4.1 Data gathering and collection

The publicly accessible datasets provided by NRCan have been consistently releasing annual performance statistics for light-duty cars in Canada for more than ten years. Covering a wide range of vehicle technologies available in the market, NRCan regularly updates these datasets, with corresponding information on several technical aspects, such as vehicle manufacturers and models, engine size, fuel type, and transmission system type. In addition, emissions and consumption data are also provided in the dataset, which are derived through a 5-cycle standard approved by the Government of Canada. Using a chassis dynamometer, the derived data for fuel consumption and CO₂ emissions for drivetrains are collected following controlled laboratory protocols to ensure comparability across vehicle technologies and model years. Unlike European standards such as the ECE-15, the EUDC, and the combined NEDC, which are somewhat modal and simplified, the NRCan standard driving cycles incorporate a broader range of driving conditions, including city and highway driving, cold temperatures, air-conditioning loads, and high-speed/aggressive behaviour, making them a credible dataset and ideal foundation for developing predictive studies [60]. NRCan can compile and validate this information before the release of annual datasets, making it one of the most comprehensive and credible sources for fuel consumption and emission benchmarking in Canada.

As mentioned earlier, the compiled “Fuel consumption rating” open databases are obtained from the official website of the Canadian Ministry of Transportation, where two datasets (for ICE and PHEV) are measured to provide specific vehicle information using mixed data types of both numerical and categorical categories and encompassing vehicle information, including the make, model, engine details, transmission type, fuel consumption rates, and CO₂ emissions [61]. These vehicles usually weigh 10,000 pounds or less and are primarily used for transporting passengers and cargo, including cars, vans, sport utility vehicles (SUVs), and pickup trucks. The original compiled data for consumption and CO₂ emissions were collected from vehicle manufacturers through laboratory standardized testing procedures, thus simulating real-world driving scenarios [35].

In this work, the authors combined two sets of ICEV and PHEV data from 2017--2025, forming one dataset containing 11 attributes and 7737 records, which differentiates this work from other related works in the literature by offering detailed insights into vehicles presently available in the Canadian market, covering both traditional combustion and electric vehicles. This merging provides a more comprehensive and holistic view of policy decisions and network planning to determine total fuel demand, which is fundamental for transportation infrastructure development, as well as identifying the level of carbon emissions produced by both vehicle categories to verify the role of PHEVs in fulfilling sustainability goals.

4.2 Exploratory data analysis

Dataset exploration is a vital step for data understanding and preparation before analysis; thus, data analysis, data visualization and correlation analysis are carried out in this subsection. Table 2 presents the descriptive statistics of the numerical features of the studied dataset, which facilitates understanding the data distribution and serves as an indicator of dataset quality. On the basis of the data provided, the average total fuel consumption is 10.639 L/100 km, of which 12.4 L/100 km are from the city, whereas the remaining 9 L/100 km are from highways. Additionally, the average automobile CO₂ emissions are 243.833 g/km, with a corresponding standard deviation of 66.2 g/km. Furthermore, Table 2 illustrates a wide range of existing engine sizes, ranging from 0.6 litres to 8.4 litres, which are expected to be significant factors influencing both emissions and consumption prediction. In addition, CO₂ experiences large variations, reflecting the existence of both efficient and high-emission vehicles due to the fusion of the ICEV and EV datasets. Moreover, the standard deviation (std) indicates that the expected value distributions of both total consumption and emissions are well defined, thus enabling accurate forecasting.

Download:

• PNG
  larger image
• TIFF
  original image

Table 2. Descriptive statistics of the combined datasets.

https://doi.org/10.1371/journal.pone.0342418.t002

Fuel consumption and CO₂ emissions are presented in Figs 2 and 3, respectively. Luxury and high-performance ICEV vehicles have the highest energy demand and emission profile among all brands. High-performance ICEV vehicles are the highest among all brands in terms of energy demand and emission profile.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 2. Average fuel consumption per vehicle brand.

https://doi.org/10.1371/journal.pone.0342418.g002

Download:

• PNG
  larger image
• TIFF
  original image

Fig 3. Average CO₂ emission per vehicle brand.

