Elsevier

Automatica

Volume 47, Issue 3, March 2011, Pages 539-544
Automatica

Brief paper
Feedback control of nonlinear stochastic systems for targeting a specified stationary probability density

https://doi.org/10.1016/j.automatica.2010.10.044Get rights and content

Abstract

In the present paper, an innovative procedure for designing the feedback control of multi-degree-of-freedom (MDOF) nonlinear stochastic systems to target a specified stationary probability density function (SPDF) is proposed based on the technique for obtaining the exact stationary solutions of the dissipated Hamiltonian systems. First, the control problem is formulated as a controlled, dissipated Hamiltonian system together with a target SPDF. Then the controlled forces are split into a conservative part and a dissipative part. The conservative control forces are designed to make the controlled system and the target SPDF have the same Hamiltonian structure (mainly the integrability and resonance). The dissipative control forces are determined so that the target SPDF is the exact stationary solution of the controlled system. Five cases, i.e., non-integrable Hamiltonian systems, integrable and non-resonant Hamiltonian systems, integrable and resonant Hamiltonian systems, partially integrable and non-resonant Hamiltonian systems, and partially integrable and resonant Hamiltonian systems, are treated respectively. A method for proving that the transient solution of the controlled system approaches the target SPDF as t is introduced. Finally, an example is given to illustrate the efficacy of the proposed design procedure.

Introduction

The control of stochastic systems for targeting specified probabilities or statistics has been studied for several decades. However, many well-established regulatory control techniques focus only on two quantities, i.e., mean and variance or covariance, as key design targets (Åström, 1970, Åström and Wittenmark, 1980, Goodwin and Sin, 1984, Lu and Skelton, 1998, Skelton et al., 1998, Wojtkiewicz and Bergman, 2001). Since the output of a nonlinear stochastic system is usually non-Gaussian, mean and variance or covariance are not enough to characterize the output process. Thus, in recent years, there has been a growing interest in the feedback control of nonlinear stochastic systems for targeting a probability density function (PDF). Karny (1996) proposed a randomized controller to minimize the mismatch between the joint PDF of a finite process realization and the target one. The controller depends on the solution of a functional equation, which is simpler than the general dynamic programming equation but still very complicated to obtain the solution. Guo and Wang (2005) proposed a pseudo proportional-integral-derivative (PID) tracking control strategy for general nonlinear stochastic systems, based on a linear B-spline model for the output PDFs. Yang, Guo, and Wang (2009) proposed a constrained proportional-integral (PI) tracking control for output probability distributions, based on two-step neural networks (NNs). Following the square-root B-spline NN approximation to the measured output PDF, the problem was transferred into the tracking of dynamic weights. Although the PID and PI tracking controls in Guo and Wang (2005) and Yang et al. (2009) are easy to implement and user-friendly, the B-spline approximation at each sample time is still of a high computational load. Wang (2003) used a nonlinear AutoRegressive and Moving Average with eXogenous (ARMAX) model with an arbitrary and bounded random input to represent general nonlinear stochastic dynamical systems and developed an optimal control to minimize the difference between the conditional output PDF and the target PDF. Guo, Wang, and Wang (2008) proposed an optimal control strategy for nonlinear ARMAX systems with channel time-delay and non-Gaussian noise to make the output PDF follow the target PDF and use a multi-step-ahead nonlinear cumulative cost function to improve the tracking performance. Crespo and Sun (2003) proposed a discontinuous non-linear feedback law to minimize the error between the output PDF and the target one for a one-dimensional stochastic continuous-time process model. Forbes et al., 2003a, Forbes et al., 2003b and Forbes, Guay, and Forbes (2004a) developed regulatory control synthesis techniques for shaping the PDF of stochastic processes. The regulatory control synthesis techniques include a 9-step algorithm, integrating PDF-shaping into a comprehensive control design procedure.

