The increasing cost and demand for energy, coupled with the need to achieve environmental sustainability goals, pose significant challenges that require a multifaceted approach. This PhD aims to address these challenges by leveraging predictive Machine Learning (ML) techniques, automating feature engineering, and integrating Decision-Focused Learning (DFL) to improve energy consumption forecasting.

The United Kingdom's largest manufacturing sector, the food and drink industry, is under significant sustainability pressure due to high energy consumption and greenhouse gas emissions. Food and Drink Cold Storage (FDCS) rooms, as a component of food supply chains, are vital for preserving perishable goods such as dairy, meat, and fresh produce by maintaining proper temperature and humidity levels. Optimising electricity use in FDCSs through accurate forecasting can improve operations and reduce energy consumption by better scheduling door openings, maintenance, and restocking. Although ML has been applied to forecast energy use in various domains such as commercial and residential buildings, its use in addressing the specific challenges of FDCS, which require stringent temperature and humidity control for food safety and quality, has been less explored. The first study of this PhD addresses this gap by proposing a tailored ML pipeline for FDCS settings capable of predicting one week (hourly) into the future and suitable for small dataset sizes. Moreover, in contrast to existing studies predominantly concerned with energy consumption forecasting, this study includes the forecasting of indoor temperature and humidity, given their essential role in preserving the quality and longevity of stored food items. Using two real-world FDCS datasets, this study provides a comparative analysis of model performance, dataset size impact on performance, and feature importance analysis. This work has been accepted at the IEEE Access journal [1].

While the first study successfully developed a tailored ML pipeline for FDCS settings, it was observed that developing such models is time-consuming and expert-dependent. Automated ML (AutoML) can streamline ML pipelines, yet it often still requires human experts to enhance performance, particularly in feature engineering. This is crucial in complex energy domains where deep learning's automatic feature extraction lacks interpretability and demands large datasets. To address this issue, the second study introduced an automated feature engineering algorithm for energy forecasting problems. This method aims to generate a comprehensive set of features for AutoML, reducing the need for domain-specific expertise for feature engineering. The proposed method, validated with eleven datasets from various energy domains, demonstrates a considerable reduction in prediction errors. This highlights its potential to improve AutoML performance and address existing feature engineering limitations in such frameworks. This work has been accepted and presented at the IEEE SMC24 conference, and the full paper will be available in the proceedings soon. Building upon these promising findings, ongoing work includes extending the experimental setup, incorporating additional datasets, and submitting the refined study to a journal.

For future work, the plan is to address the gap in integrating ML predictions into actionable decisions by leveraging DFL. Unlike the traditional predict-then-optimise approach, DFL combines prediction and optimisation into a single framework to directly enhance decision quality. This approach aims to improve both predictive accuracy and decision effectiveness by minimising energy usage and operational costs while maintaining optimal conditions, such as those found in FDCS environments. The objective of this study is to develop an end-to-end DFL framework that supports the thesis's overarching aim of improving energy consumption forecasting.