The policy buys reserve before the outage. That can raise cost or carbon on the nominal day while reducing critical loss across the forecast branches.
+10.9 vs greedy
+13.2 vs greedy
-3.2 vs greedy
60.9 → 2.1
One day, every balance exposed.
Six branches. No dice roll.
Each branch is a deterministic replay with an explicit probability and profile multiplier. The controller uses the same rules in every branch. This is a stress ensemble, not a stochastic MILP.
- Expected policy cost
- €69.7
- Worst critical service
- 86.4%
A policy you can inspect, not an optimizer you have to trust.
GridAtlas closes an energy balance every 15 minutes. Battery charge, discharge, efficiency, power, and state limits are explicit. During a grid outage, import and export are forced to zero and any remaining deficit becomes unserved load.
The look-ahead policy scores future intervals with price, carbon, and outage reserve weights. It uses quantile thresholds and a rising reserve target. There is no mixed-integer solver, unit commitment, or optimality certificate.
All time series are modelled. Outages and forecast deviations are synthetic. No observed measurements ship with this repository.