https://doi.org/10.1371/journal.pone.0342418.g003

A correlation heatmap using Pearson’s correlation coefficients is presented in Fig 4 to visualize the relationships between features and help identify feature combinations with strong correlations among all the quantitative numerical variables, such that all these variables range from highly positive to highly negative. Since fuel consumption and CO₂ emissions are the dependent variables, it is concluded from this heatmap that fuel consumption in both cities (fuel_consumption_city) and highways (fuel_consumption_hwy) is highly correlated with combined fuel consumption (fuel_consumption_comb) as well as CO₂ emissions (co2_emissions), followed by engine size (engine_size) and cylinder (cylinders). Similarly, there is a strong positive relationship between the same variables and CO₂ emissions. Additionally, some other general observations could be drawn from the feature map, such as that larger engines and more cylinders usually imply higher fuel consumption and emissions. Additionally, the observed strong correlations among several numerical predictors indicate the presence of multicollinearity, which may be useful for the selected ML models to learn their joint influence on fuel consumption and CO₂ emissions. Furthermore, the observed scale difference between correlated variables suggested the use of standardization in the model preprocessing stage to ensure balanced model training.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 4. Heatmap of quantitative independent variables.

https://doi.org/10.1371/journal.pone.0342418.g004

4.3 Data preprocessing

In this section, data are prepared for a subsequent phase of model training and building. To ensure that the dataset meets the model requirements and achieves accurate fitting and prediction results for both fuel consumption and carbon emissions, preprocessing for the dataset is performed, establishing the following objectives:

1. i. Data cleaning: Missing, null and duplicate values in the dataset leading to bias, skewness or inaccuracies in model training and outcomes must be removed. After determining whether any entries in the data were duplicates, null or missing, they were either eliminated, in the case of duplicates, or replaced with nonnull entries, in the case of missing entries. For categorical data, missing/null entries are filled with dataset mode, whereas mean values are used for numerical data.
2. ii. Feature Handling, Data Transformation & Standardization: To preserve information related to predicting consumption and carbon emissions during data preprocessing, the considered models must incorporate both numerical features and categorical features handling which can be carried out as follows:
   • Categorical feature handling: To transform categorical variables into numerical representations, one-hot encoding has been utilized for the categorical variables ‘Make’, ‘vehicle_class’, ‘Transmission’, and ‘Fuel Type’. Combined with numerical features, encoded categorical features are subsequently used by models for analysis. The resulting encoded variables are then combined with the numerical features in the dataset to form the final model input feature space. In addition, the first category of each feature is dropped to prevent multicollinearity between variables and bias.
   • Numerical feature scaling: Standardization is applied, as many regression algorithms are sensitive to scale differences between features and ensure that everyone contributes equally to the model’s performance. The numerical features were normalized via a z- score standardized scale, thereby reducing the impact of varying units and magnitudes by maintaining the data distribution shape through centering the data with a mean value of zero and standard deviation of 1. This was particularly crucial for features that have different scales, such as fuel consumption and engine size. The equation for z- scaling is as follows:

(1)

where is the transformed data variable, is the mean and is the standard deviation. In this way, the normalized value of a sample is obtained by the difference between the original value and the average of the considered records of the attribute by the standard deviation of all samples.

1. iii. Data splitting: After all the previous preprocessing stages, the original data were randomly divided into subsets for training and testing at a ratio of 80:20. This ensures reproducibility, such that consistent splitting is maintained across different program runs and prevents data leakage by ensuring that the model’s performance is assessed through unseen data.

4.4 Model building

Predicting fuel consumption and carbon emissions via data-driven approaches is regarded as a multivariate problem that requires regression techniques, as these numerical variables are influenced by multiple input features. These techniques, whether linear regression-based or nonlinear regression-based, use labelled data to create a model that captures the underlying relationships between features such as engine size, number of cylinders, transmission type, and fuel type and the numerical values of specific target variables—fuel consumption and CO₂ emissions. While linear regression models are known for their simplicity and low demand for computational resources, nonlinear models are typically more powerful, accurate and flexible, making them ideal for higher predictive accuracy, especially when dealing with complex, large and nonlinear relationships in the data.

To support data-driven decision-making in the context of sustainable transportation and emission reduction, this work conducts a comparative analysis of several regression techniques that are conducted for ICEV and PHEV data from Canada. The comparison includes linear, nonlinear, and ensembled ML-based models where the main objective is to identify the most effective models that are capable of capturing relationships between vehicle characteristics and their consecutive consumption and emissions. In total, thirteen different predictive algorithms (4 linear, 3 traditional nonlinear, and 6 ensemble ML) are examined in this work. Notably, the ensemble category includes three state-of-the-art techniques: XGBoost, CatBoost, and LightGBM. To increase the accuracy of the predictive models, the model hyperparameters are tuned after training via Grid SearchCV with 5-fold cross-validation. This approach allows the optimal model configuration for each algorithm to be identified through initial exploration of combinations of predefined parameters.

