With the promotion of the "dual carbon" goal and the construction of new power systems, the complexity and uncertainty of power systems have increased dramatically, and the distribution network is faced with the challenges of high proportion of distributed energy and asymmetric loads, such as voltage overstep and three-phase imbalance. To cope with these problems, put forward based on robust stochastic optimization (RSO) of low voltage distribution network optimal power flow (OPF) local control strategy. The strategy uses a 1-norm Wasserstein distance uncertainty set to describe distributed energy resource (DER) output uncertainty, the construction of a three-phase four-wire low-voltage distribution network robust stochastic optimization model, designed to minimize control costs network losses and take into account the worst-case expected adjustments. Convolutional neural networks (CNN) training to obtain local control strategies for distributed energy without communication infrastructure. The simulation results show that the proposed control strategy is effective and economical.