Abstract: Fully exploiting the reactive power support capability of distributed photovoltaic power supply is instrumental in addressing issues related to voltage fluctuation, voltage over-limit and new energy consumption in the distribution network caused by a high proportion of photovoltaic access. However, the reactive power output of the photovoltaic power supply will cause the junction temperature of its power device to exceed the limit or fluctuate sharply, which seriously threatens the reliable operation of the photovoltaic power supply. Therefore, this paper proposes a data-driven voltage/var optimization control strategy for the new energy distribution network considering the reliability of the photovoltaic power supply. First, a data-driven reliability evaluation method for photovoltaic power supply is proposed. This method uses the XGBoost machine learning model to calculate IGBT junction temperature, which improves the calculation efficiency of IGBT junction temperature and avoids the dependence of evaluation accuracy on IGBT parameters. Then, the voltage/var optimization model of the distribution network considering the reliability of the photovoltaic power supply is established, and the average junction temperature and junction temperature fluctuation of IGBT are introduced into the model optimization goal. Then the model is transformed into the Markov decision process, and the agent training is completed based on the deep deterministic strategy gradient algorithm. Finally, the advantages of the proposed strategy in ensuring the speed of reactive voltage optimization and improving the reliability of photovoltaic power supply are verified through simulations on the IEEE 33-bus system.