Appliance Classification and Scheduling in Residential Environments with Limited Data and Reduced Intrusiveness

The United Kingdom aims for a 78% reduction in greenhouse gas emissions by 2035, with a specific carbon budget for 2033–2037. Despite rising CO2 emissions from 2021 to 2022 due to increased energy demands, this thesis presents novel strategies to reduce residential electricity consumption, a major emissions driver. It addresses two critical gaps in energy management: First, it develops a feature extraction methodology using machine learning and deep learning for accurately classifying high-power household appliances with smart meter data. Traditional methods often require complex setups or large datasets, leading to intrusiveness and implementation challenges. This research introduces the Spectral Entropy – Instantaneous Frequency (SE-IF) method, effective with limited datasets and enhancing usability (Chapter 3). Second, it proposes an optimisation model that intelligently schedules household appliance usage to balance costs, emissions, and user comfort, incorporating renewable energy and battery storage systems. Existing scheduling techniques typically overlook significant CO2 reductions and user comfort. The thesis utilises the Multiobjective Immune Algorithm (MOIA) to demonstrate this model’s effectiveness, achieving a 9.67% cost reduction and a 16.58% decrease in emissions (Chapter 5). Chapters 4 and 5 further detail how the SE-IF method, paired with a Bidirectional Long Short-Term Memory (BiLSTM) network, achieves a 94% accuracy in identifying appliances from aggregated data and applies the multi-objective optimisation in various scenarios. This research advances the integration of energy efficiency, environmental sustainability, and user-centric solutions in smart homes, contributing significantly to national goals of reducing energy consumption and emissions.