The rapid advancement of artificial intelligence (AI) in engineering has created both opportunities and challenges for complex system modelling and control. Data-driven approaches can approximate complex behaviour but may lack physical interpretability and generalisation, whereas physics-based methods can be too computationally demanding for real-time use. This thesis develops a unified AI-orchestrated physics-informed modelling and predictive control framework integrating data-driven learning, physical knowledge, and optimisation-based decision-making. By embedding engineering knowledge directly into learning and control processes, the framework seeks to improve reliability while retaining the computational advantages of AI. It is developed progressively across energy-intensive manufacturing and battery management, covering steel reheating, single-cell charging, and multi-cell pack balancing.
Grounded in this framework, one primary contribution in this thesis is the real-time prediction of transient heat transfer in steel reheating furnaces, where nonlinear partial differential equations, limited measurements, scarce industrial data, and changing operating conditions challenge both black-box and numerical approaches. A physics-informed deep learning framework, termed EngGeneNet, is developed for 2D temperature-field prediction in large-scale intermediate steel stock. It combines a convolutional neural network (CNN) with a Crank--Nicolson-based physics-informed loss, an Engineering-Gene (Eng-Gene) module embedding the temperature-dependent Fourier number, and knowledge distillation with transfer learning. These components allow the model to use physical structure both during training and when generalising to previously unseen operating conditions. Validation on a pilot-scale walking-beam bloom reheating furnace shows accuracy comparable to the conventional alternating direction implicit (ADI) method, while reducing per-frame computation time by 96%. EngGeneNet therefore enables real-time prediction suitable for closed-loop furnace control.
Beyond modelling, a further contribution in this thesis is the battery charging control process optimisation, where charging speed and internal-temperature safety are competing objectives and historical trial data are often underused. A historical data-driven self-learning model predictive control (MPC) framework combines the trial-to-trial learning capability of iterative learning control (ILC) with MPC constraint handling. A convex mapping constraint, constructed from the convex hull of historical state trajectories, narrows the feasible solution space, while a memory function-enhanced cost function uses previous charging iterations to refine the current trajectory. This enables the controller to improve performance progressively without relying on unsafe exploratory interactions with the battery. In a lithium iron phosphate (LiFePO4) case study, the proposed method reduces charging duration by 70--100~s relative to an existing constrained generalised predictive control benchmark, strictly satisfies internal temperature constraints, and reduces solve-time variability by approximately 80% relative to standard MPC.
Finally, the framework is extended from single-cell control to multi-cell battery-pack balancing, where state-of-charge (SOC) equalisation, thermal regulation, and output-voltage stabilisation must be achieved simultaneously under coupled electrical and thermal dynamics. A 2D cost map-guided parallel-inference neural modelling and predictive control framework is proposed. A unified multi-task neural model predicts coupled states under candidate balancing actions, while an inertia-aware 2D cost map represents the different time constants of electrical and thermal states. This formulation permits fast comparison of multiple candidate actions while accounting for the distinct responses of electrical and thermal variables. Parallel Gaussian sampling and softmax-weighted aggregation determine the control action without gradient-based optimisation. Simulations under three real-world driving-cycle profiles achieve an inter-cell temperature difference below 5 degrees celsius, terminal SOC non-uniformity below 1%, voltage regulation within a 10% error band, and a per-command decision time of 0.121s.
Taken together, the proposed frameworks demonstrate that integrating physical laws, engineering knowledge, and data-driven learning can balance model fidelity, computational efficiency, interpretability, and adaptability. The progression from physics-informed modelling to historical-data-driven control and parallel multi-objective decision-making provides generalisable solutions for complex engineering systems characterised by coupled multi-physics dynamics, limited data, repetitive operation, and stringent real-time constraints.
