Machine Learning for Sustainable Shipping: Predicting Vessel CO₂ Emissions Using Random Forest Models

Carol Anne Hargreaves; Briana Wan Nee Toh

doi:10.5772/intechopen.1008820

Abstract

Predicting fuel consumption in the shipping industry is a critical task that supports optimized operations, driving both economic and environmental benefits as global demand for shipping continues to grow. However, accurately forecasting Carbon Dioxide (CO₂) emissions is challenging due to the complexity and volume of operational data. In this study, we developed and evaluated the Random Forest model to measure the prediction accuracy. The Random Forest model achieved a high predictive accuracy, with a Mean Absolute Percentage Error (MAPE) of 0.046, demonstrating its robustness in capturing the non-linear relationships in the data. Feature importance analysis within the Random Forest model highlighted vessel length, gross tonnage, width, and draft as key predictors of fuel consumption, offering valuable insights into CO₂ emissions-reduction strategies. These findings underscore the potential of machine learning to empower data-driven decisions in the maritime sector. Future work will explore further optimization of the Random Forest model, incorporate additional predictive features, and investigate real-time applications to enhance operational efficiency and sustainability. By making accurate CO₂ emission predictions, this research contributes to the industry’s efforts to achieve sustainable and environmentally responsible shipping practices.

Keywords

decarbonization
CO₂ emissions prediction
sustainability
maritime industry
machine learning
Random Forest model

Author Information

Show +

Carol Anne Hargreaves*
- National University of Singapore (NUS), Singapore, Singapore
Briana Wan Nee Toh
- National University of Singapore (NUS), Singapore, Singapore

*Address all correspondence to: carol.hargreaves@nus.edu.sg

1. Introduction

The growth of the shipping sector has intensified public concerns about decarbonization due to its significant environmental impact, particularly in terms of energy consumption and greenhouse gas (GHG) emissions. International shipping accounts for 2.4% of global anthropogenic GHG emissions, a percentage expected to rise in the coming years [1]. Discussions were held on the recent literature regarding technical and operational strategies to reduce fuel consumption for new and existing ships (hull and propeller design) [2]. Nadery et al. [3] numerically investigated a new pre-swirl stator (PSS) configuration on the ship propulsion’s performance. The KP505 propeller and the KRISO container ship (KCS) were used, and the PSS was installed before the propeller to adjust the wake inflow and to increase the performance of the propeller. The numerical results indicated that the PSS with these parameters worked well at low ship speed and increased the efficiency and decreased the delivered power.

The International Maritime Organization’s (IMO’s) 2050 targets will be achieved via radical technology shift together with the aid of social pressure, financial incentives, regulatory and legislative reforms at the local, regional, and international level. This has driven research interest in accurately predicting Carbon Dioxide (CO₂) emissions, as operational data hold immense potential to provide timely insights into a ship’s CO₂ efficiency [4].

Such data can help experts identify and address key factors contributing to CO₂ emissions, supporting efforts to reduce emissions through informed operational strategies [4]. A precise understanding of the relationship between fuel consumption and influencing factors is essential to optimize ship operations, improve energy efficiency, and develop effective navigation plans [5]. However, predicting fuel consumption efficiency is challenging due to the complexity of operational data, which are often unstructured, collinear, and non-stationary [4]. In this context, machine learning models offer promising solutions to process these data intricacies and generate reliable predictions.

2. Literature review

A common strategy for reducing fuel emissions in shipping is speed optimization, or “slow steaming” [6, 7, 8]. Using noon report data, Bal Beşikçi et al. [6] proposed an Artificial Neural Network (ANN) model for on-site optimization, which was integrated into a Decision Support System (DSS) to provide real-time fuel-efficient guidance for ship operations. Tay et al. [8] developed a Hidden Markov Model (HMM) to predict fuel consumption based on environmental conditions, enabling decision-makers to optimize ship speed for maximum fuel efficiency. Meanwhile, Bassam et al. [7] compared a range of regression techniques, including Multiple Linear Regression (MLR), tree ensemble models, and Support Vector Machines (SVMs). Their study underscores the inherent collinearity and non-linearity in ship operational data [4], highlighting the value of non-parametric models like tree-based regression and Gaussian Process Regression (GPR) for capturing complex data patterns effectively.

