Smart Activation of a Dynamic PCM-Insulation Systems in Building Envelopes

Abstract

Buildings are among the largest consumers of energy worldwide, with space cooling accounting for a substantial portion of their operational demand. In typical residential and commercial buildings, HVAC (Heating, Ventilation, and Air Conditioning) systems alone can represent around 40% of total energy consumption. Phase change materials (PCMs) have therefore attracted significant interest as a means of storing and releasing thermal energy to reduce cooling demand. Previous work by our research group showed that a dynamically actuated PCM-insulation system with a cylindrical geometry can achieve very high cooling-load reductions, reaching up to 101% under carefully optimized, rule-based operation. Despite this strong performance, the reliance on fixed activation schedules makes such systems sensitive to weather variability and limits their effectiveness under realistic operating conditions. This thesis builds on that earlier work by exploring how the dynamic PCM system can be operated in a more flexible and adaptive manner. An integrated framework is proposed that combines Internet of Things (IoT) sensing with physics-informed machine learning to predict indoor temperature evolution and support data-driven control decisions. Environmental and system data are collected through an IoT-oriented architecture, while a time-series dataset is generated using COMSOL Multiphysics simulations, capturing indoor and outdoor temperatures, PCM rotation angles, and equivalent thermal resistance capacitance parameters. To make the predictions more stable over long periods, the modeling approach focuses on predicting changes in indoor temperature rather than absolute temperature values. Simple physical insights about how the PCM system behaves are then built into the model through carefully chosen input features. These include rotation-independent thermal parameters and smoothed representations of the outdoor temperature, which help the model better reflect the system’s thermal memory. Model performance is evaluated using a leave-one-experiment-out cross-validation approach across different operating scenarios. The results show that simple linear models give more stable and reliable predictions than more complex methods, such as ensemble models and recurrent neural networks, when basic physical information is included. Across the tested scenarios, the models typically predict indoor temperature with mean errors between 0.2 and 0.8 °C and achieve coefficients of determination (R²) above 0.8 in the most representative operating conditions, while remaining accurate over several days, especially when temperature variations follow a repeating pattern. Based on these results, a simple control strategy is proposed to decide how the PCM system should be rotated while maintaining indoor comfort and avoiding unnecessary actuator movement. Although the study relies on simulation data, the findings suggest that combining dynamic PCM systems with IoT sensing and simple, physics-aware prediction models can reduce reliance on traditional HVAC systems and support more energy-efficient building operation.