Marina Dorokhova
The energy industry is going through challenging times of disruptive changes caused by decarbonization, decentralization, and digitalization. As the energy value chain is restructuring itself to accommodate the growing penetration of renewables, increasing number of independent power producers, and augmented self-consumption, new energy management approaches are required to accomplish energy transition. With the Fourth Industrial Revolution underway, it becomes evident that digitalization is the key to increase energy efficiency and ensure stable, reliable, and secure operations of the electric grid. Due to the energy industry's massive inertia, most energy utilities are missing the real momentum for unleashing large-scale digitalization enabled by Information and Communication Technology (ICT). This thesis proposes a set of ICT-based software applications, models, and tools aiming to bridge the gap between strategic roadmaps focused on the energy industry's digitalization and their actual implementation in real-world scenarios through digital energy services. The research is conducted within the designed ICT-based smart building and smart community frameworks, the modular structure and scalability of which can serve as a backbone for future digital energy management solutions. The introduction of a novel unsupervised load disaggregation approach helps raise awareness of one's energy-related behavior and understand what drives the energy usage in residential households without compromising privacy and security. Showcasing algorithm's performance on real-world datasets from Norway and Germany highlights compliance with state-of-the-art disaggregation accuracy and reduced computational costs. The development of machine learning-based supervised and unsupervised building occupancy forecasting algorithms with prediction accuracies beyond 97% helps identify best-suited windows for energy-saving opportunities and deliver insights into one's presence and absence patterns. Built on top of that occupancy-centric rule-based heating and air conditioning automation algorithm strives to unlock the buildings' massive potential for energy savings without compromising the occupants' thermal comfort. Simulations on real-world datasets collected in Portugal demonstrate more than 15% potential energy savings. Zooming out from smart buildings towards smart communities, we focus on the important role of intelligent green mobility in supporting further digitalization of the electric power sector. To overcome the inconveniences posed by the sparsity of charging infrastructure and facilitate the adoption of Electric Vehicles (EVs), we present a reinforcement learning (RL)-based EV-specific routing method that guarantees paths' energy feasibility in a graph-theoretical context. Consequently, we propose several deep RL algorithms to control EV charging with the aim to increase renewables' self-consumption and EV drivers' satisfaction. Benchmarking against rule-based and model predictive control demonstrates RL's superior computational performance and better fitness for future mobility systems. Finally, we introduce an innovative decentralized blockchain-supported framework that enables secure and reliable accounting of energy exchanges within the smart community. Implementing it in a demonstration site in Switzerland shows blockchain's potential to reduce EV charging costs, transform the market's business model, and facilitate the large-scale deployment of EVs.
EPFL2021