Another prominent research area involves estimating a ship’s propulsion power to enhance operational efficiency [9]. Dos Santos Ferreira et al. [9] explored six algorithms to model the relationship between propulsion power and fuel consumption, leveraging ship operational variables, such as Speed Over Ground, Course Over Ground, Speed through Water, Shaft Revolutions Per Minute (RPM), mean draft, and trim. Their study found that ensemble models, specifically gradient-boosted trees and random forests, delivered superior precision in fuel consumption prediction compared to least squares methods and Multilayer Perceptron Neural Networks, supporting their applicability for future fuel efficiency studies [9].

In recent years, shipborne emissions have become a growing environmental threat. The International Maritime Organization has implemented various rules and regulations to resolve this concern. The Ship Energy Efficiency Management Plan (SEEMP), Energy Efficiency Design Index (EEDI), and Energy Efficiency Operational Indicator (EEOI) are examples of guidelines that increase energy efficiency and reduce shipborne emissions. The main engine shaft power (MESP) and fuel consumption (FC) are the critical components used in ship energy efficiency calculations.

A study by Tayfun Uyanık et al. [10] aimed to predict the MESP and FC of a container ship with the help of data-driven methodologies utilizing actual voyage data to assist in the calculation process of the ship’s energy efficiency indexes appropriately. The algorithms’ prediction success was measured using the Root Mean Square Error (RMSE), Mean absolute Error (MAE), and R-Squared (R²) error metrics. When the simulation results were analyzed, the Deep Neural Network and Bayes algorithms predicted MESP best with 0.000001 and 0.000002 RMSE, 0.000987 and 0.000991 MAE, and 0.999999 R², respectively, while the Multiple-Linear Regression and Kernel Ridge algorithms estimated FC best with 0.000208 and 0.000216 RMSE, 0.001375 and 0.001471 MAE, and 0.999999 R², respectively.

Collectively, these studies reveal a growing preference for machine learning and advanced regression methods in the prediction of fuel consumption and emissions, with a particular emphasis on handling the non-linear and collinear nature of operational data. This body of work lays the groundwork for developing robust models that improve fuel efficiency, reduce emissions, and support sustainable practices in the shipping industry.

3. Objective of study

Despite extensive research into the factors influencing carbon emissions in the shipping industry, the impact of ship draft and length on emissions remains underexplored. Most studies have centered on variables, such as ship speed, fuel type, and environmental conditions, with limited attention to how structural characteristics might affect emissions efficiency.

Three responses that reduce energy consumption and CO₂ emissions in maritime transport are slower speeds, larger vessels, and slender hull designs (Dr Elizabeth Lindstad, et al. [11]). She also demonstrated that increasing the length of a 63,000 Dead Weight Tonnage (DWT) supramax vessel from 200 m to 220 m, thereby creating a “slender supramax” with a reduced block coefficient (Cb) from 0.875 to 0.806, can yield significant power consumption savings, 15–20% on typical northern European sea routes, while maintaining the same operational speed (“Half Measures: Could “Slender” Ships Be the Answer to IMO’s GHG Targets?,” n.d.). This evidence underscores the potential of ship length as a critical design factor in reducing emissions, providing a compelling case for investigating how vessel design elements, such as slenderness and draft, might contribute to achieving emissions reductions.

This study aims to address this gap by focusing on the role of draft and length in shipping emissions, thereby broadening the scope of emission-reducing strategies beyond traditional operational factors. By understanding the influence of these design attributes, the research has the potential to inform more sustainable ship designs and contribute to industry-wide decarbonization efforts.

4. Methodology

4.1 Data collection and description

This study draws on a comprehensive dataset comprising two main components: Automatic Identification System (AIS) data and vessel emissions information. Focusing on CO₂ emissions data from 2021, we sourced emissions data from the European Union (EU) Monitoring, Reporting, and Verification (EU MRV) platform [12], which offers standardized insights into emissions generated by vessels operating in EU waters.

The dataset, referred to as vessel info, includes key characteristics such as gross tonnage and age of ships, with each vessel identifiable through its unique International Maritime Organization (IMO) number. For the AIS data, we utilized information from MarineCadastre.gov, extracting vessel traffic data from the first and last days of each month in 2021. After initial processing, we integrated these data with the EU MRV emissions and vessel information. This integration, based on matching IMO numbers, ensured a clean and unified dataset.

Our final dataset provides operational, structural details and emissions data of each vessel, including Length, Draft, Gross Tonnage, CO₂ emissions per distance traveled, Distance traveled, time spent at sea, fuel consumption, and CO₂ emissions by voyage. This dataset allows us to explore not only the operational metrics tied to emissions but also how vessel-specific attributes may influence CO₂ output, enabling a more robust analysis of emission trends and potential efficiencies.