Linear-Based Regression Models:

One of the most recognized and well-known models in statistics and ML is linear regression-based models, of which four different models are investigated, including linear regression, Ridge, Lasso, and Elastic Net, which can be briefly described as follows:

1. a). Linear regression (LR): This is a baseline model indicating associations among variables by best fitting a linear equation to data such that the error difference between the actual and predicted values is minimized [62]. LR is fast and simple yet may exhibit poor generalization performance in the case of large datasets and features [63].
2. b). Penalty-based models include Ridge, Lasso and Elastic net regression models, which are often referred to as regularized versions of LR. Through the introduction of a penalty term, these types of models work better with high-dimensional data than do LR models by imposing constraints on the size of the model coefficients, referred to as shrinkage and regularization. In Lasso regression, the least absolute shrinkage and selection operator is applied, thus controlling model complexity by eliminating unnecessary features [54]. Ridge regression, on the other hand, adds a penalty to the loss function that is equivalent to the square of the magnitude of the coefficients, making it useful when there is multicollinearity between variables [64]. In contrast, Elastic Net aims to capture the strengths of both Ridge and Lasso by combining both penalties [65].

Linear regression techniques minimize the ordinary least squares loss, as specified in (2), where n is the total number of samples, is the observed actual value for the -th sample, is the vector of model parameters, and is the residual prediction error. For penalty-based models, however, this equation is extended by introducing a regularization term, as presented in (3), to control model complexity, such that controls the strength of regularization and represents the penalty function to balance model bias and variance. Ridge, Lasso, and elastic net regression apply L2 according to (4), L1 as in (5), and combined penalties in (6), such that determines the relative weighting between sparsity L1 and coefficient shrinkage L2 in the elastic net formulation.

(2)

(3)

(4)

(5)

(6)

1. i. Nonlinear-Based Regression Models:

In this work, 9 nonlinear ML models are investigated, including 3 traditional nonlinear ML and 6 ensemble ML models, which can be briefly described as follows:

1. a). Conventional nonlinear: This study covers 3 nonlinear ML models, including SVR, GP and DTs. SVR, which is a kernel regression technique, is recognized as one of the most robust and accurate methods among the data mining algorithms [66]. Predictions of continuous variables can be achieved by identifying the hyperplane that optimally maximizes the margin between the data points and hyperplane, producing robust predictions. Similarly, GP models are also a kernel-probabilistic strategy for learning correlations of numerous variables in the data learning set, making them particularly ideal for nonlinear forecasting problems [67]. They make predictions via kernels and offer uncertainty metrics for those predictions via the Bayesian framework [68], making them good candidates for small datasets. DT regression, on the other hand, predicts the target value of a variable by learning decision rules inferred from the input features. Using a group of logical decision rules, data are split into smaller subsets on the basis of feature values, creating a tree-like structure. Splits are chosen to minimize the variance of the target variable; thus, they are suitable for complex and nonlinear data representations [69].
2. b). Ensemble ML models: For ensemble models, both bagging and boosting algorithms are compared in this work, using GB, HGB, RF, with particular emphasis on state-of-the-art boosting architectures such as XGBoost, CatBoost, and LightGBM. Boosting strategies, such as GB, XGBoost, CatBoost and HGB, employ the core principle of sequential learning, where each new model corrects the errors of previous models [51]. GB combines weak learners of DTs to generate a strong learner that can make accurate predictions. To minimize the overall prediction error, the gradient descent algorithm is fitted on the residual errors of the previous learners in a sequential manner, thus achieving high accuracy with a customizable loss function [21]. XGBoost is a tree-based GB model that fits a decision tree to the negative gradient of the objective function in each iteration. Through the addition of a regularization penalty term, the model’s coefficients are altered, which influences the prediction outcome, making them ideal not only for handling missing data but also for faster training owing to their parallel computation capability. On the other hand, the GB approach is applied similarly to HGB but uses a histogram-based approach for binning features, where the dataset’s feature histograms are used to construct decision trees [70], thereby enhancing both computational efficiency and model performance. To better handle categorical variables, CatBoost is another ensembled algorithm specially designed for categorical boosting that improves model accuracy and robustness by mitigating overfitting through ordered boosting while avoiding target leakage during training. Unlike traditional one-hot encoding, which relies on transforming categorical variables into binary variables, CatBoost replaces categorical values with statistically derived values through target-based encoding [71], thus allowing features to be handled without the need for an extensive preprocessing phase, making it more suitable for large datasets with categorical features. Similarly, LightGBM, another gradient-boosting technique, is based on gradient-based one-sided sampling (GOSS), exclusive feature bundling (EFB), and histogram and leafwise growth. These design optimizations increase their training efficiency, enable them to consume less memory and provide predictions with high accuracy, particularly with highly dimensional datasets [72].