It should be noted that all the design techniques proposed in Crespo and Sun (2003), Forbes et al., 2003a, Forbes et al., 2003b, Forbes et al., 2004a, Guo and Wang (2005), Guo et al. (2008), Karny (1996), Wang (2003) and Yang et al. (2009) are approximate. The disadvantages of these design techniques are: (1) there is no guarantee that the output PDFs of the system will be exactly the same as a given one; (2) the explicit expressions of the feedback controllers are difficult to obtain and the computational loads are very high; (3) only low-dimensional stochastic process models or some typical systems were presented in the references.

An ideal approach to control the design of nonlinear stochastic systems for targeting a specified PDF is based on the exact solution of the systems. The output of a nonlinear system to Gaussian white noise is a diffusion process and the transition PDF of the output is governed by the Fokker–Planck–Kolmogorov (FPK) equation. The exact transient solutions of the FPK equation have been obtained only for very special one-dimensional nonlinear stochastic systems. The exact stationary solutions of the FPK equation, however, have been obtained for a variety of nonlinear stochastic systems (Lin and Cai, 1995, Zhu, 2006). Among them, five classes of exact stationary solutions have been obtained for the dissipated MDOF Hamiltonian systems (Zhu et al., 1990, Zhu and Yang, 1996). All these exact stationary solutions can be treated as the basis for designing feedback control of nonlinear stochastic systems to target a specified SPDF. However, so far only the exact stationary solution of a one-dimensional nonlinear stochastic system has been used for designing the feedback control of a system to target a specified SPDF (Forbes, Guay, & Forbes, 2004b).

In the present paper, an innovative procedure for designing the feedback control of MDOF nonlinear stochastic systems to target a specified SPDF is proposed, based on the technique for obtaining the exact stationary solutions of the dissipated MDOF Hamiltonian systems. This paper is organized as follows: In Section 2, the technique for obtaining the five classes of exact stationary solutions of the dissipated MDOF Hamiltonian systems is briefly reviewed. Then, the design procedure for the feedback control of MDOF nonlinear stochastic systems to target a specified SPDF is proposed in Section 3. In Section 4, a method for proving that the transient output of the controlled system approaches the target SPDF as t is presented. In Section 5, an example is worked out to illustrate the design procedure and its effectiveness. Finally, conclusions and suggestions for future studies are given in Section 6.

Section snippets

Exact stationary solution

Consider a MDOF nonlinear stochastic system expressed by the following dissipated Hamilton equations Q̇i=HPiṖi=HQicij(Q,P)HPj+fik(Q,P)Wk(t)i,j=1,2,,n;k=1,2,,m where Qi and Pi are generalized displacements and generalized momenta, respectively; Q=[Q1,Q2,,Qn]T, P=[P1,P2,,Pn]T; H=H(Q,P) is a Hamiltonian; cij(Q,P) are coefficients of damping; fik(Q,P) are amplitudes of random excitations; Wk(t) are Gaussian white noises in the sense of Stratonovich with correlation functions E[Wk(t

Design of control for targeting a specified SPDF

Consider a controlled MDOF nonlinear stochastic system. Suppose that it can be expressed as the following controlled, dissipated Hamiltonian system Q̇i=HPiṖi=HQicij(Q,P)HPjui(Q,P)+fik(Q,P)Wk(t)i,j=1,2,,n;k=1,2,,m where ui=ui(Q,P) are feedback control forces to be designed and the other notations are the same as those in Eq. (1). A derivation of Eq. (6) as that from Eq. (1) to Eq. (4) leads to the following Itô stochastic differential equations dQi=HPidtdPi=[HQi+mij(Q,P)HPj+

Output of the controlled system

Substituting ui(1) and ui(2) determined in the last section into Eq. (7) leads to the following Itô stochastic differential equations of the controlled system dQi=H̄PidtdPi=[H̄Qi+mij(Q,P)H̄Pj+ui(2)(Q,P)]dt+σik(Q,P)dBk(t)i,j=1,2,,n;k=1,2,,m. It is readily proved by using the technique described in Section 2 that system (15) has the same exact stationary solution as the target SPDF in Eq. (8). [QT,PT]T is a vector diffusion process with the elliptic differential generator L=H̄piqi[

