MIT researchers have introduced a new problem-solving framework that could significantly speed up power grid decision-making. The approach, called FSNet, is designed to solve the AC Optimal Power Flow (AC-OPF) problem more quickly while still meeting all grid constraints, a requirement for system safety and reliability.
Power operators rely on AC-OPF to calculate the most economical way of dispatching generation without overloading transformers, lines or equipment, and without violating voltage or balance conditions. Conventional solvers can achieve accurate results, but they become slow and computationally costly as networks grow and renewable generation makes grid conditions more dynamic.
FSNet addresses this by splitting the task into two stages. A deep neural network first predicts a potential solution based on historical or simulated data. This initial output is then passed to a feasibility-focused optimisation routine, which adjusts the prediction just enough to satisfy equality and inequality constraints such as power balance, voltage limits, line ratings and generator boundaries.
Benchmark tests on systems ranging from IEEE 14-bus to large multi-hundred-bus models show that FSNet can cut calculation time from hours to minutes or seconds, while maintaining feasibility. Pure machine-learning models, by comparison, produced fast but often infeasible outputs, and traditional solvers were significantly slower. In some complex cases, FSNet also reached higher-quality solutions by avoiding local minima that can delay deterministic solvers.
The researchers note that the framework does not guarantee global optimality, but its speed and reliability make it suitable for real-time control, rapid re-dispatch and integration of variable renewable resources. The method could also be applied to other constrained optimisation domains beyond the power sector, including autonomous systems, finance and logistics.
Source: EE Power
