Enhancing energy demand forecasting and data imputation using deep learning : an integrated approach

Abstract

This PhD thesis introduces an integrated approach that leverages deep learning techniques to advance household electricity demand forecasting and data imputation within the UK energy sector. The research focuses on creating a novel system incorporating state-of-the-art machine learning solutions for electricity demand processing and prediction. The study involves data collection from appropriate electricity demand datasets, conducting comprehensive exploratory data analysis to uncover underlying patterns. A framework is established to process these datasets, encompassing data imputation, outlier handling, transformations, and feature scaling. A novel missing value imputation model is developed, employing a Transformer neural network and a K-means clustering algorithm to address missing data effectively. Subsequently, a forecasting framework for short-term residential load prediction is presented. This modelling framework integrates a Bayesian optimisation strategy, feature decomposition techniques, feature engineering, and percentile-based bias correction algorithms with a CNN-LSTM network to enhance prediction accuracy. The research contributes significantly to the field of household electricity demand forecasting and data imputation by offering a scalable and transferable framework. The application of these methodologies yields valuable insights, not only for the UK energy sector but also for broader applications, enabling precise predictions and efficient demand data processing. The findings promote energy efficiency and sustainable energy management practices.