Example

Consider a 4-DOF non-linear stochastic system governed by the following Itô stochastic differential equations dQi=PidtdPi=(ωi2Qi+diPi+ui)dt+dBi(t)+QldBl(t)di=ci0+ci1H1+ci2H2+ci3H3+ci4H4ω12=1,ω22=1.41,ω32=1.21,ω42=0.49i=1,,4;l=4,,1 where cij(i=1,,4;j=0,,4) are constants, and Bi(t),Bl(t) are independent Wiener processes with intensities 2Dii,2Sii, respectively. The Hamiltonian and first integrals of the Hamiltonian system associated with the uncontrolled system (17) are H=i=14HiHi=(pi2+ωi2

Concluding remarks

In the present paper the technique for obtaining the exact stationary solutions of MDOF nonlinear stochastic systems has been applied for the first time to design feedback control of MDOF nonlinear stochastic systems for targeting a pre-specified SPDF. The proposed control procedure has a number of advantages: it is simple yet yields an exactly analytical control law; the control can be designed offline and no online measurement is required; the physical implications of the feedback control are

C.X. Zhu is currently a Ph.D. candidate at the School of Mechanics, Civil Engineering and Architecture, Northwestern Polytechnical University. She received her B.S. degree from the same university in 2007. She has worked in the Department of Mechanics in Zhejiang University for two years since 2008. Her research interests include stochastic optimal control and process control.

References (21)

There are more references available in the full text version of this article.

Cited by (24)

  • Decentralized PI tracking control for non-Gaussian large-scale interconnected distribution systems

    2016, Journal of the Franklin Institute
    Citation Excerpt :

    In [5] the first-order stochastic distribution process was discussed, while the linear model with PID controller was studied in [6]. The tracking control was proposed in [7] for a specified stationary density function. The discrete-time PI controller in SDC and the two-step intelligent modeling method were established in [8,9], respectively.

  • Robust tracking controller design for non-Gaussian singular uncertainty stochastic distribution systems

    2014, Automatica
    Citation Excerpt :

    In these cases, variance is not sufficient to characterize the statistical properties of random output. Motivated by this problem, some new approaches called stochastic distribution control (SDC), such as the output SDC (Wang, 2000), the system SDC or fully probabilistic control design (Herzallah & Karny, 2011; Karny, 1996; Karny & Kroupa, 2012) and the process SDC (Annunziato & Borzi, 2013; Forbes & Forbes, 2004; Pigeon, Perrier, & Srinivasan, 2011; Zhu & Zhu, 2011; Zhu, Zhu, & Yang, 2012), were proposed which directly control the probability density function (PDF) of the concerned. Those types of stochastic control systems are much more general than the traditional ones and can therefore be applied to either Gaussian or non-Gaussian dynamic systems.

View all citing articles on Scopus

C.X. Zhu is currently a Ph.D. candidate at the School of Mechanics, Civil Engineering and Architecture, Northwestern Polytechnical University. She received her B.S. degree from the same university in 2007. She has worked in the Department of Mechanics in Zhejiang University for two years since 2008. Her research interests include stochastic optimal control and process control.

W.Q. Zhu is a professor of Zhejiang University and a member of Chinese Academy of Sciences. He graduated from Northwestern Polytechnical University in 1961 and the Graduate School of the same university in 1964. In the last two decades, he visited University of Wisconsin-Madison, MIT, Florida Atlantic University and State University of New York at Buffalo, in USA, University Blaise Pascal in France, Kyoto University in Japan, etc. His research interest is in nonlinear stochastic dynamics and control, and he has proposed and developed a Hamiltonian theoretical framework of nonlinear stochastic dynamics and control.

The work reported in this paper is supported by the National Natural Science Foundation of China under Grant No. 10772159, 11072212 and No. 10932009, the Zhejiang Natural Science Foundation of China under Grant No. Y7080070 and the Research Fund of Northwestern Polytechnical University under Grant No. JC200802 and JC200937. The material in this paper was not presented at any conference. This paper was recommended for publication in revised form by Associate Editor George Yin under the direction of Editor Ian R. Petersen.

View full text