In contrast, bagging algorithms such as RF employ parallel learning during data training, build multiple decision trees on different subsets of the data and features, and then aggregate their predictions. This reduces both variance and bias in the model and improves the ability of the model to generalize[73]. Compared with single decision trees, predictions of the bootstrapped subsets of data are combined and averaged, making them more immune to overfitting and more robust to noise and outliers, resulting in higher predictive accuracy. Thus, these methods are well suited for datasets with large feature counts (high dimensionality) and can handle missing values in the data without requiring explicit imputation.

4.5 Hyperparameter tuning

To test all the algorithms effectively, all the ML models were trained via the same unified pipeline incorporating preprocessing, model definition, and hyperparameter search, following the flowchart presented in Fig 1(B). For the models where hyperparameter tuning is possible, cross-validated GridSearchCV with 5-fold validation has been implemented. Conversely, standard configurations, such as linear regression and GP, have been adopted for models without parameter grids. After the hyperparameters were determined, each algorithm was retrained using the corresponding set of best hyperparameters obtained from GridSearchCV, thus minimizing the cross-validation error [74].

4.6 Model evaluation

The choice of the most appropriate model with the best performance is usually driven by a trade-off between flexibility and accuracy. While linear models are inherently simple and highly interpretable, their simplicity can be a limitation when dealing with complex, nonlinear datasets, eventually affecting their predictive performance. In contrast, more complex data correlations can be captured through nonlinear ML models but at the expense of greater chances of overfitting. Combining multiple models through ensemble ML, on the other hand, offers a compromise of enhanced predictive accuracy while mitigating the shortcomings of individual linear or nonlinear models, thus offering a balanced, middle ground solution. For these reasons, this section employs a study of the previously mentioned models in an attempt to find the most suitable modelling strategy that will present the best trade-off for predicting energy consumption and carbon dioxide emissions for EVs.

To assess the effectiveness of each model along with its accuracy, consistency, and ability to predict actual values, various performance metrics are analysed. In this energy prediction problem, the performance of each ML algorithm was evaluated via three regression evaluation metrics, namely, the root mean squared error (RMSE), the mean absolute percentage error (MAPE), and the R-square (, as expressed in equations (7–9).

To determine the RMSE, the total number of data points is divided by the square root of the sum of the squared deviations between the predicted and actual values. Thus, the RMSE penalizes larger errors and has the same unit as the target variable, which makes it more interpretable; however, it is sensitive to outliers [75]. Given that is the mean of the actual target values and that is the model’s predicted value, the RMSE can be presented as in (7) by:

(7)

The MAPE aims to quantify the prediction error as a percentage of the actual value of the target variable. Consequently, the error is standardized and presented in percentage values, making it comparable and interpretable across many data points, as depicted by equation (8). This is beneficial in scenarios where the target variables exhibit considerable variability and where prediction mistakes might substantially impact the overall performance of the model [76].

(8)

R-squared, also known as the coefficient of determination, is a statistical metric for evaluating how well a model fits the actual dataset. A value between 0 and 1 is determined by dividing the model’s explained variance by the overall data variance, where a greater number denotes a better fit [77]. Given that R2 reflects how well independent variables explain the dependent variable, R² is also often referred to as accuracy and can be presented as in (9).

(9)

5.1 Energy consumption prediction

Table 4 compares the performance of the 13 tested models; accordingly, their predictive accuracy and robustness are evaluated by collectively assessing each model’s error magnitude, relative prediction deviation, and overall goodness-of-fit. For the linear model category, Ridge regression achieved the best performance compared with the other techniques, with a recorded RMSE of 0.8766, a MAPE of 6.62 and an R² value of 0.9282, followed by linear regression, with similar prediction accuracies. With respect to the nonlinear model category, the DT model performed better than the SVR and GP models did, with the lowest RMSE of 1.32923, MAPE of 3.210663, and the highest R² of 0.966042; moreover, the DT model outperformed all the linear models. For the ensemble ML category, the RF attained the best R² score of 0.9772 and the lowest RMSE and MAPE of 0.5455 and 3.71, respectively, indicating greater generalizability and reliability than the other ensemble ML models and superior performance among all the considered predictive models.

