Optimizing Building Energy Efficiency through Explainable AI

The increasing urgency of climate change necessitates innovative solutions for energy management. This doctoral thesis explores the optimization of building energy flows through explainable artificial intelligence and addresses the need for efficient energy management systems that can adapt to complex, non-linear systems while maintaining real-time performance and scalability. For that, a comprehensive review of current energy flow optimization methods is given together with their strengths and limitations, highlighting the need for a more flexible and efficient approach. Genetic programming for symbolic regression is then proposed as a solution, as it offers the ability to generate interpretable mathematical models that optimize energy flows effectively and nearly optimal. The research questions focus on leveraging genetic programming to overcome existing energy management challenges and ensuring the explainability of the resulting models. The methodology involves the development a novel simulation-based optimization approach which enables the training of an arbitrary number of energy flow controllers. The adaptability of this approach is demonstrated through the optimization of both, simple electric systems and complex thermal-electrically coupled systems. Key contributions include the invention of multi-tree as well as adaptive operators for genetic programming, which allow a simultaneous optimization of multiple parameters within a single solution candidate. The adaptive operators additionally dynamically adjust the application of crossover and mutation based on the training progress, improving optimization results and training duration. Empirical evaluations show that the proposed approach achieves statistically significant improvements in energy cost reduction compared to state-of-the-art methods. Additionally, the explainability of the trained controllers is validated through detailed analyses, ensuring that the decision-making processes are transparent and understandable. In conclusion, this research advances the field of energy management by providing a scalable, efficient, and explainable AI-based solution, which will help to achieve global climate goals and enhance the sustainability of energy systems.