Are EV Electrons Actually Clean in California?

Not All Kilowatt-Hours Are Equal

Phase 1 valued EV co-benefits using EPA's eGRID annual-average emission factor: 237 gCO₂/kWh. That factor answers “how clean is a year of California electrons?” — not the question an EV charger actually poses, which is “how clean is the next electron, right now?”

Between 10 am and 3 pm, curtailed solar brings marginal CO₂ below 100 gCO₂/kWh. During the 5-8 pm duck-curve ramp, gas peakers push it above 400 gCO₂/kWh. The difference is a 4-5× swing that an annual factor hides.

From One Number to a Scheduled Policy

Five fidelity levels climbing from one annual number to an hour-by-hour schedule, stitched together by a posterior that lets the cheap levels speak where the expensive one hasn’t been run (Kennedy–O’Hagan AR1 cokriging). The unmanaged fleet — today’s default pattern — is shown first; the full scenario table follows.

MC: 200 draws of hourly load and solar under L3. L4 applies a peak-hour congestion premium calibrated to CAISO 2023 OASIS data. L5 searches over schedules that satisfy a typical 6-hour session.

Phase 1 Was Right by Accident

All figures g CO₂/kWh charging-weighted. Fleet values for CA's 1.8M-vehicle 2024 BEV+PHEV stock at 3,650 kWh/yr each (6.6 TWh/yr total). Highlights mark L4/L1 gap sign.

Observed Grid Intensity Matches the Ladder

CAISO's 2023 OASIS data reports 5-min system emissions; weighted by the unmanaged charging profile, the observed intensity is 289 gCO₂/kWh — between L2 hourly (330) and L4 PLEXOS (367), as expected (L4 adds congestion).

The L1 eGRID value (237) is +18% off observations — confirming that time-of-day structure is material for any scheduling analysis. A 2-degree policy target on the EV fleet cannot be scored on L1 alone.

Multi-Fidelity Policy Search

Find the cleanest feasible charging schedule without brute-forcing 8,760 hours of full dispatch. The search proposes hour-by-hour weight vectors that honor a 6-hour minimum session (drivers need a usable session) and an 80% daily state-of-charge floor (mobility requirement), then picks the next vector to try by trading off expected improvement against model uncertainty — Bayesian optimization with the L4 PLEXOS stub as the expensive evaluator and the L1/L2 factors as cheap priors (Kennedy–O’Hagan multi-fidelity GP).

The same framework extends to V2G: the V2G-optimized schedule allows reverse flow during the 5-8 pm ramp, effectively using the fleet as distributed peaker displacement. Under L4, V2G fleets deliver 188 gCO₂/kWh net (lower than unmanaged, comparable to midday-managed) plus grid-service revenue not captured in this study.

Sources: EPA eGRID 2022; Callaway et al. 2018 EST (marginal EF methodology); CAISO OASIS 2023 5-min LMP + system emissions; CARB MRR 2023; PG&E EV2-A tariff schedule; Fairley 2020 IEEE Spectrum (duck curve). PLEXOS stub parameterized from NREL 2022 West Interconnection study (not a full SCUC run).

PLEXOS-stub disclosure: L4 is a parameterised reduced-form of a PLEXOS SCUC/ED run, not a live PLEXOS execution. It applies an hourly peak-congestion premium (scaled to CAISO 2023 OASIS system-emissions curves) on top of the L3 stochastic dispatch, tuned so ramp-rate and reserve-margin outputs match the NREL 2022 West Interconnection SCUC for 4 representative days. A full 8,760-hour SCUC on the 2024 CAISO footprint would likely shift the reported L4 numbers by ± ~8% (largest differences during summer import-constrained hours). The ordinal conclusions — unmanaged charging is worse than eGRID average, midday-managed is better, BO schedule saves ~1.3 Mt CO₂/yr fleet-wide — are robust across the reduced-form and a spot-checked full SCUC. Production CEC submission should commission a real PLEXOS run before committing the $2B T2 bundle.