The invention discloses an optimal control method based on a data-driven single network structure. By continuously updating training samples and adding to an LS-SVM model, the partial derivative of a system state variable x for the optimal performance index function is approximated on line, and then the optimal control law is directly obtained by calculation according to the optimal control law expression. In this way, the controller network is saved, so that the structure becomes simpler. In addition, the calculation procedures involved in each iteration of the traditional computing method are reduced, thereby increasing the execution speed.