems.