Abstract: With the rapid deployment of 5G mobile communication, the construction of numerous 5G base stations offers significant flexibility resources for the power system. By utilizing the spare stored energy of these 5G base stations as schedulable energy storage resources, the adverse effects of wind power generation's randomness and volatility on the system can be mitigated. The optimization scheduling problem of active distribution networks containing distributed wind power generation is the focus of this article. Firstly, a base station reliability evaluation model is established to systematically evaluate the real-time schedulable capacity of the base station energy storage based on two factors: the potential power interruption and the power outage recovery time. Furthermore, an improved Twin Delayed Deep Deterministic policy gradient (TD3) algorithm, based on a Variational Autoencoder (VAE) model, is utilized to minimize the system operation cost by solving the optimal charging and discharging strategy of the 5G base station energy storage. The energy storage status of multiple base stations is represented in the form of hidden variables to mine the hidden associations in the data, thereby reducing the model solving complexity and improving the algorithm performance. By iteratively solving to converging, the multi-base station energy storage(MBSES) system achieves continuous action control, and the personalized charging and discharging strategies tailored to the actual operating conditions are developed for each base station. Finally, the effectiveness of the proposed method is verified through numerical examples.