Abstract
Accurate vehicle-level CO₂ emission prediction is essential for regulatory planning and greener powertrain design. This study presents an integrated FA-ACO-LSTM framework for estimating vehicle-level tailpipe CO₂ emissions from standard vehicle descriptors. The contribution lies in combining leakage-aware Binary Firefly feature selection, mixed-variable FA-ACO hyperparameter tuning, repeated-run evaluation, and post hoc interpretability into a unified prediction pipeline. Using a public fuel-consumption dataset sourced from the official Canadian vehicle fuel-consumption and emissions registry (open government portal) (22,556 vehicles), we adopt a 70/20/10 train/validation/test split and perform wrapper-based feature selection with Binary Firefly Algorithm (BFA). The FA-ACO scheduler jointly tunes hyperparameters for five candidate learners (LSTM, CNN, GRU, MLP, TabNet) and selects the best. FA-ACO-LSTM attains MSE = 0.0099, MAE = 0.0791, MedAE = 0.0664, MAPE = 0.84%, and R² = 98.71% on the test set, surpassing FA-ACO-CNN (R² = 93.55%), FA-ACO-GRU (90.03%), FA-ACO-MLP (87.60%), and FA-ACO-TabNet (84.88%). Compared with untuned baselines, FA-ACO improves LSTM from R² = 96.53% to 98.71% and reduces MSE by 45%. BFA yields the lowest average error among feature selectors versus BMWO, BFO, and BGWO, and an ANOVA across models confirms significance (F = 30.2, p < 0.0001). These results represent a fuel-consumption-informed prediction scenario because the selected feature set includes COMB (L/100 km), HWY (L/100 km), and COMB (mpg), which are strongly associated with certified CO₂ emissions. When all fuel-consumption indicators were excluded before feature selection and training, model performance decreased from [Formula: see text]to [Formula: see text]. The ablation confirms that fuel-consumption proxies account for a substantial part of the main predictive performance, while non-consumption vehicle attributes retain meaningful predictive value, and indicates that, under the adopted fixed feature-ordering representation and FA-ACO tuning protocol, LSTM achieved the best performance among the evaluated neural and tabular baselines. However, because the dataset is cross-sectional rather than temporal, the LSTM should be interpreted as a feature-interaction learner rather than a temporal forecasting model.