Credible CO ₂ Comparisons: A Machine Learning Approach to Vehicle Powertrain Assessment

ABSTRACT

Decarbonising road transport requires comparisons transparent about what is controlled. We present a machine‐learning framework for operational assessment of internal‐combustion‐engine vehicles (ICEVs) and electric vehicles (EVs), on a common instantaneous ‐equivalent scale. Rather than treating equal speed as proof of identical operation, the estimand is a comparison conditioned on the observed shared covariates. Data include speed, acceleration, temperature, and actuation channels, but not road grade, payload, wind, driver identity, gear position or transmission state; hence, the study is a methodological framework and readiness assessment, not a causal cross‐powertrain comparison. All experiments use complete‐trip train/validation/test splits before window construction, training‐only scaling, fixed split seed 20260801 and validation‐selected checkpoints. The long short‐term memory (LSTM) serves as a recurrent reference because both context‐to‐actuation and actuation‐to‐emissions mappings may depend on operating history; the framework remains model agnostic and benchmarks recurrence against validation‐selected non‐recurrent estimators. A two‐stage search screens 72 LSTM configurations, refines 15 candidates and evaluates one task‐specific configuration over five seeds for four emissions tasks and the BMW i3 context‐to‐actuation task. The LSTM is numerically lower than the selected non‐recurrent baseline for BMW i3 and marginally lower for QX50, whereas random forest is lower for Blazer and ridge is substantially lower for Pacifica. In the EV feature task, LSTM is lower for torque and histogram gradient boosting is marginally lower for throttle. Pairing feature and emissions checkpoints by seed shows that replacing measured actuation with predicted torque and throttle increases mean trip‐level emissions MAE from to g/s. The mean increase is g/s, with a paired hierarchical 95% bootstrap interval of [0.00871, 0.01245]. The mixed benchmark results support estimator selection by dataset, task, output and temporal context rather than universal recurrent superiority, and the proxy step remains a statistically resolved source of uncertainty.