PID神经网络改进研究

The Study of Momentum PID Neural Network

PID神经网络(PIDNN)是将神经网络和PID控制规律融为一体的新型前向神经元网络。对于多数系统,PIDNN可以利用已有PID控制规律的先验知识确定网络权重初值,使系统得到稳定快速的控制。当权重初值选择为随机数时,收敛速度变慢,同时可能陷入局部最小。针对这一类系统,提出附加动量项的改进算法,克服权重初值取随机数带来的问题。系统仿真结果证明改进后的PIDNN系统性能得到了明显改善。

PID neural network is a new type of feed-forward neural network, in which neural networks integrate with PID algorithm. For most systems, the apriority knowledge of PID algorithm can be used to choose the initial weights, so that system has been steady and rapid control. Without the apriority knowledge, initial weights usually use the random number, and the convergence slows down, while likely to fall into local minimum. The paper proposes an improved algorithm formula which has momentum coefficients to overcome the problems caused by the random initial weights. The distinct improvement of the PIDNN control system is proved by simulation results.