Control strategy for dynamic sustainable systems

Tze Chao Chiam, Purdue University

Abstract

This study investigates the development of strategies for controlling a NASA ALS/NSCORT system. The water subsystem is the focus due to the complexity of the overall system, and because water is one of the most expensive and crucial resources for human survival. A water-treatment model is constructed based on BVAD published by NASA. Using a simulation built, various aspects, or conditions, of the system are captured hourly. These conditions form the system state. Policies are defined in terms of water-treatment efficiencies. One policy is to be applied to the system at the beginning of every hour. The goal of this study is to develop a strategy for choosing the “best” policy in order to minimize the water deficiencies experienced by the crewmembers, while minimizing operational costs. The Markov Decision Process is used to aid the decision on choosing the best policy. State-space exploration techniques are developed. The Controlled Random Walk technique developed inherits the unbiased nature of Pure Random Walk while allowing controls to be imposed to visit a larger portion of the state space within a given time. Together with re-initialization of system conditions, this technique is shown to outperform some other techniques tested. From the data captured, state transition probabilities are calculated. To ensure that these probabilities represent the dynamics of the system, a probability convergence measure is defined. Based information theory, this measure provides an indication of whether or not the probabilities obtained have converged. The Policy Iteration algorithm by Howard is used to obtain the best policies. Several scenarios based on a test case are constructed to assess the performance of these policies. Results show that the best policies outperformed cases where fixed policies as well as random policies are used. Due to the incomplete knowledge of the state space, policies from Policy Iteration do not cover the entire state space. See5.0 algorithm is used to generate decision trees based on these policies to cover the state space. Comparisons made between cases that use both methods for choosing control policies show that there are no statistical differences between the system performances.

Degree

Ph.D.

Advisors

Yih, Purdue University.

Subject Area

Industrial engineering|Systems science

COinS