A novel data-driven approach to enhance analyse flexibility in engineering systems design under uncertainty
File(s)
Author(s)
Caputo, Cesare
Type
Thesis
Abstract
This thesis presents the development of a novel data-driven approach to analyse flexibility in complex engineering systems design under uncertainty, with a focus on Deep Reinforcement Learning. The proposed approach complements existing methodologies particularly for concept generation and design space exploration. The procedure, in contrast to others, allows for the systematic evaluation of more sources of uncertainty and flexibility enablers by exploiting the power of neural network function approximation. The validity of the methodology is incrementally tested through three primary case studies. Firstly, an infrastructure system which exhibits tensions between modular deployment and economies of scale. Results show substantial improvements in economic value compared to all benchmarks. Statistical analysis of the generated flexibility strategy confirm the development of more dynamic and creative solutions than through alternatives Secondly, the design of a waste to energy system is tackled, coordinating decision-making across different plants in the city. Results confirmed significantly improved expected value under uncertainty as compared to benchmarks. A larger decision making space can be considered through the proposed approach, further enhancing performance under uncertainty, and highlighting its value for highly dimensional problems. Accordingly, the first ever design and planning of mobile micro-grid energy systems for nomadic communities is conducted, a highly complex problem. The results show substantial improvements in economic feasibility, system sustainability, and resilience metrics. A novel computational framework for data-driven analysis and formulation of policy tools under uncertainty is also presented and applied to the mobile energy system case study in line with ambitious governmental decarbonization and energy access objectives. Finally, a distributional perspective is considered to more dynamically tailor flexibility strategies to project specific risk-tolerance levels. Several avenues for interpretability are presented, providing a roadmap for practical use. Nonetheless, further work is necessary to continue refining and validating the novel approach to maximize its future impact.
Version
Open Access
Date Issued
2023-11-06
Date Awarded
2024-05-01
Copyright Statement
Attribution-NonCommercial 4.0 International Licence (CC BY-NC)
License URL
Advisor
Cardin, Michel-Alexandre
Del Rio Chanona, Antonio
Sponsor
Engineering and Physical Sciences Research Council
Grant Number
EP/R513052/1
Publisher Department
Dyson School of Design Engineering
Publisher Institution
Imperial College London
Qualification Level
Doctoral
Qualification Name
Doctor of Philosophy (PhD)