Reduced Order Modeling of a Heat Exchanger with a Stacking Ensemble to reduce Computational Inefficiencies

Abstract

Reduced Order Modeling is a technique for reducing the computational complexity of a model while preserving the expected fidelity within a controlled error. One of the techniques used to create a Reduced Order Model (ROM) is Artificial Neural Networks (ANN). A successful approach to reducing the variance of ANN model prediction is to train multiple models instead of a single model and to combine the predictions from these models, which is commonly called Ensemble learning. When the predictions from the multiple models are combined using another regression model, it is called Stacking ensemble. This paper studies the effectiveness of using Genetic programming algorithm in taking the outputs of each model as input and attempting to learn how to best combine the input predictions to make a better output prediction.The above-mentioned approach is used to create a ROM for a crossflow heat exchanger steady-state component. There are 6 inputs parameters namely Cold & Hot inlet temperature, Cold & Hot outlet pressure and Cold & Hot inlet flow. There are four outputs namely Hot & Cold outlet temperature and Hot & Cold inlet pressure. A multi-input single output (MISO) ROM is created for each of the outputs. There are 3 different configurations of ANNs used to cover a good range of the Hyperparameter values. The output from each of the ANNs is then combined using Genetic Programming Algorithm. The Overall model has an R2 value of above 95% for each of the outputs. The ROM thus created can run simulations at a much faster rate. The ROM of the HX component is a black box and can be shared with third party without any concerns over propriety information loss.