Rooftop solar, electric vehicle charging and heat pumps are pushing low-voltage distribution grids past limits they were never designed for. Managing that means curtailing generation or demand when the network is about to be overloaded, and the operator has to do it with sparse observability, noisy measurements and a grid model that does not quite match reality.

Prior reinforcement learning approaches learn the whole thing end to end, mapping observations directly to curtailment actions. This work separates detecting congestion from deciding what to do about it.

Decoupling buys something specific in a safety-relevant setting. You can inspect why the system believes there is a problem separately from whether its response is sensible, and an operator who has to justify curtailing a customer's solar output needs exactly that separation. It also means a detection failure and a control failure are distinguishable rather than presenting as one opaque wrong action.