Download:

• PNG
  larger image
• TIFF
  original image

Table 4. Energy consumption prediction results.

https://doi.org/10.1371/journal.pone.0342418.t004

Conclusively, RF stands out as the top performer among all tested algorithms, demonstrating its excellent ability to predict vehicle energy consumption with minimum prediction errors and strong prediction accuracy while capturing the variance in the data as well as verifying the supremacy of ensemble ML models over their linear and traditional nonlinear counterparts.

For a clearer interpretation of the comparative results, Fig 5 illustrates the performance of all the evaluated models across the RMSE, MAPE, and R² metrics. The figure demonstrates the relative positioning of different model families, illustrating how ensemble-based approaches consistently cluster in regions of lower prediction error and higher fitting performance compared with linear and traditional nonlinear models. This visualization complements the quantitative results presented in Table 4 by further showing the overall performance trends rather than individual metric values, thus offering a better intuitive comparison of model robustness and predictive behaviour across all studied model categories.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 5. Fuel consumption performance metrics of all ML models.

https://doi.org/10.1371/journal.pone.0342418.g005

In order to assess model fitting performance, a visual representation of actual versus predicted parity plots are illustrated in Fig 6, where an ideal performance when actual and predicted values align is represented by the 1:1 dashed line. For models with a high degree of accuracy, more clustering should be observed around the 1:1 line, indicating high generalizability. For linear models, performance is moderately aligned with the ideal lines but with noted widening scatter with highly consuming vehicles. Elastic net exhibits the weakest fit among all linear architectures, with increased dispersion confirming the model’s limitations in capturing nonlinear interactions in the dataset and the presence of systematic bias. For nonlinear models, including SVR, GP, and DT, tighter clustering is observed in the parity plots, especially for low- and medium-fuel-consuming vehicles employing the SVR and GP models, thereby indicating better performance capabilities than those of linear approaches. Conversely, DTs show high signs of possible overfitting, which can be observed by the very high dispersion around the ideal parity line, thus highlighting the limited model’s generalization abilities. Ensemble-based approaches clearly exhibit the most consistent parity behaviour, where GB, HSG, XGBoost, CatBoost, and LightGBM show strong predictive alignment with minor deviations in some models. Notably, CatBoost, HGB, and RF exhibit the closest adherence to the ideal 1:1 line, indicating high predictive accuracy and low systematic error.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 6. Actual versus predicted consumption plots: (A) Linear regression, (B) Ridge, (C) Lasso, (D) Elastic Net, (E) SVM, (F) DT, (G) GP, (H) RF, (I) GB, (J) HGB, (K) XGBoost, (L) CatBoost, (M) LightGBM.

https://doi.org/10.1371/journal.pone.0342418.g006

5.2 CO2 emissions prediction

Table 5 shows the results of the prediction of CO₂ emissions for all vehicle types. Again, the three previously considered statistical metrics are utilized to evaluate the performance of the tested three models’ categories, representing 13 predictive algorithms. Among all linear models, linear regression and Ridge regression are more accurate than Lasso is, with a higher R² value of approximately 0.988 and lower RMSE and MAPE values of approximately 3.6 and 1.7, respectively. Elastic net underperformed, with the highest RMSE of 14.64, MAPE of 5.384, and lowest R² value of 0.954. In contrast, the nonlinear models show better overall performance, particularly DT (RMSE of 2.205, MAPE of 1.125) and SVR (RMSE of 2.1997, MAPE of 1.342555), with R² values higher than 99% for both, indicating strong performance and better reliability than those of the GP model and verifying the superiority of the nonlinear models over the linear models. The ensemble ML models, on the other hand, stand out as having the best performance, with the RF model attaining the lowest RMSE and MAPE values of 1.6296 and 0.42937, respectively, and the highest R² value of 0.998771 among all 13 considered models.

