Digital Twin Calibration with Operational Data

Abstract

Recent advances in numerical algorithms and computing power have brought digital twin technology to the forefront of operating and controlling systems. As virtual counterparts to physical systems, digital twins enable real-time monitoring, diagnosis, and forecasting of the system’s states. In many applications, digital twins are typically represented by physics-based computer models with complex mathematical functions. Successful digital twins must represent the physical systems’ behavior correctly, where parameter calibration is essential. Parameter calibration focuses on estimating parameters within digital twins to ensure that the simulation outputs reasonably match those of the physical process.

This dissertation leverages operational data for parameter calibration while addressing the research challenges posed by data size and complexity. Given the data, it proposes an efficient and robust framework for digital twin calibration, utilizing optimization-based procedures that seamlessly integrate uncertainty quantification methods from statistical inference.

A key motivating example in this dissertation is building energy systems. The building energy model, developed by the U.S. Department of Energy’s National Renewable Energy Laboratory, has been used to simulate energy use under various weather conditions and operational scenarios. To obtain accurate simulations, it is necessary to calibrate parameters that specify building conditions, such as heating, ventilation, air conditioning, lighting, process loads, and more.

This dissertation first focuses on digital twin calibration when parameters influence different datasets. For instance, in the building energy model, cooling (heating) season parameters are associated with the data collected during the cooling (heating) season only, whereas other global parameters are employed with the entire data. Under these circumstances, the parameters should be calibrated using the corresponding data only. This study proposes a new multi-block calibration approach that considers such heterogeneity. To address multiple data blocks, this study considers multiple loss functions, each for a block of parameters that use the corresponding dataset and estimate the parameters using a nonlinear optimization technique.

Subsequently, in analyzing the building energy simulation, a systematic discrepancy between simulated and measured energy use is observed, showing a daily periodic pattern that prevents precise estimation of energy demand. This study proposes a new bias-correction methodology by modeling the discrepancy with a time-series model. Concurrently, the heterogeneous variance observed in electricity loads is addressed by introducing weights via the iteratively reweighted least squares algorithm.

Lastly, this study proposes a novel calibration method for effectively estimating a large number of parameters in digital twins. Recently, Bayesian optimization has gained attention for its potential in digital twin calibration, especially when each simulation run is not computationally negligible. However, it faces challenges when handling a large number of parameters. A possible remedy is to selectively focus on influential parameters, simplifying a complex, high-dimensional task into a more tractable, lower dimensional endeavor. This study develops a new method that ranks parameter importance to enable stochastic dimension reduction effectively, utilizing the multi-armed bandit approach. By accounting for unequal importance among parameters, the developed approach generates accurate surrogate models tailored to the reduced dimension and guides efficient exploration of the parameter search space in the Bayesian optimization procedure.