Online Demand Response Programs for Constrained Local Electricity Systems

The Centre for Machine Intelligence was set up the Department of Electronics and Computer Science of the University of Southampton to develop a coherent and ethical approach to research and technology transfer in the areas of Artificial Intelligence, Machine Learning, and Autonomous Systems. The Centre brings together researchers interested in both fundamental research as well as high impact applications.

Centre for Machine Intelligence

CMI Seminar on novel Demand Response (DR) program to mitigate grid congestion

University of Southampton Highfield Campus. B32/3077

The Centre for Machine Intelligence (CMI) at the University of Southampton invites you to a seminar on the topic of "Online Demand Response Programs for Constrained Local Electricity Systems". There are 40 places for in-person attendance (coffee and refreshments provided) and 100 places for online attendance.SpeakerProf. Dr. Dirk Neumann Albert-Ludwigs-Universität FreiburgAbstract:Renewable energies require new approaches to the operation of power systems. In this talk, I present a novel Demand Response (DR) program to mitigate grid congestion. The program stands out by adjusting electricity prices online and solely based on observed aggregate electricity consumption. It thereby avoids privacy concerns and expensive information system infrastructure investments. Additionally, it can cope with non-elastic and time-interdependent demand—which is typical for storage devices and industrial processes. ) model four flexible load types and provide a unified framework of load modeling under uncertainty and with time interdependencies.Numerical experiments show that the DR program achieves considerable and stable cost savings of 40% to 60% in comparison to conventional DR programs. Moreover, the solution approach based on Deep Reinforcement Learning reaches these savings after a short learning period corresponding to only 25 simulation days.Regarding load flexibility, I find that the responses of the four load types strongly differ depending on prices and the lag between the announcement and the application of price changes (`notification interval’) which is an important design parameter. The results imply that system operators should flexibilize their DR programs with the help of learning algorithms and tailor their program to the local load composition, congestion frequency, and forecasting quality.