Abstract: The access of multiple distributed sources and loads leads to enhanced voltage volatility in the distribution network.Meanwhile, the uncertainty fluctuation in the voltage of the upper main grid also affect the voltage characteristics of the distribution network. In order to effectively deal with the voltage fluctuations of the main grid and the distribution network, this paper proposes a multi-timescale voltage control framework for active distribution networks based on the combination of data-driven and model solving. In the slow time scale, considering the voltage fluctuation of the main grid, a multiple-feeder environment with a noninfinity system of the upper main grid is constructed, and the voltage control problem in this environment is modeled as an adversarial Markov process. During the training process, the voltage of the main grid is perturbed with a projected gradient descent algorithm. The Bayesian deep Q network algorithm is utilized to sense the voltage fluctuation of the upper main grid and realize the fast control of taps of the on-load tap changer. In the fast time scale, the reactive power output of the photovoltaic inverter is controlled based on the traditional second-order cone optimization method. The case results show that the method can accurately sense the voltage fluctuation of the upper main grid, realize model-free voltage control of the distribution network in a very short time, and ensure that the voltage of each node is maintained within the safety range.