JAX-based Gray-Box Modeling Framework for Hybrid Thermal Building Models

Modeling the thermal dynamics of buildings is crucial for effectively controlling HVAC systems. However, the realization of corresponding models can be quite cumbersome, identifying multi-zone systems algorithms for gray-box models can be quite complex and machine learning methods require large amounts of data. Therefore, we propose a hybrid approach that combines fully observable gray-box models with a neural network corrector. In this study, we present a multi-zone modeling approach that uses a JAX-based gray-box modeling framework and validate it using an EnergyPlus simulation. In the second part of our research, we examine the data requirements of this method. For the purpose of comparison and explainability, we test our method on four different fully observable state space models. We demonstrate that gray-box modeling, when combined with a corrector term, can lead to highly general and explainable multi-zone thermal building models.