The rising cost and demand for energy, coupled with the need to meet environmental sustainability goals, present significant challenges that require a multifaceted approach. Accurate Energy Consumption Forecasting (ECF) emerges as a tool that enables more informed energy management by predicting future consumption patterns, allowing decision-makers to take pre-emptive action and implement effective planning strategies to reduce energy use while minimising waste and emissions. Whilst Machine Learning (ML) methods have become widely adopted for ECF due to their capacity to learn complex patterns from data without explicit programming, developing these models remains a resource-intensive task.

One of the primary reasons is that developing accurate ECF models requires effective Feature Engineering (FE). This requirement is particularly demanding because FE is a time-consuming process that is prone to human error and relies heavily on domain expertise to transform raw data into informative features that enhance model performance. This task is needed as raw data are often not ideal in their original form for algorithms to learn effectively. Additionally, real-world datasets are often small or limited due to data collection limitations, privacy issues, or resource constraints. In such scenarios, FE can compensate for limited data by extracting useful features, thereby maximising the utility of available data and enhancing computational efficiency by eliminating noisy or irrelevant information. Although state-of-the-art Automated Machine Learning (AutoML) frameworks have streamlined ML model development through automated model selection and hyperparameter tuning, they often assume that data preparation and feature generation have been completed, leaving FE largely dependent on human practitioners. This limitation has led to growing interest in automated FE methods. While there have been attempts to develop automated FE approaches, existing methods do not capture the domain-specific temporal patterns and consumption characteristics inherent in energy data. Additionally, these methods typically focus on maximising predictive accuracy, whereas they should ultimately serve downstream decision-making processes and operational outcomes in predict-then-optimise settings, such as minimising costs in energy management systems.

This thesis, therefore, aimed to address these challenges by developing ML models for ECF that minimise reliance on domain knowledge and can be applied across diverse energy systems while maximising both forecasting accuracy and downstream decision quality. Three key research objectives guided this thesis: (A) establishing baseline ECF ML models and investigating the role of domain knowledge in FE; (B) developing an automated FE method tailored for ECF that minimises the domain expertise needed for FE and seamlessly integrates with AutoML for fully automated ECF models; and (C) leveraging automated FE to optimise downstream tasks.

The first contribution of this thesis establishes a comprehensive ML pipeline capable of predicting one week into the future and suitable for small dataset sizes, evaluated on two novel real-world energy datasets. Through extensive investigation across different feature extraction and selection approaches, the proposed pipeline highlights the role of domain knowledge in FE, thereby establishing a strong empirical foundation for FE automation in later investigations. The second contribution introduces AutoEnergy, a novel automated FE algorithm specifically tailored for ECF problems. AutoEnergy automatically generates interpretable features from timestamps and past consumption values through rule-based transformations, integrating seamlessly with AutoML frameworks for fully automated ECF modelling. This method was extensively evaluated across eighteen diverse real-world energy datasets spanning residential, commercial, industrial, renewable, and grid power domains, demonstrating strong generalisation potential. On average, it achieves forecasting error reductions of 19.52% to 84.72% compared to baseline AutoML and established FE methods, while running 1.31x to 4.41x times faster than competing approaches. Finally, the third contribution leverages AutoEnergy to improve the nascent Decision-Focused Learning (DFL) framework, with a validation through a novel energy storage system data in real-world settings. Results show that incorporating AutoEnergy further improves the DFL performance by achieving 22.9-56.5% lower operating costs compared to the same models without it.

This thesis advances the field of ECF by demonstrating that domain-specific automated FE can significantly reduce expertise barriers while improving both predictive accuracy and decision quality. The development of AutoEnergy provides practitioners with a method that minimises the time and expertise required for ECF model development, while the integration with DFL ensures that forecasting improvements translate to tangible operational benefits. These contributions have broader implications for energy management systems facing similar challenges of expertise requirements, data scarcity, and the need for automated, reliable decision-support tools in the transition to smart, sustainable energy systems.