Simulation Parameter Calibration with Field Operational Data: Methods and Applications

Abstract

This dissertation study is concerned with the parameter calibration in computer models. Computer models are typically developed based on physics-based principles. Integrating computer models with real-world data provides an enabling tool for understanding and optimizing system performance. Computer models often require various parameter values to be appropriately specified. However, physical laws to accurately identify those parameters are often unavailable or insufficient in many applications. Parameter calibration is a procedure to identify those unknown parameters with observational data, aiming to improve the accuracy of computer models and make them represent near-exact replicas of real systems. Existing studies of computer model calibration typically build surrogate models to impute physical data under one key assumption: insufficient numbers of physical experiments, and/or computationally expensive computer trials. Advanced computing technology, however, now allows us to collect very large volumes of data. For example, it is possible to collect more than 50,000 data points for a single wind farm’s annual operation. When there is an abundance of both field observational and computer-generated data, this dissertation study describes new quantitative schemes to solve the calibration problem, by reconciling statistical theories and tools within the proposed optimization framework, which contributes to the following scientific advancements: • Development of a new stochastic optimization approach, enhanced with stratified sampling, to guide the search for the calibration parameter in a computationally efficient manner. • Enhancement of stratified sampling through dynamic partitioning for achieving the best computational resource allocation. • Development of a new non-parametric functional model for local calibration when the parameter depends on the input. We first define the problem as a stochastic optimization framework and employ stochastic gradient descent to iteratively refine the calibration parameters with randomly selected subsets of data. We propose a stratified sampling scheme that enables choosing more samples from noisy and influential sampling regions, thus reducing the variance of the estimated gradient for improved convergence. Then we reformulate the stochastic optimization framework such that the optimal partitioning, followed by the optimal resource allocation, can be integrated. Our approach, which dynamically partitions the input domain to identify the most influential variables and their partitioning locations, enables us to obtain the best subset of data in the parameter search process. Lastly, motivated by the reality that no physics laws or domain knowledge are available to define the functional relationship between input and calibration parameter in some applications, we propose a new, data-driven non-parametric model that does not require any pre-specified functional form. The developed methods are applied to the wake effect computer model in the wind power system. A well-calibrated wake effect model can help find the optimal wind farm layout. However, due to the lack of a methodology that can handle large-scale data, most studies have been limited to using small-scale datasets, which are not statistically representative. The outcomes of this research can potentially improve the layout design of new wind farms, as the wake decay parameter in the computer model is calibrated with data collected under various environmental conditions.