In the net-zero transition of the UK by 2050, electricity demand will increase and more renewable power generation will be installed in industrial plants. The bulk electricity system also faces the challenges of increased total and peak demand, increased difficulty in balancing supply and demand, and increased network issues. The flexibility of industrial plants, i.e., the ability to change the normal electricity generation/consumption patterns, can be utilised to address these challenges, through improving the utilisation of renewable power generation onsite and providing balancing and network services to the bulk electricity system. However, the scheduling and control for tapping this flexibility are subject to great difficulty due to significant uncertainties and computational complexity.

Digital twins are systems of advanced sensing, communication, simulation, optimisation and control technologies, and can provide updating system states and prediction, based on which data-driven approaches can be developed to tackling the uncertainties and computational complexity in scheduling and control. Specifically, a kernel-learning based method is proposed to characterise the uncertainty sets, and an artificial neutral network based method is proposed for predictive control of industrial plants in real-time operation.

A test digital twin platform is established in the lab to demonstrate and assess the proposed data-driven solutions. The platform adopts a two-level structure, with the upper-level global digital twin for whole-plant level predictive control and lower-level local digital twins representing industrial processes, renewable power generation and energy storage systems. The measurements are taken from sensors or a data generator which produces mimic data flow. Two industrial case studies with real data are tested on the platform. One case is an industrial site with a number of bitumen tanks and PV panels, and the other is a paper mill with onsite wind turbines and battery storage.