To Örebro University

In the domain of building energy systems, many district heating control strategies rely on simplistic approaches that fail to account for the unique thermal characteristics of individual buildings. Traditional district heating control methods apply generic linear relationships between outdoor and supply temperatures, disregarding the significant variations in building-specific factors such as architectural design, geographical location, and occupancy patterns. These conventional approaches often result in substantial energy waste and suboptimal indoor comfort conditions. The challenge of optimizing energy consumption in residential buildings requires sophisticated methods that can adapt to diverse building characteristics while maintaining occupant comfort. Current prediction models treat buildings as homogeneous entities, leading to significant performance gaps between predicted and actual thermal behaviour. This limitation becomes particularly pronounced when attempting to develop control strategies that can generalize across different building types and geographical contexts. This dissertation addresses these challenges by developing a comprehensive framework for context-aware district heating optimization using advanced machine learning techniques. Context, in this work, refers to building-specific characteristics such as architectural design (building type, size, floors), and geographical location (climate zone, city) that uniquely influence each building’s thermal behavior and heating requirements. The approach systematically incorporates building-specific contextual information including rise type, geographical location, and architectural features into both predictive modelling and control optimization. We propose an adaptive context-aware transformer architecture that dynamically learns contextual representations through innovative embedding mechanisms, enabling the model to capture complex relationships between building characteristics and thermal behaviour patterns. Building upon these predictive foundations, we extend the framework to control optimization through two complementary approaches. First, we develop a context-enhanced district heating control system using dual-output prediction and multi-objective optimization that balances energy efficiency with comfort maintenance. Second, we introduce a novel reinforcement learning framework that integrates the adaptive context-aware transformer with Deep Q-Networks to discover optimal control policies without being constrained by historical control patterns. The outcome of this dissertation advances building energy systems through machine learning approaches that can adapt to individual building characteristics while optimizing across multiple objectives. Our proposed methods demonstrate how context-aware modelling can transform district heating systems from reactive, generic control to proactive, building-specific optimization. We evaluate and validate our methods experimentally in real-world scenarios across diverse building types and geographical locations, underscoring the effectiveness and practical applicability of our approaches for reducing energy consumption while maintaining occupant comfort.