Robust digital design of continuous-time nonlinear control systems using adaptive prediction and random-local-optimal NARMAX model

摘要

In this paper the time-delay and uncertainty of continuous-time (CT) systems are considered, and it is suggested that input and output of a discrete-time (DT) Neural Plant Model (NPM) and recursive neural controller have scaling factors which limit the value zone of measured data from a system. Adapted scaling factors cause the tuned parameters to converge to obtain a robust control performance. However, the proposed Random-Local-Optimization (RLO) design for a model/controller uses off-line initialization to obtain a near global optimal model/controller. Other important issues are the considerations of cost, greater flexibility, and highly reliable digital products for these control problems. This issue of DT control design for CT plant is more difficult than that of CT control design for CT plant, because of the need to process the modeling error between the CT plant and DT model. The input-delay, uncertainty, and sampling distortion of a CT nonlinear power system need to be solved by developing a digital model-based controller. Here, this is called the DT tracking control design of CT systems (DT-CT). Therefore, the DT structure of the adaptive controller for the CT nonlinear power system should be designed as a kind of feed-forward-Recursive-Predictive controller (FRP). First, due to the problem of delays, a digital neural controller with feed-forward of the reference signal and a Nonlinear Auto-Regressive Moving Average eXogenous (NARMAX) neural model design is adopted to reduce this difficulty. The most important contribution is that the more reasonable and systematic two-stage control design, the CT nonlinear delayed system to be controlled is modeled using a NARMAX technique with the first-stage (off-line) method by the proposed global optimal network algorithm and second-stage (on-line) adaptive steps. Second, the dynamic response of the system is controlled by an adaptive NARMAX neural controller via a sensitivity function. A theorizing method is then proposed to replace the sensitivity calculation, which reduces the calculation of Jacobin matrices of the BP method. Finally, the feed-forward input of reference signals helps the digital neural controller to improve the control performance, and the technique works to control the CT systems precisely.