Transfer learning in demand response: A review of algorithms for data-efficient modelling and control

A number of decarbonization scenarios for the energy sector are built on simultaneous electrification of energy demand, and decarbonization of electricity generation through renewable energy sources. However, increased electricity demand due to heat and transport electrification and the variability associated with renewables have the potential to disrupt stable electric grid operation. To address these issues using demand response, researchers and practitioners have increasingly turned towards automated decision support tools which utilize machine learning and optimization algorithms. However, when applied naively, these algorithms suffer from high sample complexity, which means that it is often impractical to fit sufficiently complex models because of a lack of observed data. Recent advances have shown that techniques such as transfer learning can address this problem and improve their performance considerably — both in supervised and reinforcement learning contexts. Such formulations allow models to leverage existing domain knowledge and human expertise in addition to sparse observational data. More formally, transfer learning embodies all techniques where one aims to increase (learning) performance in a target domain or task, by using knowledge gained in a source domain or task. This paper provides a detailed overview of state-of-the-art techniques on applying transfer learning in demand response, showing improvements that can exceed 30% in a variety of tasks. We observe that most research to date has focused on transfer learning in the context of electricity demand prediction, although reinforcement learning based controllers have also seen increasing attention. However, a number of limitations remain in these studies, including a lack of benchmarks, systematic performance improvement tracking, and consensus on techniques that can help avoid negative transfer.

Abstract licence: CC BY-NC-ND

40 more references, whose wording the paper’s licence keeps private.

A person counts once a day per paper. We keep no cookie and no address: a reader is a hash that changes every day. Any month under five reads shows no number.