Non-solution power flow diagnosis method for AC/DC hybrid power grid based on topology-aware graph neural network

With the continuous evolution of power grids, the issue of power flow unsolvability in hybrid AC–DC grids, arising from topological changes, abrupt load variations, or control failures, has become increasingly prominent. To overcome the black-box nature inherent in conventional artificial intelligence methods, this work explicitly models identifiable manifestations of physical constraints through interpretable graph neural network (GNN) layers. Consequently, this paper develops a data-driven and physics-model-fused diagnostic approach: the topology-aware graph neural network (TAGNN). A dual diagnostic strategy based on DC-side harmonic injection and AC-side voltage sensitivity is proposed. This establishes explicit mappings between topological characteristics and power flow feasible regions, thereby revealing the intrinsic relationship between grid topological structures and unsolvable power flow conditions. The approach addresses the challenge of solving multivariate nonlinear equations within coupled AC–DC Supervisory Control and Data Acquisition systems. The simulation experiments, designed to replicate grid fault scenarios, demonstrate the superior performance enhancement achieved by integrating the topology attention layer with the gated recurrent unit within the TAGNN framework. Compared against the current mainstream benchmark methods, the proposed model yields an average improvement exceeding 10% across key test metrics. These results validate the model’s capability to establish standardized diagnostic procedures and provide a practical utility tool for grid planning and operational analysis within power system computer aided design environments.