Ph.D., Colorado State University, August 2015
Co-Major Professors: H. J. Siegel and Anthony A. Maciejewski
This dissertation examines resource allocation optimization in the areas of Smart Grid and high-performance computing (HPC). The primary focus of this work is resource allocation related to Smart Grid, particularly in the areas of aggregated demand response (DR) and demand side management (DSM). Towards that goal, a framework for heuristic optimization for DR in the Smart Grid is designed. The optimization problem, denoted Smart Grid resource allocation (SGRA), controls a large set of individual customer assets (e.g., smart appliances) to enact a beneficial change on the electric power system (e.g., peak load reduction). In one part of this dissertation, the SGRA heuristic framework uses a proposed aggregator-based approach. The aggregator is a for-profit entity that uses information about customers' smart appliances to create a schedule that maximizes its profit. To motivate the customers to participate with the aggregator, the aggregator offers a reduced rate of electricity called customer incentive pricing (CIP). A genetic algorithm is used to find a smart appliance schedule and CIP to maximize aggregator profit. By optimizing for aggregator profit, the peak load of the system is also reduced, resulting in a beneficial change for the entire system. Visualization techniques are adapted, and enhanced, to gain insight into the results of the aggregator-based optimization. A second approach to DR in the Smart Grid is taken in the form of a residential home energy management system (HEMS). The HEMS uses a non-myopic decision making technique, denoted partially-observable Markov decision process (POMDP), to make sequential decisions about energy usage within a residential household to minimize cost in a real-time pricing (RTP) environment. The POMDP HEMS significantly reduces the electricity cost for a residential customer with minimal impact on comfort.
The secondary focus of the research is resource allocation for scientific applications in HPC using a dual-stage methodology. In the first stage, a batch scheduler assigns a number of homogeneous processors from a set of heterogeneous parallel machines to each application in a batch of parallel, scientific applications. The scheduler assigns machine resources to maximize the probability that all applications complete by a given time, denoted the makespan goal. This objective function is denoted robustness. The second stage uses runtime optimization in the form of dynamic loop scheduling to minimize the execution time of each application using the resources allocated in the first stage. It is shown that by combining the two optimization stages, better performance is achieved than by using either approach separately or by using neither.
The specific contributions of this dissertation are: (a) heuristic frameworks and mathematical models for resource allocation in the Smart Grid and dual-stage HPC are designed, (b) CIP is introduced to allow an aggregator profit and encourage customer participation, and (c) heuristics and decision-making techniques are designed and analyzed within the two problem domains to evaluate their performance.