Informal Systems Seminar (ISS), Centre for Intelligent Machines (CIM) and Groupe d'Etudes et de Recherche en Analyse des Decisions (GERAD)
Reinforcement Learning Applications in Power Systems
Department of Electrical Engineering , Ecole Polytechnique de Montreal
March 24, 2017 at 3:00 PM
McConnell Engineering Room 437
Reinforcement Learning is an area of machine learning concerned with how an agent can pick its actions in a dynamic environment to transition to new states, in such a way that maximizes a cumulative reward. Reinforcement Learning methods allow the development of algorithms to learn online the solutions to optimal control problems for dynamic systems that are described by difference equations. These involve two-step techniques, known as policy iteration (PI) or value iteration (VI).
The recent decades showed accelerated evolution of Reinforcement Learning applications in power Systems. Reinforcement Learning and optimization techniques are utilized to assess the security of the electric power systems and to enhance Microgrid performance. Adaptive learning methods are employed to develop control and protection schemes. Transmission technologies with High-Voltage Direct Current (HVDC) and Flexible Alternating Current Transmission System devices (FACTS) based on adaptive learning techniques can effectively help to reduce transmission losses and CO2 emissions.
In this talk, applications of Reinforcement Learning are highlighted for three research problems in power systems. First, Reinforcement Learning is used to develop distributed control structure for a set of distributed generation sources. The exchange of information between these sources is governed by a communication graph topology.
Second, an online adaptive learning technique is used to control the voltage level of an autonomous Microgrid. The control strategy is robust against any disturbances in the states and load. Only partial knowledge about the Microgrid's dynamics is required. Finally, Q-Learning with eligibility traces technique is adopted to solve the power systems non-convex Economic Dispatch problem with valve point loading effects, multiple fuel options, and power transmission losses. The eligibility traces are used to speed up the Q-Learning process.