Download:

• PNG
  larger image
• TIFF
  original image

Table 5. CO₂ emission prediction results.

https://doi.org/10.1371/journal.pone.0342418.t005

For a better visual comparison, Fig 7 illustrates the distribution of prediction errors and goodness-of-fit metrics for the CO₂ emission estimation task across all evaluated ML models, such that the relative performance trends among different model families are clear. Similar to fuel consumption estimation, ensemble-based approaches generally achieve lower prediction errors and higher accuracy compared with linear and traditional nonlinear techniques. These findings further confirm the superiority of ensemble-based ML algorithms in terms of robustness and predictive consistency, particularly RF, which shows a consistent, top performing behaviour across both fuel consumption and CO₂ emission prediction tasks, through maintaining the lowest prediction errors and highest accuracy score.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 7. CO₂ Emission performance metrics of all ML models.

https://doi.org/10.1371/journal.pone.0342418.g007

Finally, Fig 8 shows the actual versus predicted plots for all 13 tested modelling approaches. Notably, these plots reveal solid distinctions between all model categories being studied in this analysis. While some linear models, such as Linear, Ridge and Lasso, show acceptable data alignments between the predictions and true emission values, Elastic Net clearly deviates, especially with mid- and high-range CO₂-emitting vehicles, thus reflecting weak model generalizability. For nonlinear models, SVR achieves acceptable performance by maintaining clear clustered data scatter. In contrast, GP and DT show high dispersion and scattering, especially with high-CO₂-emitting vehicles. In general, all the ensemble ML models outperform the linear and nonlinear models, where GB, XGBoost and CatBoost perform strongly but with slight variance, whereas HGB and RF stand out for their tight clustering, with RF showing excellent and consistent accuracy. These visual diagnostics reinforce the quantitative findings reported earlier, confirming that ensemble learning—especially RF—offers a reliable and robust framework not only for fuel consumption prediction, but also for CO₂ emission estimation by effectively maintaining stable predictive behaviour across diverse vehicle technologies.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 8. Actual versus predicted CO₂ emissions plots: (A) Linear regression, (B) Ridge, (C) Lasso, (D) Elastic Net, (E) SVM, (F) DT, (G) GP, (H) RF, (I) GB, (J) HGB, (K) XGBoost, (L) CatBoost, (M) LightGBM.

https://doi.org/10.1371/journal.pone.0342418.g008

5.3 Statistical significance testing

Besides visual parity plots and pointwise error calculations, it is important to assess whether the evaluated performance differences are statistically meaningful rather than data-split dependent. For this reason, statistical significance analysis was performed in this subsection, where the dataset was partitioned using an 80:20 split with proportional representation of vehicle categories, while model tuning and evaluation were further reinforced through 5-fold cross-validation to ensure robustness and eliminate any likelihood of overfitting to a particular data split. Accordingly, the mean value of each evaluation metric across the folds was computed according to (9).

(9)

where represents the performance metrics obtained from folds, where =5 for 5-fold validation.

To verify that the best-performing model offers a statistically significant improvement over a baseline approach, formal statistical significance testing was conducted. Since the RF model managed to outperform all the other tested models for consumption and emission estimation, its performance was benchmarked against the baseline Linear regression model.

As depicted in Fig 9(A), for consumption prediction from the 5-fold cross-validation results, the RF consistently outperforms the linear regression baseline across all evaluation metrics, scoring a lower average RMSE of 0.5548 compared with 0.8760 for the baseline linear metric, and subsequently, it is statistically significant, with a p value of 1.72 × 10 ⁻ ⁵. Similarly, a considerable decrease in the MAPE is exhibited in the RF model of 3.642%, whereas it is 6.597% for the linear model, corresponding to p = 6.17 × 10 ⁻ ⁹. With respect to the coefficient of determination, RF achieved a higher mean R² value of 0.9704 than 0.9262 for linear regression, achieving a statistically significant score of p = 2.93 × 10 ⁻ ⁵. Additionally, the standard deviation (STD) of each metric was computed to assess the stability of model performance across folds as (10):

Download:

• PNG
  larger image
• TIFF
  original image

Fig 9. Statistical significance testing (A) Fuel consumption, (B) CO₂ prediction.

https://doi.org/10.1371/journal.pone.0342418.g009

(10)

In addition, narrow confidence intervals were depicted across all five folds for consumption, with a 95% CI of 0.5112–0.5983 for RMSE and a 95% CI of 3.489%–3.794% for MAPE for RF, thus further confirming the stability and robustness of the RF model’s performance across all folds. These intervals correspond to low standard deviations of 0.035 for the RMSE and 0.123% for the MAPE, further confirming the stability and robustness of the RF model across all folds.

