一种改进的近似动态规划方法及其在SVC的应用

An improved approximate dynamic programming and its application in SVC control

近似动态规划的基本思想是通过近似计算代价函数,从而避免动态规划中的"维数灾"问题。随机选取初值使得近似动态规划方法需要多次的学习才能最终收敛,极大地限制了在实际系统中的应用。针对上述问题,提出一种基于改进PID神经网络的直接启发式动态规划算法,将初始执行网络与PID控制器之间建立起一种等价关系,因此可以利用已经设计好的PID控制器来指导其初值选取,从而使算法收敛性大大提高。改进的神经网络与常规PID神经网络相比,结构简单且具有更好的扩展性,性能上具有更强的鲁棒性。对4机2区系统的静止无功补偿器附加阻尼控制进行仿真测试,仿真结果表明基于改进PID神经网络的直接启发式动态规划算法和初值选取方法的有效性,并且在部分状态反馈和延时两种情况下有着很好的控制效果。

The main idea of approximate dynamic programming（ADP） is approximately computing cost function to avoid the curse of dimension.However,it needs many times learning to converge due to the randomly choosing initial weights.So it is greatly limited in the application.This paper presents a direct heuristic dynamic programming（DHDP） based on an improved proportion integration differentiation PID neural network（IPIDNN）.This method constructs an equivalent between the initial action network and PID controller.Therefore,well-designed PID controller can guide the initial weights choosing,so that the convergence of this algorithm will be remarkably improved.Moreover,compared with the traditional PID neural network,the configuration of IPIDNN is flexible and easy to expand,as well as a better robust performance.The simulation results show the validity of this algorithm and initial weights choosing method by the static var compensator（SVC） supplementary control in four-machine two-area system.It also has a good performance in the circumstance of partial state feedback and state delay.