Artificial intelligence for decision making in energy demand-side response

Abstract

This thesis examines the role and application of data-driven Artificial Intelligence (AI) approaches for the energy demand-side response (DR). It follows the point of view of a service provider company/aggregator looking to support its decision-making and operation. Overall, the study identifies data-driven AI methods as an essential tool and a key enabler for DR. The thesis is organised into two parts. It first provides an overview of AI methods utilised for DR applications based on a systematic review of over 160 papers, 40 commercial initiatives, and 21 large-scale projects. The reviewed work is categorised based on the type of AI algorithm(s) employed and the DR application area of the AI methods. The end of the first part of the thesis discusses the advantages and potential limitations of the reviewed AI techniques for different DR tasks and how they compare to traditional approaches. The second part of the thesis centres around designing machine learning algorithms for DR. The undertaken empirical work highlights the importance of data quality for providing fair, robust, and safe AI systems in DR — a high-stakes domain. It furthers the state of the art by providing a structured approach for data preparation and data augmentation in DR to minimise propagating effects in the modelling process. The empirical findings on residential response behaviour show better response behaviour in households with internet access, air-conditioning systems, power-intensive appliances, and lower gas usage. However, some insights raise questions about whether the reported levels of consumers’ engagement in DR schemes translate to actual curtailment behaviour and the individual rationale of customer response to DR signals. The presented approach also proposes a reinforcement learning framework for the decision problem of an aggregator selecting a set of consumers for DR events. This approach can support an aggregator in leveraging small-scale flexibility resources by providing an automated end-to-end framework to select the set of consumers for demand curtailment during Demand-Side Response (DR) signals in a dynamic environment while considering a long-term view of their selection process.