Similarly, Fig 9(B) for CO₂ prediction shows the results of the 5-fold cross-validation process, again, with the RF outperforming the baseline Linear regressor. Statistical significance testing revealed that the mean RMSE values decreased from 3.5648 to 5.1767 for the RF model, with p = 2.00 × 10 ⁻ ³. In addition, similar improvements are observed in the MAPE, which decreases from 1.718% for Linear regression to 0.429% for RF, with p = 4.39 × 10 ⁻ ⁶). Moreover, RF achieved a higher mean coefficient of determination, with an R2 value of 0.99877, compared with 0.98760 for Linear regression and p = 6.32 × 10 ⁻ ⁶. Similar to the consumption estimation, the emission estimation RF model’s 95% confidence intervals further validate its robustness across the 5 folds, with a mean RMSE CI of 3.3328--5.1767 and a MAPE CI of 0.848%−1.304%, corresponding to standard deviations of 0.742% and 0.184%, respectively, thereby validating robustness across all cross-validation folds.

6.1 Quantitative analysis between the proposed models and related work

As shown in Table 1, which is presented in Section 3, several studies have been conducted using the same data source, although they vary in terms of temporal coverage and modelling objectives. Nevertheless, the proposed approach excels in predicting both emissions and consumption rather than estimating one of them as well as depending on versatile data for both vehicle types rather than only one, which in turn increases its adaptability, scalability and generalizability. For further validation of the proposed models, a quantitative analysis is carried out, comparing the proposed results with those of their counterparts, which use the same dataset. Tables 6 and 7 provide this comparative analysis in terms of evaluation metrics (RMSE, MAPE and ) for forecasting CO2 emissions and energy consumption, respectively.

Download:

• PNG
  larger image
• TIFF
  original image

Table 6. CO2 emission prediction comparison with previous studies.

https://doi.org/10.1371/journal.pone.0342418.t006

Download:

• PNG
  larger image
• TIFF
  original image

Table 7. Fuel consumption prediction comparison with previous studies.

https://doi.org/10.1371/journal.pone.0342418.t007

Almost all related works focused exclusively on CO₂ emission prediction, neglecting fuel and energy consumption, except for the work in [35], which resembles the proposed approach as the sole study to provide a comprehensive analysis covering both fuel consumption and CO₂ emission prediction. However, the study applied the CNN-based DL approach with the smallest dataset among all the methods, featuring lower accuracies, i.e., R² values of 71.74% and 70%, than the proposed RF-based ensemble ML approach, with R² values of 71.74% and 70% for emissions and consumption, respectively, verifying the superiority of the proposed prediction models.

Other related works that considered only carbon emissions can be categorized as follows. GHG emission prediction in [49–51,53] was achieved through the adoption of several data-driven modelling strategies via Linear regression, nonlinear regression, and ensemble ML models. Compared with other regression approaches, the latter achieves the highest accuracies; GB, XGBoost, and CatBoost, and GB, RF and XGBoost are applied in [49–51,53], as demonstrated in Table 6. Comparatively, the RF-based ensemble ML applied in the proposed approach resulted in lower RMSE and MAPE values, verifying its excellent prediction performance. On the other hand, the research conducted in [55] targeted PHEVs only for predicting CO2 emissions via artificial neural network (ANN) models, making it the only study to include PHEVs in their analysis of the dataset. Nevertheless, the study did not include ICEV vehicles, which limits the complete comparative understanding of emission performance across different vehicle technologies. On the other hand, the highest accuracies in each [54,56,58] were achieved by applying DL-based approaches, i.e., the gate recurrent unit (GRU), bidirectional long short-term memory (Bi-LSTM) and light multilayer perceptron (MLP) approaches. Despite their high accuracy values with respect to computed errors, they achieve lower R² values than the proposed RF ensemble ML models do, which feature less complexity. In fact, the performance metrics for the ensemble ML model architecture were close to those of the DL models, thus reflecting that the added complexity of the DL models is not sufficiently justified by a significant improvement in prediction accuracy. Finally, both studies in [52,57] neglected categorical features, thus omitting contextual variables that could enhance the model’s predictive capabilities, resulting in the absence of class-related emission trends, potentially affecting the model’s generalizability.

Compared with related methods, the proposed predictive approach in this paper proposes generalized data-driven models for both carbon emissions and fuel consumption prediction on the basis of mixed data of both ICEVs and PHEVs. Compared with other ensemble ML models developed in the literature, the proposed RF-based approach achieves higher accuracies. Moreover, it achieves accuracies that are competitive with those of the DL models developed in related works, yet it has a less complicated realization and computational burden.

