Design and Implementation of a novel deep learning model to forecast household energy consumption

Amidst the escalating global energy demand and the imperative for sustainability, accurate energy consumption forecasting has become critical for ensuring grid stability, optimizing resources and advancing renewable energy integration. The residential sector, accounting for a substantial share of global electricity consumption, poses unique challenges due to its dynamic consumption patterns and increasing adoption of decentralized energy generation and demand-side flexibility. Traditional forecasting models, including statistical and conventional machine learning approaches, often fail to address the complexities of modern energy consumption behaviours, necessitating innovative solutions. This thesis introduces JITtrans (Just In Time Transformer), a novel deep learning model designed specifically for time-series forecasting in the energy sector. By leveraging the self-attention mechanism, JITtrans captures both short- and long-term dependencies in large-scale, heterogeneous energy datasets. Additionally, clustering techniques are integrated to segment residential consumers on the basis of energy usage profiles, enabling tailored and scalable predictions that accommodate behavioural diversity. The contributions of this research are threefold: (1) systematic exploratory analysis of smart meter data to uncover key insights into customer behaviour and their implications for energy management; (2) development of a Transformer-based forecasting model optimized for energy systems; and (3) integration of clustering techniques to improve prediction accuracy and scalability. The findings offer actionable insights for energy utilities, facilitating demand-side management, operational planning and the integration of renewable energy sources. Experimental evaluation using proprietary smart meter data from an Italian power company demonstrates the superior performance of JITtrans over state-of-the-art models, including LSTM, CNN-LSTM and traditional Transformers. The proposed framework achieves significant improvements in prediction accuracy, robustness and adaptability across diverse consumption patterns. Beyond methodological advancements, this study bridges critical gaps in the energy sector by providing grid operators, policymakers and energy providers with tools to optimize demand management, enhance grid reliability and drive sustainability. The proposed framework lays a foundation for resilient, intelligent energy infrastructures capable of addressing the challenges of an evolving global energy landscape.