Record-Driven Calibration of Cooling-Tower Performance Digital Twins Using Hiera-UNet

Abstract

Although design equations and commissioning tests are typically used to start cooling-tower digital twins, variations in fill resistance, nozzle distribution, fan condition, instrumentation bias, and other factors cause their accuracy to decrease. Although the records are asynchronous, regime-dependent, and primarily comprise ordinary operating data rather than controlled test findings, plant historians have gathered the information required for recalibration. A record-driven Hiera-UNet model for ongoing performance calibration of wet cooling-tower digital twins is presented in this study. A low-order heat- and mass-transfer twin is contrasted with historian tags arranged into multi-resolution operational windows. Seasonal environmental variations, shift-scale loading, and rapid actuator changes are all represented by a hierarchical attention encoder. In order to calculate corrections for outlet water temperature, approach, efficacy, and the four identifying twin characteristics, a U-Net style temporal decoder uses encoder features and physical residual channels. The calibration is limited by energy-balance, monotonicity, smoothness, and uncertainty losses. Fourteen months of one-minute recordings from six tower cells at the two industrial sites served as the foundation for the technique. Hiera-UNet maintained a 90% interval coverage of 91.3% on the held-out station while lowering the approach error by 71.6% and the outlet-temperature mean absolute error for the uncalibrated twin to 0.29 °C from 1.18 °C. Additionally, fill-performance decrease at 4.1% and simulated sensor bias at 0.12 °C are found. The findings of the ablation experiment show that physical residual limitations and hierarchical encoding have added the most accuracy. The end product is a digital twin calibration layer that can be audited without changing the engineering framework.