Image recognition-based PV hosting capacity enhancement via distribution network retrofit

The assessment and enhancement of photovoltaic (PV) hosting capacity using roof area analytics and classical optimization approaches face several challenges, like low accuracy and poor credibility. This paper proposes a novel image recognition-based hybrid optimization method to assess the allowable PV integration capacity and determine the cost-effective planning strategy for network retrofit. Specifically, it improves the geometry-aware YOLOv11 framework by a dual-phased regularization to precisely extract rooftop surface areas, thereby establishing spatially-constrained PV capacity boundaries. Then, by introducing transformer insulation aging formulas and security chance constraints, a probabilistic optimization model is developed to quantify the PV hosting capacity under varying risk tolerance levels. Additionally, the infrastructure capacity expansion is formulated as an integer decision variable and incorporated into this model to economically retrofit the distribution network for foreseeable PV-rich scenarios. Furthermore, to decrease computational complexity, the model is relaxed using second-order cone programming (SOCP) and decomposed into a master-subproblem structure via Benders Decomposition (BD). The simulation results demonstrate that this method enables precise identification and full exploitation of PV hosting capacity.