Lie, Tek TjingLi, WeihuaTito, Shafiqur RahmanParkash, Barkha2026-05-052026-05-052026http://hdl.handle.net/10292/21025Electricity is an essential resource in the modern world, underpinning nearly all aspects of daily life. In recent years, the residential electricity consumption profile has evolved significantly due to the increasing penetration of non-linear loads such as electric vehicles (EVs) and localised generation from rooftop photovoltaic (PV) systems. These changes pose new challenges for utility operators, who must manage infrastructure while ensuring a stable balance between electricity demand and supply. Accurate load forecasting is, therefore, critical to avoid mismatches that may compromise grid reliability and efficiency. In the residential sector, electricity consumption patterns vary widely between households, influenced by socio-demographic factors such as income, household size, and occupancy patterns. These contribute to diverse load profiles, necessitating a deeper understanding of their relationship with electricity usage to develop more accurate and responsive forecasting models. This research advances residential load forecasting by developing a novel top-down (TD) hierarchical forecasting framework. Traditional TD methods typically use fixed historical ratios for disaggregation, which can fail to adapt to dynamic consumption patterns. In contrast, the proposed framework leverages input features from both the aggregated level and the target sub-level, along with cross learning features, to forecast loads across multiple layers of hierarchy. Two model variations are proposed: a two-stage model, where separate neural models are trained for aggregated and sub-level forecasts with the second model utilising output of first model, and an end-to-end model that integrates all inputs into a single learning structure. These approaches enable accurate forecasting across different levels of hierarchy without relying on static assumptions, improving both scalability and performance. A second key contribution is the integration of socio-demographic features and EV charging data into the forecasting models. By incorporating variables such as household characteristics and EV charging behaviours, the study evaluates the impact that these have on the forecasting accuracy. The third major contribution is the investigation of EV charging prices and their interaction with charging time and load forecasting accuracy. As market liberalisation has introduced time varying and dynamic pricing structures, consumers are increasingly responsive to price signals. The thesis analyses how these pricing models affect EV charging patterns and, subsequently, the performance of the proposed forecasting models. It highlights that price driven charging behaviours introduce additional uncertainty and volatility into load profiles, which must be accounted for in effective forecasting strategies. Together, these contributions offer a robust and adaptable framework for residential load forecasting that accommodates evolving consumption patterns. The proposed methodology is not only scalable and data efficient but also tailored for practical deployment in modern power systems, offering utility providers and stakeholders a valuable tool for planning and operational decision making.enThesisOpenAccess