As the core of optimal dispatching decision-making for distribution networks，optimal power flow （OPF） urgently requires fast calculation methods under large-scale grid frameworks.A deep learning-based OPF solving method oriented toward feasibility restoration was proposed.First，an OPF solving architecture based on state-control variable decomposition was constructed，and an OPF state variable solving model was built based on a deep neural network.Second，to address the problem that OPF results based on deep neural networks fail to satisfy control variable constraints，samples with inequality constraint violations were screened to construct a correction sample set；considering actual physical constraints and supply-demand balance relationships，joint correction constraint conditions based on overall control variables were proposed to establish correction intervals.Finally，the Sinkhorn algorithm based on multi-marginal distributionswas used to adjust control variable solutions，and constraint-violating variables were iteratively projected into correction intervals to meet actual physical constraint conditions.The proposed method is verified based on an improved IEEE 123-node distribution network case.Experimental results show that the proposed method can effectively achieve feasibility restoration of control variables，balance the mean absolute error of each control variable，and improve solution accuracy.