Scientific machine learning for predictive modelling and optimisation of physical systems
File(s)
Author(s)
Li, Siyi
Type
Thesis or dissertation
Abstract
The application of machine learning to physical systems has emerged as a powerful paradigm for accelerating scientific discovery and enabling practical optimisation at scale. This thesis presents three contributions to scientific machine learning for predictive modelling and optimisation of physical systems, spanning wind energy and electric vehicle charging.
The first contribution introduces a graph neural network surrogate model for wind turbine wake modelling, trained on high-fidelity Reynolds-Averaged Navier-Stokes (RANS) computational fluid dynamics data. By operating directly on unstructured meshes, the proposed model eliminates the need for interpolation onto uniform grids, preserving the multi-scale spatial structure of the flow field. The trained model achieves 99.71% accuracy in predicting velocity flow fields and enables farm-scale power prediction within seconds, compared to many CPU hours required by traditional CFD simulations.
The second contribution proposes a graph transformer surrogate model for wind farm power prediction and yaw steering optimisation. Encoding wind farms as fully connected graphs and processing them through a graph transformer, the model achieves 99.8% accuracy on unseen farm layouts and wind conditions, and identifies optimal yaw angle configurations with more than 99% accuracy when coupled with genetic algorithms, while reducing optimisation time by over an order of magnitude.
The third contribution presents a deep learning framework for predicting electric vehicle direct current fast charging profiles and durations from minimal real-world input data. Trained on over 900,000 charging sessions, the model achieves 95\% accuracy in predicting charging duration using six data points collected within five minutes of session initiation, with uncertainty quantification and real-time updating capabilities.
These contributions demonstrate that carefully designed machine learning architectures, informed by the structure of the underlying physical problem, can achieve both high accuracy and practical deployability across diverse physical systems.
The first contribution introduces a graph neural network surrogate model for wind turbine wake modelling, trained on high-fidelity Reynolds-Averaged Navier-Stokes (RANS) computational fluid dynamics data. By operating directly on unstructured meshes, the proposed model eliminates the need for interpolation onto uniform grids, preserving the multi-scale spatial structure of the flow field. The trained model achieves 99.71% accuracy in predicting velocity flow fields and enables farm-scale power prediction within seconds, compared to many CPU hours required by traditional CFD simulations.
The second contribution proposes a graph transformer surrogate model for wind farm power prediction and yaw steering optimisation. Encoding wind farms as fully connected graphs and processing them through a graph transformer, the model achieves 99.8% accuracy on unseen farm layouts and wind conditions, and identifies optimal yaw angle configurations with more than 99% accuracy when coupled with genetic algorithms, while reducing optimisation time by over an order of magnitude.
The third contribution presents a deep learning framework for predicting electric vehicle direct current fast charging profiles and durations from minimal real-world input data. Trained on over 900,000 charging sessions, the model achieves 95\% accuracy in predicting charging duration using six data points collected within five minutes of session initiation, with uncertainty quantification and real-time updating capabilities.
These contributions demonstrate that carefully designed machine learning architectures, informed by the structure of the underlying physical problem, can achieve both high accuracy and practical deployability across diverse physical systems.
Version
Open Access
Date Issued
2026-04-08
Date Awarded
2026-08-01
Copyright Statement
Attribution-NonCommercial-ShareAlike 4.0 International Licence (CC BY NC-SA)
Advisor
Piggott, Matthew
Arcucci, Rossella
Publisher Department
Department of Earth Science & Engineering
Publisher Institution
Imperial College London
Qualification Level
Doctoral
Qualification Name
Doctor of Philosophy (PhD)