Rising energy costs and decarbonisation targets demand smarter, computationally efficient AI-driven energy management. This thesis advances Energy Consumption Forecasting (ECF) by reducing reliance on expert-centric Feature Engineering (FE), a dominant bottleneck in Machine Learning implementation, particularly under small-data constraints. First, it proposes a reproducible end-to-end ECF pipeline for one-week-ahead forecasting. It conducts a systematic ablation study of FE on two real-world datasets, quantifying where domain knowledge contributes most and establishing an automation-ready baseline. Second, it introduces AutoEnergy, a novel, domain-aware, expert-free automated FE algorithm for ECF problems. AutoEnergy composes interpretable temporal representations from timestamps and historical consumption using rule-based operators and integrates with AutoML to enable fully automated ECF modelling. Across 18 heterogeneous real-world datasets spanning residential, commercial, industrial, renewable and grid domains, AutoEnergy reduces forecasting error by 19.52% to 84.72% relative to baseline AutoML and established AFE approaches, while improving computational efficiency by reducing runtime by a factor of 1.31 to 4.40. Third, it operationalises decision-aware forecasting via Decision-Focused Learning (DFL) by coupling prediction with optimisation for energy storage systems. An AutoEnergy-DFL framework is introduced to jointly forecast demand and electricity prices and optimise energy storage scheduling to minimise cost; validated on a real-world dataset and practical system configuration, it reduces operating cost by 22.9% to 56.5% compared with equivalent DFL models without AutoEnergy. Overall, the thesis delivers an automated, interpretable and decision-oriented ECF workflow suitable for diverse energy systems and resource-constrained settings.