RESEARCH · RESEARCH · #2314
Learned control variates for rare grid-outage estimation at EV charging stations
The paper proposes using a learned binary classifier as a control variate to reduce variance when estimating very small feeder-overload probabilities caused by public EV charging. The authors derive variance-reduction as an explicit function of classifier recall and false-positive rate, show learned CVs remain effective in extreme rare-event regimes where Gaussian load proxies fail, demonstrate ~25% variance removal with a weak single-feature learner and ~60% with boosted trees across three orders of outage probability, and report that applying the learned CV to Surabaya's feeder network reduces city-wide blackout days by one-third versus plain Monte Carlo at the same simulation budget.
KEY POINTS
- The paper proposes using a learned binary classifier as a control variate to reduce variance when estimating very small feeder-overload probabilities caused by public EV charging.
- The authors derive variance-reduction as an explicit function of classifier recall and false-positive rate, show learned CVs remain effective in extreme rare-event regimes where Gaussian load proxies fail, demonstrate ~25% variance removal with a weak single-feature learner and ~60% with boosted trees across three orders of outage probability, and report that applying the learned CV to Surabaya's feeder network reduces city-wide blackout days by one-third versus plain Monte Carlo at the same simulation budget.
- ML-based control variates can substantially cut expensive simulator runs and give planners more accurate, computationally affordable estimates of very rare but high-impact grid outages driven by EV charging.
WHY IT MATTERS
ML-based control variates can substantially cut expensive simulator runs and give planners more accurate, computationally affordable estimates of very rare but high-impact grid outages driven by EV charging.