Multifidelity optimization methods for interconnected multidisciplinary systems

Hamdan, Bayan

Permalink

https://hdl.handle.net/2142/121233 Copy

Description

Title

Multifidelity optimization methods for interconnected multidisciplinary systems

Author(s)

Hamdan, Bayan

Issue Date

2023-07-13

Director of Research (if dissertation) or Advisor (if thesis)

Wang, Pingfeng

Doctoral Committee Chair(s)

Wang, Pingfeng

Committee Member(s)

• Krishnan, Girish
• Allison, James
• Tran, Huy Trong

Department of Study

Industrial&Enterprise Sys Eng

Discipline

Industrial Engineering

Degree Granting Institution

University of Illinois at Urbana-Champaign

Degree Name

Ph.D.

Degree Level

Dissertation

Keyword(s)

• Multifidelity modeling
• surrogate modeling
• multidisciplinary design optimization.

Abstract

This dissertation presents novel solution schemes for dynamic, multi-stage modeling of complex interconnected systems. For complex systems with partial observability and limited stochastic data, tools that maximize the utilization of the available data and implement adaptive sampling techniques are needed to enhance the selection of sampled data as well as the understanding of the system. In this work, such tools are proposed and applied to realistic case studies that highlight the importance of their use and the applicability of the methods proposed. The main contribution of this work is to create an online, dynamic monitoring tool that can continuously update the predicted health of complex systems and conduct inference across the entire system. In order to accurately model dynamic systems with highly stochastic parameters, multi-stage modeling is used. However, multi-stage modeling constructs a model based on the understanding of the uncertainty in that point of time. This is not always applicable for dynamic, long-term systems, where the understanding of the uncertain parameters changes with time. A novel, Dynamic Multi-Stage Design framework is proposed to incorporate new information for multi-stage models and coordinate previous and future decisions. The proposed framework is implemented on a realistic IEEE 30-bus system where the stochastic input is updated over time. The results show a benefit to updating the scenarios for stochastic applications. The results also show that as the stochasticity of problems increase, so does the benefit of dynamically updating the discretization of the stochastic parameters and coordinating the decisions made over stages. Moreover, to illustrate the applicability of the proposed framework, a simulation-based battery pack immersion cooling application was implemented. A finite element (FE) model was designed in COMSOL Multiphysics software and used to develop a surrogate model that could be used in the proposed framework. The temperature readings of the battery were introduced to the model dynamically to influence the design of the battery pack as well as the coolant flow rate. The case study highlights the benefit of using the proposed framework for control co-design of physical systems as well as numerical models. A major limitation to accurately representing highly dynamic systems is the high computational burden required to optimize the models. This computational burden is due to the introduction of boundary conditions, decision variables and constraints. Although their addition can improve the representation of the original system, these models can become too costly to be practical. Many tools exist to mitigate the high computational cost of complex systems models, such as Multi-Disciplinary Design Optimization (MDO) methods. However, such methods rely on extracting easily decomposable disciplines or subproblems from the original problem, which can be difficult in systems that are not perfectly modular. The benefit of MDO methods is that by decomposing the problem, the subproblems become more manageable to solve and parallelization techniques can be used. To facilitate the use of such methods to non-decomposable model structures, a meta-heuristic decomposition solution scheme is proposed to relax the complicating connections across disciplines in the model structure and solve a decomposition-based scheme. The proposed scheme is implemented on benchmark MDO problems to test the benefit of the proposed scheme over traditional MDO solution schemes given comparable decomposition structures. The results show that the proposed scheme can improve the computational time, as well as the optimal solution value obtained. The proposed meta-heuristic solution scheme is beneficial to handle interconnected subproblems in complex system models; yet, if the complicating connections between the subproblems are implicit, then new tools are needed to handle the coordination between the subproblems. Surrogate modeling methods can be used to represent unknown functional forms mathematically, this can then enable the use of MDO techniques for decomposition through the introduced meta-heuristic. However, when very limited information is known for these black-box connections, all available information should be leveraged to accurately represent the function. An adaptive, online extended multi-fidelity surrogate modeling approach is introduced, AO-MFNets, that can incorporate heterogeneous data and fuse it using an underlying Bayesian Network scheme to relate the different data sources to each other. This method can accurately estimate implicit functions with limited function evaluations. The proposed extension is tested against existing multi-fidelity functions and shows a benefit in functional representation, especially when data is incorporated dynamically. The proposed surrogate modeling method is tested on benchmark numerical examples and shown to provide a lower model estimation error than existing methods. An adaptive scheme is also proposed to select the next best function evaluation to reduce the error. The methods and frameworks introduced in this work can all aid in accurately modeling intricate and complex systems. However, a comprehensive dynamic modeling method is needed to monitor the state of complex, partially observable systems that are comprised of many components. These tools are vital for systems that cannot be continuously monitored and whose component interactions are not always known. In this work, a Multi-Scale Multi-fidelity Bayesian Learning (MMBL) tool is introduced to dynamically incorporate heterogeneous data and use it to update both component-level, as well as system-level state prediction. The proposed tool uses a Dynamic Bayesian Network (DBN) to relate the different components of the system and conduct inference across the network. It also interfaces with the proposed AO-MFNets multi-fidelity surrogate modeling tool to utilize available data and update the state predictions for observable and partially observable nodes. The proposed MMBL tool is applied to a case study of an off-shore production well with many components. Commercial failure rates are used for the components initially, then the failure rates are updated based on the AO-MFNets method given different sources of failure information for one of the components. This new information is then sent to the DBN to update the predictions of the state of components over the system and the probability of system leakage. The proposed model was validated against the literature and was shown to both accurately portray the system, as well as incorporate new model capabilities, such as inference, in addition to the literature. Model reduction schemes were also evaluated to facilitate solving large-scale problems using this method. Applying these schemes was shown to drastically reduce the computational cost of the problem. The set of tools and methodologies proffered by this research facilitate the modeling of interconnected, complex systems that are dynamic and partially observable in nature. The tools proposed have been studied and tested using both, numerical examples for tractability, in addition to realistic case studies to emphasize their applicability. The ideas presented have been discussed in this research and summarized along with future work directions.

Graduation Semester

2023-08

Type of Resource

Thesis

Handle URL

https://hdl.handle.net/2142/121233

Copyright and License Information

Copyright 2023 Bayan Hamdan

Owning Collections