With the increasing penetration of distributed photovoltaic energy, the inherent three-phase unbalanced problem of a distribution network is becoming more serious. This brings adverse effects on power quality and economic operation of the system. In addition, a high proportion of photovoltaic leads to a more complex physical structure and operation mode of the distribution network, resulting in it being difficult to apply the current three-phase unbalanced optimization method that relies on precise topology and line parameters. Therefore, this paper proposes a three-phase unbalanced optimization method of a distribution network with a high proportion of photovoltaic based on a data-driven power flow model. First, a dual-stage attention-based recurrent neural network is used to establish the data-driven power flow model, and the functional relationship between the variables in the three-phase power flow constraint is fitted. At the same time, a graph feature embedding method is proposed to embed the partially known topology information into the model to improve the fitting accuracy. Secondly, the three-phase unbalanced optimization model is reconstructed based on a trained data-driven power flow model. Finally, the conditional gradient descent method is used to analyze the model, and a modified IEEE 33-node distribution network is taken as an example to verify the effectiveness of the proposed method.