Abstract
High operational costs and significant energy penalties remain the primary
barriers to the widespread deployment of carbon capture, utilisation and storage.
To achieve economic viability in a net-zero energy system, carbon capture
plants must transition from steady-state baseload operation to flexible dispatch,
allowing energy consumption to adapt to time-varying electricity prices. However,
existing modelling paradigms struggle to meet these combined economic and
technical requirements: mechanistic models remain computationally demanding
for real-time optimisation, while conventional data-driven models lack the
physical reliability required for safety-critical operation. This thesis argues that
one of the key barriers to the industrial deployment of artificial intelligence
in this context is not only predictive accuracy, but structural trust. Initial
studies within the thesis demonstrate that while deep learning architectures
offer superior predictive power compared to classical baselines, they effectively
function as black-box models. This limitation was quantified when evaluating
model performance on industrial data from a pilot plant. The results exposed a
feasibility gap: while linear models proved computationally efficient and robust
to noise, they inherently lacked the capacity to capture complex non-linear
thermodynamics. Conversely, unconstrained deep learning exhibited stochastic
instability and increased computational cost, limiting their suitability for real-
time control-oriented applications. To address this conflict between linear
robustness and non-linear fidelity, this thesis develops a novel physics-informed
spatio-temporal graph neural network. This architecture extends the modelling
framework by embedding the plant’s process flow diagram directly into the
graph topology. The findings suggest that aligning algorithmic topology with
plant structure can improve the reliability and interpretability of data-driven
models, supporting the development of physically consistent digital twins for
carbon capture systems.