Physically Consistent Neural Networks for building thermal modeling: Theory and analysis

Due to their high energy intensity, buildings play a major role in the current worldwide transition. Building models are ubiquitous since they needed at each stage of life buildings, i.e. for design, retrofitting, and control operations. Classical white-box models, based on physical equations, bound follow laws physics but specific design underlying structure might hinder expressiveness hence accuracy. On other hand, black-box better suited capture nonlinear building dynamics thus can often achieve accuracy, require lot data not physics, problem that is particularly common neural network (NN) models. To counter this known generalization issue, physics-informed NNs have recently been introduced, where researchers introduce prior knowledge ground them avoid classical NN issues. In work, we present novel architecture, dubbed Physically Consistent (PCNN), which only requires past operational no engineering overhead, including linear module running parallel NN. We formally prove such networks physically consistent – by even unseen with respect different inputs temperatures outside neighboring zones. demonstrate performance case study, PCNN attains an accuracy up 40% than physics-based resistance-capacitance model 3-day long prediction horizons. Furthermore, despite constrained structure, PCNNs attain similar validation data, overfitting training less retaining tackle issue.