HxPINN: A hypernetwork-based physics-informed neural network for real-time monitoring of an industrial heat exchanger

Abstract

The importance of heat exchangers in achieving sustainability goals has significantly increased across industries, given their crucial role in maintaining energy efficiency. However, lack of effective monitoring hardware for industrial heat exchangers often leads to suboptimal performance and unplanned stoppages. Physics-based numerical solvers, while accurate, are often computationally intensive and require frequent adjustments for monitoring. On the other hand, machine learning or data-driven models face a stiff challenge as they rely exclusively on data which is scarce due to lack of sensors. We present HxPINN, a hypernetwork-based physics-informed neural network (PINN) model for real-time monitoring of an industrial heat exchanger. Our approach incorporates a hypernetwork-based framework, enhancing a domain-decomposed PINN to learn the heat exchanger’s thermal behavior under dynamic operating conditions. HxPINN minimizes the need for retraining PINNs for different operating conditions and significantly reduces the inference time compared to existing approaches, without significant loss of accuracy vis-à-vis physics-based simulations. The methodology was tested on an Air Preheater (APH), a regenerative heat exchanger used in thermal power plants. The mean absolute error in APH internal temperatures with HxPINN when compared with benchmark physics-based simulation was found to be well below 5 °C across all the operating conditions. HxPINN has also accelerated the inference by up to 35–40x compared to existing approaches including transfer learnt PINNs and physics-based numerical solution. The methodology holds great potential for predictive maintenance of industrial equipment using digital twin systems providing efficient, adaptable, and real-time monitoring capabilities.