The integration of large-scale distributed renewable energy sources brings new challenges to the active distribution network (ADN), including the three-phase imbalance problem, unexpected voltage violations and increased line losses. However, due to the incomplete installation of measurement equipment in the current distribution network, it is difficult to accurately obtain the load data of some nodes. Therefore, the traditional ADN voltage control method based on global observation is difficult to meet the actual control requirements. To solve these problems, a fast Volt/Var control (VVC) evolutionary algorithm with a deep learning surrogate model is proposed. In the development of the algorithm, a highway neural network is first designed as the surrogate model to accurately fit the mapping relationship between limited measured load information, voltage regulation control strategy and system performance indices. Then, the trained surrogate model is embedded into the iterative optimization process of the non-dominated sorting genetic algorithm, and the voltage deviation rate, three-phase unbalance degree and line losses indicators are directly calculated, and the data-driven distribution network VVC strategy can be quickly obtained. A modified IEEE 123-node three-phase distribution network is employed to verify the advantages and efficiency of the proposed algorithm. This work is supported by the National Natural Science Foundation of China (No. 51977133).