Explainable AI for deriving predictive and trustworthy operational rules for building energy control

Innovative solutions for optimizing energy use and reducing carbon emissions lead to increasing complexity of the building energy systems, influenced by varying conditions such as weather, user behavior, energy availability and grid requirements. This poses a challenge to traditional control strategies based on simple "if then" rules. To maximize energy efficiency or control buildings with complex objectives, new intelligent strategies are needed that can use prediction and optimization calculations. Traditional expert rule systems are becoming too complex for building operators to understand, necessitating a shift to advanced model-based control (MPC) techniques. However, their demand for sophisticated building automation infrastructure and systems expertise are challenging. Our research addresses these challenges by fostering Explainable Artificial Intelligence (XAI) techniques to extract new knowledge from existing data, such as measurements, simulations or expert knowledge, and translate it into simple, comprehensible rules. The aim is to create a method that approximates the behavior of complex optimization-based control strategy through a learning-based rule base. This results in an easily implementable, humanly understandable, and thus trustworthy rule-based control system that adapts to environmental changes and achieves a near optimal control. To investigate the potential of the developed concept, we apply it to the use case grid supported heat pump control and compare it with other control approaches such as MPC in term of control performance and trustworthiness. The testing is conducted within a hardware-in-the-loop experiment in a heat pump laboratory, ensuring a rigorous assessment of the method's effectiveness and reliability.