The large number of layered multi-point accesses of wind and solar energy sources has brought a negative impact of voltage limit to the operation of a distribution network. It also increases the difficulty of reactive power collaborative optimization of different voltage levels. Considering the regulation requirements of different voltage levels, a multi-layer coordinated reactive power optimization method based on equivalent model is proposed from the perspective of dimension reduction equivalence of the model. Considering two station systems, one containing reactive power adjustment equipment and one not, this method uses a neural network to carry on the training to the massive station area correlation operation data to obtain the station area fitting model, forming two types of uncontrollable and controllable platform equivalent models. These two fitting models are used to replace the physical models of the two types of station areas, enabling the station areas to participate in the reactive power optimization of the distribution network in a data-driven manner. This forms a single voltage level object number mixed optimization scheduling model composed of the feeder physical model and the station areas fitting model. The original multi-layer reactive power optimization problem is transformed into a single-layer system optimization problem, and then the single-layer system optimization scheduling problem is solved. It can reduce the scale of optimization calculation for distribution network systems, and combine reactive power optimization mathematical models with data-driven methods to achieve multi-level coordinated reactive power optimization of distribution network feeders and station areas. The feasibility, effectiveness and superiority of the proposed method for solving the problem of voltage over-limit caused by distributed wind-photovoltaic access are verified by an example. This helps to ensure the economic, safe and stable operation of the distribution network.