Nonlinear Model Predictive Control with Probabilistic Models
نویسندگان
چکیده
Nonlinear Model Predictive Control (NMPC) is a powerful control framework, which strongly relies on a good model of the system dynamics. In the case, such a model is not available apriori, non-parametric regression using Bayesian regression or Gaussian Processes (GPs) have been shown promising in inferring the dynamics from collected data. An advantage of Bayesian methods and GPs over other regression methods is the availability of a predictive distribution expressing the uncertainty about the true function induced by the finite amount of observations. Although recent work indicates that propagation of this uncertainty can be used to design robust controllers, it has not been considered in NMPC yet. This thesis presents an approach to robust Semi-Implicit NMPC of Bayesian linear models and Gaussian Process dynamics subject to control constraints. The propagation of the uncertainty is done by means of the Moment-Matching (MM) technique to track the central moments of the state distribution and a recent approximation framework is used for fast online NMPC. Although the approach has several advantages from a theoretical perspective, its performance on a highly nonlinear benchmark system is worse than expected. Several possible sources for the degradation are investigated and discussed. Zusammenfassung Nonlinear Model Predictive Control (NMPC) ist eine mächtige Methode der Steuerung und Regelung. Diese setzt allerdings ein gutes Modell des Systemverhaltens voraus. Wenn ein solches Modell nicht verfügbar ist, können Verfahren der nicht-parametrischen Regressionsschätzung, besonders Gaussian Processes (GPs), erfolgreich eingesetzt werden um das Systemverhalten aus Beobachtungen zu schätzen. Bayes’sche Modelle und Gaussian Processes zeichnen sich gegenüber anderen Methoden dadurch aus, dass sie die Möglichkeit bieten mittels einer Wahrscheinlichkeitsverteilung die Unsicherheit in der Vorhersage, bei einer möglicherweise geringen Anzahl von Beobachtungen, zu quantifizieren. Obwohl in aktuellen Veröffentlichungen bereits gezeigt werden konnte, dass mithilfe der Fortpflanzung dieser Verteilung über Zeit robuste Regler entworfen werden können, wurde diese Methode bisher noch nicht in NMPC verwendet. In dieser Arbeit wird ein Ansatz zur Semi-Impliziten NMPC von Bayes’schen lineare Modellen sowie Gaussian Processes bei zusätzlichen Steuerungsbeschränkungen vorgestellt. Die Unsicherheit wird hierbei durch die Moment Matching (MM) Technik fortgesetzt, während die Online-Steuerung durch ein modernes Approximationsverfahren erfolgt. Obwohl der vorgestellte Algorithmus von theoretischer Seite vielversprechend ist, zeigte sich die tatsächliche Regelungsqualität auf einem nichtlinearen Testsystem als schlechter als erwartet. Die möglichen Ursachen für das schlechte Abschneiden werden deshalb dargestellt und diskutiert.
منابع مشابه
مدل ترکیبی تحلیل مؤلفه اصلی احتمالاتی بانظارت در چارچوب کاهش بعد بدون اتلاف برای شناسایی چهره
In this paper, we first proposed the supervised version of probabilistic principal component analysis mixture model. Then, we consider a learning predictive model with projection penalties, as an approach for dimensionality reduction without loss of information for face recognition. In the proposed method, first a local linear underlying manifold of data samples is obtained using the supervised...
متن کاملPredictive control with Gaussian process models
This paper describes model-based predictive control based on Gaussian processes. Gaussian process models provide a probabilistic nonparametric modelling approach for black-box identification of non-linear dynamic systems. It offers more insight in variance of obtained model response, as well as fewer parameters to determine than other models. The Gaussian processes can highlight areas of the in...
متن کاملWhich Methodology is Better for Combining Linear and Nonlinear Models for Time Series Forecasting?
Both theoretical and empirical findings have suggested that combining different models can be an effective way to improve the predictive performance of each individual model. It is especially occurred when the models in the ensemble are quite different. Hybrid techniques that decompose a time series into its linear and nonlinear components are one of the most important kinds of the hybrid model...
متن کاملHybrid model predictive control of a nonlinear three-tank system based on the proposed compact form of piecewise affine model
In this paper, a predictive control based on the proposed hybrid model is designed to control the fluid height in a three-tank system with nonlinear dynamics whose operating mode depends on the instantaneous amount of system states. The use of nonlinear hybrid model in predictive control leads to a problem of mixed integer nonlinear programming (MINLP) which is very complex and time consuming t...
متن کاملAn ANOVA Based Analytical Dynamic Matrix Controller Tuning Procedure for FOPDT Models
Dynamic Matrix Control (DMC) is a widely used model predictive controller (MPC) in industrial plants. The successful implementation of DMC in practical applications requires a proper tuning of the controller. The available tuning procedures are mainly based on experience and empirical results. This paper develops an analytical tool for DMC tuning. It is based on the application of Analysis of V...
متن کاملA Linear Matrix Inequality (LMI) Approach to Robust Model Predictive Control (RMPC) Design in Nonlinear Uncertain Systems Subjected to Control Input Constraint
In this paper, a robust model predictive control (MPC) algorithm is addressed for nonlinear uncertain systems in presence of the control input constraint. For achieving this goal, firstly, the additive and polytopic uncertainties are formulated in the nonlinear uncertain systems. Then, the control policy can be demonstrated as a state feedback control law in order to minimize a given cost funct...
متن کامل