6.2 Proposed work contributions

A detailed comparative analysis of the model results verified the superiority of the ensemble ML-based RF models over other regression alternatives. Moreover, compared with other predictive models developed in related works that use datasets from the same source, the proposed RF-based models in this paper have the following capabilities:

• The emission and consumption predictions of the proposed models rely on two combined datasets provided by NRCan covering both ICEVs and PHEVs, unlike most related studies, which depend on their developed model for the analysis of only ICEV datasets. Hence, the proposed models are considered more versatile prediction tools for different types of light-duty vehicles.
• Rather than predicting only vehicle emissions in related works, the proposed work developed two models, one to predict emissions and the other to predict fuel consumption, which adds to its significance in vehicle design and policymaking.
• Compared with the ensemble ML models derived in related works, the approach proposed in this paper achieves higher accuracies, verifying their superiority. Moreover, compared with other alternatives that apply DL, the proposed ensemble ML models can capture both linear and nonlinear relationships with competing accuracies without relying on overly complex architectures.

Consequently, this study provides a robust foundation for developing evidence-based environmental and vehicle design policies in any region on the basis of similar relative datasets, paving the path for industries to adapt and innovate to anticipate economic and environmental impacts. This serves well the increasing commitment to the United Nations’ Sustainable Development Goals (SDGs), particularly those focused on climate action, clean affordable energy and sustainable urban development.

6.3 Real-time deployment and policy simulation considerations

The development of an accurate and reliable model for EV energy consumption and emissions is a fundamental requirement, not only for vehicle-level control but also for making EV policies and regulations, infrastructure planning, and policy simulation. In practical deployment scenarios, such models must meet additional challenges related to computational efficiency, hardware compatibility, and scalability, particularly when integrated into real-time vehicle systems or used for extensive fleet-level and regulatory analyses.

Despite recent advances in data-driven modelling, deploying predictive models in real-time vehicle systems and large-scale policy simulations is influenced by a wide range of interacting factors, including vehicle dynamics, road conditions, weather, auxiliary loads, traffic conditions, and driving styles [78]. In real-time applications, due to the strict delay constraint of data communication, instantaneous predictions are often required for decision-making, which results in stringent computational constraints and hardware feasibility [79]. While ensemble-based ML models offer improved accuracy and robustness, their real-time deployment in embedded vehicle systems must balance prediction fidelity against runtime efficiency.

With respect to fleet-level analyses and policy simulations, scalability and interpretability are critical considerations. Model transparency and adaptability to evolving vehicle technologies and operating conditions are essential, particularly since they rely on offline or batch processing. These considerations impose certain limitations on the current study, particularly when it transitions from offline model development to real-world operational environments and policy-driven use cases.

6.4 Future perspectives

With respect to future perspectives, this work could be extended to the following:

1. i. Fuel cell vehicles (FCVs)

Fuel cell vehicles (FCVs) represent a vital pathway toward large-scale zero-emission mobility, particularly because of their high efficiency and clean hydrogen-based operation. However, accurate parameter estimation and robust lifespan prediction of PEM fuel cell stacks are essential for ensuring reliable performance under real-world dynamic conditions [80]. These capabilities enable improved energy management, enhanced durability assessment, and more effective control strategies.

Therefore, extending future work to include FCV systems—supported by advanced estimation and degradation-prediction methods—will significantly contribute to the development of sustainable, hydrogen-powered transportation solutions.

1. ii. Hybrid ML–optimization architectures

The current machine learning–based forecasting framework can be extended by integrating advanced optimization techniques to further increase the prediction accuracy and support decision-making [81]. By accurately estimating system parameters, metaheuristic optimization approaches enable predictive ML models to generate reliable forecasts under varying driving cycles and operating conditions. Moreover, these algorithms can be integrated with HEV predictive frameworks to develop energy management strategies to minimize fuel usage, support the design of low-emission control policies, and contribute to sustainable transportation goals.

Hence, future work can explore hybrid ML–optimization architectures to automatically tune model parameters, identify optimal energy-management strategies, and generate more robust emission-minimization scenarios under varying driving conditions. Additionally, future studies may investigate the real-time implementation of these hybrid models, evaluate their performance on larger and more diverse datasets, and assess their applicability across different vehicle platforms to improve scalability and generalizability.