Accuracy, generalizability, and extrapolation ability of physics-based, data-driven, and hybrid models for real-life cooling towers

Abstract

This study compares the five simulation models of real-life cooling towers in terms of model accuracy, generalizability, and extrapolation ability; a physics-based model (#1), a data-driven model (#2), a physics-informed hybrid model (#3), a physics-guided hybrid model (#4), and a transfer-learning (TL) hybrid model (#5). For this purpose, we gathered relevant data from target cooling towers for approximately three years (May 2020 to March 2023) at a sampling time of 1-h. Regarding the model accuracy, the four models (#2–4) exhibited good-enough predictive performance by achieving a cvRMSE of 15.9–18.0 % for predicting the cooling tower heat removal and MAE of 0.6–0.7 °C and R² of 0.96–0.98 for predicting cooling water outlet temperature, while the physics-based model exhibits the lowest accuracy (cvRMSE: 64.2 %, MAE: 2.9 °C, R²: 0.66). As a generalizability test, we applied the five models developed for the target system to a nearby identical system. The four models (#2–4) exhibited the reduced accuracy, close to the required level by ASHRAE Guideline 14 (lower than 30.0 % for hourly prediction). The generalizability study implicates that models #2–5 need to be re-adjusted against the unseen system data. Regarding the extrapolation ability, only the physics-based and TL-hybrid models exhibited physical consistency between the cooling water outlet temperature and on the number of fans. It is noteworthy that the TL-hybrid model (#5) performs best in terms of the model accuracy, generalizability, and extrapolation ability beyond the training data.