In order to improve the accuracy and computational efficiency of power system real-time state estimation, solve the problems of frequent, voltage fluctuations in grids and sharp increase in the uncertainty of power flow distribution in grids, a power system state estimation method based on Deep Neural Networks and approximately linear Proline Networks models were proposed, and its application in the power grids is researched. In this method, the mixed system measurement data was obtained through a particle filter algorithm to acquire a sample set, the training sample was used to train the proposed hybrid model, and finally the test samples were input into the established model to obtain the estimation result of the system state. Simulation result on the load data in IEEE118 bus system show that the power system state estimation method based on the proposed hybrid model not only allows rapid training for massive data, but also effectively avoids overfitting. The accuracy and computational efficiency of real-time state estimation compared with the Gauss-Newton method, are both improved. It can be seen that the proposed method has the application value in real-time state estimation of power systems.