Credible CO2 Comparisons: A Machine Learning Approach to Vehicle Powertrain Assessment

TL;DR

This work tackles the challenge of credibly comparing CO2 emissions across ICEVs and EVs under real-world driving by introducing a cross-domain translation framework that preserves the driving context while isolating technology-specific responses. It decomposes each domain into a feature model and an emissions model, yielding a unified instantaneous emissions metric $e_t$ (g/s) and enabling counterfactuals such as determining what an EV would emit under an ICEV trajectory. The approach is instantiated with high-frequency data and LSTM models, validated through a three-stage process including proxy validation, and demonstrates accurate ICEV predictions (MAE around 0.1–0.5 g/s per vehicle) and strong EV proxy performance (MAE ≈ 0.028 g/s for emissions with negligible degradation from predicted inputs). The framework provides a scalable, data-driven basis for credible technology comparisons in real-world conditions, with potential to inform policy and market decisions on vehicle carbon performance. These results establish readiness for cross-domain counterfactual analyses, enabling pointwise and trip-level comparisons without requiring simultaneous testing across technologies.

Abstract

Decarbonizing road transport requires consistent and transparent methods for comparing CO2 emissions across vehicle technologies. This paper proposes a machine learning-based framework for like-for-like operational assessment of internal combustion engine vehicles (ICEVs) and electric vehicles (EVs) under identical, real-world driving conditions. The approach isolates technology-specific effects by holding the observed speed profile and environmental context fixed, enabling direct comparison of powertrain performance. Recurrent neural network models are trained independently for each domain to learn the mapping from contextual driving variables (speed, acceleration, temperature) to internal actuation variables (torque, throttle) and instantaneous CO2-equivalent emission rates. This structure allows the construction of counterfactual scenarios that answer: What emissions would an EV have generated if it had followed the same driving profile as an ICEV? By aligning both vehicle types on a unified instantaneous emissions metric, the framework enables fair and reproducible evaluation of powertrain technologies. It offers a scalable foundation for credible, data-driven assessments of vehicle carbon performance under real-world operating conditions.

Figures (11)

• Figure 1: Stage 1: Domain-specific training pipeline. For each domain (EV and ICEV), an Emissions Model and a Feature Model are trained independently.
• Figure 2: Stage 2: Proxy validation pipeline. The performance of the emissions model is tested when its inputs (torque, throttle) are supplied by the feature model, quantifying the error introduced by the proxy.
• Figure 3: ICEV emissions model training/validation loss for Chrysler Pacifica. Loss (MSE) vs. epoch shows rapid convergence and stable generalization. Final MAE $\approx$ 0.300, MSE $\approx$ 0.497 (validation set).
• Figure 4: ICEV emissions model training/validation loss for Chevrolet Blazer. Training and validation curves remain tightly coupled after the initial transient. Final MAE $\approx$ 0.153, MSE $\approx$ 0.072 (validation set).
• Figure 5: ICEV emissions model training/validation loss for Infiniti QX50. The model stabilizes after the early epochs without signs of overfitting. Final MAE $\approx$ 0.097, MSE $\approx$ 0.046 (validation set).