Keeping the lights on in a decarbonising world: physics, uncertainty, time, and how AI can help

Decarbonising the electricity supply requires rapid, large-scale integration of renewable energy across transmission, distribution, and consumer premises, most of it connected through power-electronic inverters. These inverter-based resources (IBRs) behave very differently from the rotating machines of conventional power stations, and they are fundamentally changing how the power system behaves. Demand is changing too. Consumers are becoming active participants through electric vehicles, heat pumps, and smart devices, while AI data centres add large loads with demand patterns the system was never designed for.

In this talk, I argue that these changes lead to challenges to keep a stable and reliable electricity system. The challenges are along three dimensions: physics (as IBRs introduce faster and less familiar dynamics), uncertainty (arising from variable generation and the behaviour of millions of flexible devices), and time (as decisions must be made faster and at finer resolution). Recent blackouts show what can happen when decision-support tools designed for a conventional system, together with inconsistent planning and regulatory practices, fail to keep pace.

Drawing on my research, I discuss how AI, in the broad sense rather than large language models alone, can help. Neural networks trained with the physics of the grid built in can learn to solve large optimisation problems, such as how to run the system securely at least cost, far faster than conventional methods. Data-driven methods can also help quantify risk when the future is uncertain. Finally, I ask what it takes to trust AI-driven decision support, so that we can keep decarbonising the electricity system while keeping the lights on.

Behzad Kazemtabrizi is Professor of Electrical Engineering in the Department of Engineering at Durham University. His research brings together power systems modelling, mathematical optimisation, machine learning, and uncertainty quantification to create fast, dependable, and scalable computational tools for renewable-rich, flexible electricity networks. His current focus is physics-informed AI and surrogate models that accelerate computationally demanding problems such as optimal power flow, security-constrained optimal power flow, unit commitment, and stability assessment, whilst respecting the physical and operational constraints of the power system under study. He has previously developed the Flexible Universal Branch Model (FUBM) and the Convex Flexible Branch Model (CFBM), versatile formulations for power flow and optimal power flow analysis of hybrid AC/DC networks. He has also worked on multidisciplinary projects with industry and with colleagues in Mathematical Sciences and Computer Science. This work has developed new methods for assessing risk and uncertainty in long-term planning decisions and for supporting the economic and reliable integration of offshore wind. He has a successful track record of supervising several PhD students and has been involved as an investigator in research projects worth more than £5M to date.