A data-driven framework for quantifying and characterising the energy flexibility of residential buildings

A holistic and scalable evaluation of the residential building stock flexibility potential is necessary for its integration into the future smart grid. However, the lack of standardised indicators to characterise building flexibility along with the lack of robust and scalable approaches beyond building simulation to evaluate the associated DR potential prevents this potential from being fully realised. This is because the residential building consumption pattern heterogeneity poses significant barriers to electricity aggregators due to the range of technologies involved, the interactions between the various energy conversion systems and the end-users, occupant preferences, and other boundary conditions. To unlock the flexibility potential of the residential building stock, it is necessary to assess the DR potential of the diverse energy systems involved on an integrated common basis while considering a transferable and end-user tailored methodology. This thesis addresses these problems by presenting a fundamental framework for characterising the potential energy flexibility of residential buildings considering both simulation-based and data-driven approaches. Specifically, a novel indicator set is proposed to assess the flexibility potential of several thermal and electrical systems commonly found in residential buildings on an integrated basis. This framework is expanded beyond building simulation to analyse energy flexibility utilising a robust ensemble learning framework with short-term time horizons. In addition, a Bayesian deep learning framework is developed to consider the various uncertainties related to building load demand and onsite electricity generation. Finally, a set of probabilistic flexibility indicators is developed to establish prediction bounds for the DR potential of various energy systems, as well as potential thermal comfort deviations arising from harnessing building flexibility. The flexibility potential of the various building energy systems is assessed by using the developed machine learning frameworks and referenced against a physics-based model. The latter is a calibrated white-box model of an all-electric residential building developed using EnergyPlus. The proposed indicators have the potential to be used to all residential building types, depict the DR potential of various power modulation strategies, and illustrate it concisely and uniformly. Additionally, simulation results indicate that occupant preferences have a significant impact on the building electricity profile and, consequently, the associated flexibility potential. Further, different occupancy profiles and prediction horizons may result in the selection of different features for each target variable. Considering the deterministic performance of the two data-driven models developed, it is shown that the heterogeneous ensemble model outperforms the Bayesian deep learning model for all target variables and prediction horizons evaluated. Finally, the energy shifting capability related to the downward flexibility of the heat pump and the battery flexibility can be accurately estimated by using any of the machine learning models developed. The data-driven energy flexibility quantification and characterisation framework developed may be of interest to electricity aggregators since it allows them to assess or optimise building portfolios in an occupant-centred manner, and ultimately identify and engage end-users with high flexibility potential. Using this framework, the flexibility of each energy system can be optimally harnessed and peak demand consumption can be shifted to off-peak periods or periods of excess onsite electricity generation. This feature can facilitate the integration of the residential building stock flexibility potential into the future smart grid, maximising the use of renewable energy sources while mitigating potential production and distribution capacity constraints and promoting energy security and decarbonisation.