Regression Surrogate Model Research Articles

Accelerated Monte Carlo system reliability analysis through machine-learning-based surrogate models of network connectivity

The two-terminal reliability problem in system reliability analysis is known to be computationally intractable for large infrastructure graphs. Monte Carlo techniques can estimate the probability of a disconnection between two points in a network by selecting a representative sample of network component failure realizations and determining the source-terminal connectivity of each realization. To reduce the runtime required for the Monte Carlo approximation, this article proposes an approximate framework in which the connectivity check of each sample is estimated using a machine-learning-based classifier. The framework is implemented using both a support vector machine (SVM) and a logistic regression based surrogate model. Numerical experiments are performed on the California gas distribution network using the epicenter and magnitude of the 1989 Loma Prieta earthquake as well as randomly-generated earthquakes. It is shown that the SVM and logistic regression surrogate models are able to predict network connectivity with accuracies of 99% for both methods, and are 1–2 orders of magnitude faster than using a Monte Carlo method with an exact connectivity check.

Integrating building performance simulation in agent-based modeling using regression surrogate models: A novel human-in-the-loop energy modeling approach

Building Performance Simulation (BPS) is an established method used in the design phase of buildings to predict energy consumption and guide design choices. Despite their advanced abilities to model complex building systems, BPS tools typically fail to account for different and changing energy use characteristics of building occupants, contributing to important prediction errors. In parallel, Agent-Based Modeling (ABM) has emerged in recent years as a technique capable of capturing occupants’ dynamic energy consumption behaviors and actions. However, ABM lacks the building simulation capabilities to account for the complexity of various building systems in energy calculations. This research proposes a new modeling framework that integrates BPS in ABM using trained regression surrogate models. The framework is unique in its ability to (1) simulate energy use attributes of building occupants and facility managers, (2) translate those attributes to robust energy consumption estimates, and (3) help quantify the impact of uncertainty in human actions on the performance of the built environment. The framework is tested and illustrated in a case study on a prototype office building. Results indicate that providing occupants with control over their building systems can mitigate the effect of uncertainty in human actions on the performance of the built environment.

Multifidelity Surrogate Modeling of Experimental and Computational Aerodynamic Data Sets

This study presents a multifidelity surrogate modeling approach, combining experimental and computational aerodynamic data sets. A multifidelity cokriging regression surrogate model is used. This study highlights how lowfidelity data from computations contribute to improving surrogate models built with limited high-fidelity data from experiments. Various types of sampling design for low fidelity data are also examined to study the impact of characteristics of the sampling design on the final surrogate models. Replication, blocking, and randomization techniques originally developed for design of experiments are used to minimize random and systematic errors. Surrogate models representing the performance of an inverted wing with counter-rotating vortex generators in ground effect are constructed, where design variables of the wing ride height and incidence and the response of sectional downforce are examined. A cokriging regression containing 12 experimental and 25 computational data points sampled with a Latin hypercube design shows the best performance here, capturing general characteristics of the target map well.