Martin Doff-Sotta

Future power grids must operate reliably despite the uncertainty from renewable energy and flexible demand. A new generation of smart-grid devices—including smart transformers, storage, and converters—can regulate voltage, support frequency, and absorb fluctuations, enabling greater renewable integration. Unlocking this potential requires fast decision-making under uncertainty, which current methods cannot deliver. The leading framework, stochastic Model Predictive Control (MPC), makes safe automated decisions by solving optimisation problems online. Yet, in large nonlinear grids, these problems are too complex to solve efficiently. This project overcomes this with novel inputconvex neural networks that represent nonlinear dynamics as difference-of-convex functions, transforming intractable optimisation into convex programs solvable efficiently with existing solvers.  Another challenge is forecasting, required by decision-making algorithms. State-of-the-art foundation models such as Chronos predict time-series with zero-shot accuracy on unseen datasets but typically rely only on past demand. Chronos will be extended to incorporate weather and time, yielding richer MPC predictions. While international research increasingly applies AI to energy systems, most approaches lack formal safety guarantees. This project differs by combining the performance of AI with the rigour of control theory—delivering algorithms that are provably-safe and scalable, enabling resilient grids capable of integrating much larger shares of renewable energy.