Robust Gaussian Process Regression and its Application in Data-driven Modeling and Optimization

• Author / Creator

  Ranjan, Rishik

• Availability of large amounts of industrial process data is allowing researchers to explore new data-based modelling methods. In this thesis, Gaussian process (GP) regression, a relatively new Bayesian approach to non-parametric data based modelling is investigated in detail. One of the primary concerns regarding the application of such methods is their sensitivity to the presence of outlying observations. Another concern is that their ability to predict beyond the range of observed data is often poor which can limit their applicability. Both of these issues are explored in this work.

  The problem of sensitivity to outliers is dealt with by using a robust GP regression model. The common approach in literature for identification of this model is to approximate the marginal likelihood and maximize it using conjugate gradient algorithm. In this work, an EM algorithm based approach is proposed in which an approximate lower bound on the marginal likelihood is iteratively maximized. Models identified using this method are compared against those identified using conjugate gradient method in terms of prediction performance on many synthetic and real benchmark datasets. It is observed that the two approaches are similar in prediction performance. However the advantages of EM approach are numerical stability, ease of implementation and theoretical guarantee of convergence.

  The application of proposed robust GP regression in chemical engineering is also explored. An optimization problem for an industrial water treatment and steam generation network is formulated. Process models are constructed using material balance equations and used for data reconciliation and steady state optimization of the cost of steam production. Since the overall network is under manual operation, a dynamic optimization framework is constructed to find a set point change strategy which operators can use for minimizing steam production cost. Dynamic models for process units and tanks are integrated into this framework. Some of these models are identified using proposed robust GP regression method. Extrapolation ability of identified GP models is improved by applying a suitable GP kernel structure and by using some ad hoc scaling techniques. Based on the application of robust GP regression to an industrial optimization problem, it is shown that non-parametric data-based modelling can be successfully integrated with process optimization objectives.

• Subjects / Keywords

  • SAGD
  • EM algorithm
  • Optimization
  • Gaussian process regression
  • Robust identification
  • Outliers

• Graduation date

  Fall 2015

• Type of Item

  Thesis

• Degree

  Master of Science

• DOI

  https://doi.org/10.7939/R3Q81513G

• License

  This thesis is made available by the University of Alberta Libraries with permission of the copyright owner solely for non-commercial purposes. This thesis, or any portion thereof, may not otherwise be copied or reproduced without the written consent of the copyright owner, except to the extent permitted by Canadian copyright law.