Heating System Process Identification Using Physics-Based Neural Networks

This paper presents a framework for the process identification of heating systems integrating a physics-based network with a classical neural network. The core of the physics-based network is defined by energy balance equations. Integrating physics within the overall neural network can, e.g., enhance model accuracy and provide more reliable predictions across a wider range of input signals than those present in the training set. Process identification was performed using three distinct neural network models: a physics-based network combined with a feedforward neural network (PBNN), a feedforward neural network with an added residual component (FNNR), and a purely physics-based network (PBN). The comparative analysis demonstrated that the PBN and PBNN models outperformed the FNNR model, offering better accuracy and reliability. Based on the study, the PBN model was selected as the most suitable candidate for further research in implementing digital twins. The results show the importance of incorporating physics into process identification for robust process modeling, particularly in capturing the dynamics of the process.