Dynamic heterogeneous hypergraph attention convolutional network for critical node identification in distribution networks with high PV penetration

• High-order interactions are modeled for critical node identification. • Enhance adaptability via feature–structure co-evolution in dynamic topologies. • Improve robustness and interpretability with multi-dimensional evaluation system. The large-scale integration of distributed photovoltaic (PV) units into distribution networks has introduced challenges such as bidirectional power flows, dynamic reconfiguration, and stochastic fluctuations, thereby significantly increasing the complexity of system operation and control. Accurate identification of critical nodes thus becomes a fundamental prerequisite for effective supervision and control, and consequently for maintaining the stable operation of the power grid. To this end, this paper proposes an end-to-end identification framework based on a dynamic heterogeneous hypergraph attention convolutional network (DHHACN). Specifically, three categories of heterogeneous hyperedges are constructed, including voltage correlations, power interactions, and PV disturbance responses. This design overcomes the limitations of conventional binary graph structures and enables a more precise representation of multi-dimensional inter-node associations. A heterogeneous hypergraph convolution mechanism with hyperedge-type awareness is then designed to integrate multi-source features through differentiated message-passing processes. Furthermore, a feature-driven dynamic hypergraph updating strategy is introduced to achieve adaptive structural evolution. On this basis, a multi-dimensional evaluation system is developed, which combines four indicators, namely voltage coupling degree, composite voltage support index, power transfer sustainability, and PV disturbance resilience, together with an adaptive threshold determination algorithm. This design enables the quantitative assessment of node criticality without reliance on prior topological information. Experimental results on the modified 33-bus and 69-bus test systems demonstrate that the proposed approach exhibits strong applicability under high PV penetration conditions and superior stability compared with traditional methods under disturbance scenarios, offering valuable technical references for system stability and effective control.