Multi-Agent Mechanisms and Learning in Energy Applications

The thesis addresses challenges in multi-agent systems in which they consist of multiple autonomous agents where they may have different information and/or different interests in a shared environment. Most, if not all, large-scale complex systems such as the Internet, transportation systems, or power systems can be viewed as an ecosystem for interaction among self-interested or distributed individuals that we loosely call agents. On one hand, we contribute to a fundamental understanding of multi-agent systems that we study two classes of multi-agent dynamic games, namely Mean Field Games (MFG) and Markov Potential Games (MPG). MFG framework is a promising approach when systems involve a large number of agents, and the framework gets more accurate as the number of agents grows. MPG framework is a class of games in which there exists an optimal deterministic joint policy. While, in standard Markov Decision Processes, this is always true, in games, it is not always guaranteed in the presence of other agents. One property of such determinism is that policies are unconditional and, hence, other agents may not need collect additional information on other agents' behaviors. In both games, we propose computation and learning methods to identify an equilibrium of games. On the other hand, we contribute mechanism designs for Demand Response (DR) frameworks in the operation of power systems and empirical studies on their efficient operation. DR is increasingly gaining attention in electricity markets as a mechanism to provide ancillary support to reduce price volatility and to enhance grid resilience under intermittent renewable energy. However, DR programs remain limited in the residential sector, one of the largest sectors in the U.S. To facilitate the deployment of residential DR programs, we propose three necessary components. First, we propose a proper DR mechanism where socially desirable behaviors of agents are anticipated. Secondly, we empirically study the physical and market operational impacts of DR programs on a synthetic representation of the Electric Reliability Council of Texas (ERCOT) electricity grid, and measure efficiency of successful DR programs. Lastly, we provide a testbed to integrate residential participants (agents) where householders can participate in real-time DR programs