4.2 Preprocessing the data

In this study, we applied one-hot encoding to process the categorical variables such as ship_type and AgeGroup. One-hot encoding is a widely used technique for handling categorical variables in machine learning. It transforms each unique category value into a new binary column, where the presence of a category is indicated by a value of “1” and its absence by a “0” [13]. This approach prevents any unintended intercorrelation between variables, ensuring each category is represented independently and avoids misinterpretation.

For example, if a ship is classified as a Bulk Carrier in the ship_type column, a new column specifically for Bulk Carrier will show a value of “1,” while all other ship types will display a “0” in this column. This one-hot encoding method was similarly applied to the AgeGroup variable, allowing each age category to be represented as a distinct binary column.

4.3 Model used: Random Forest

The Random Forest method, introduced by Breiman [14], improves model accuracy by creating an ensemble of decision trees, each trained on randomly selected subsets of data. This randomness—either in the choice of features or in the combination of features at each node—reduces the likelihood of overfitting while allowing the model to capture complex patterns. Even as the number of trees grows, the Random Forest model’s generalization error stabilizes, converging due to the Strong Law of Large Numbers, which limits overfitting [14].

In Random Forests, predictions are aggregated from numerous decision trees, which enhance robustness and reduce variance compared to a single decision tree model [15]. This approach is also resilient to noise and outliers, making it well suited for real-world datasets with irregularities [14].

A key feature of Random Forests is bagging (bootstrap aggregating), where each tree is trained on a bootstrapped sample of the training data. This process involves drawing samples with replacement, creating multiple training sets that increase model robustness. Bagging generates a collection of estimators on these diverse samples and averages their predictions to form a final output, improving stability and reducing variance [16].

The procedure of bootstrap aggregating is given by the following steps [14]:

Generate Bootstrap Samples: For each tree 𝑏 in the ensemble, generate a bootstrap sample {(y1b,x1b),,…,,(xNb,yNb)} by randomly drawing with replacement from the training data, where b=1,…,B.
Train Each Tree: For each bootstrap sample, fit a decision tree fb̂(x) by minimizing the L2 loss function:

minfb̂(x)∑i=1N(yib−fb̂(xib))2E1

Aggregate Predictions: Once all trees are trained, aggregate their predictions to obtain the final output. The bagging estimate is calculated as:

f̂(x)bagging=1B∑b=1Bfb̂(x)E2

By combining the outputs of multiple decision trees, Random Forests achieve a balance of accuracy and robustness, avoiding reliance on any single decision tree’s prediction while reducing generalization error. This ensemble approach makes Random Forests a powerful choice for predictive tasks in data-driven fields.

4.4 Model optimization method: Grid search cross-validation

To optimize the model, we employed Grid Search Cross-Validation (CV) to systematically identify the best hyperparameters. In this approach, a set of hyperparameters is passed on to the GridSearchCV function along with predetermined values for each parameter. We define a dictionary, where each hyperparameter is assigned a list of possible values to explore [17].

The Grid Search process uses cross-validation to evaluate the model’s performance across all possible combinations of these hyperparameters, ensuring a thorough search of the parameter space. For each combination, metrics such as accuracy and loss are computed, allowing us to assess the model’s effectiveness for different parameter settings. Finally, the hyperparameter combination that yields the best performance, as indicated by the highest accuracy or lowest loss, is selected [17].

4.5 Model performance indicators

To assess the performance of our models in predicting CO₂ emissions, we employed three key performance indicators: Mean Absolute Percentage Error (MAPE), Root Mean Square Error (RMSE), and Mean Absolute Percent Accuracy (Accuracy). These indicators provide insight into the accuracy and reliability of the models by comparing the actual and predicted CO₂ emissions per distance, denoted as yi and yî, respectively [18].

Mean absolute percentage error (MAPE): MAPE measures the average percentage error between the actual and predicted values, offering a normalized metric that highlights model accuracy over multiple observations. It is calculated by dividing the absolute error for each instance by the observed value, as shown in Eq. (3) [19]:

MAPE=100n×∑i=1n|yî−yiyi|E3

Root mean square error (RMSE): RMSE is widely used to evaluate model performance due to its interpretability; it uses the same units as the predicted variable, which allows it to directly reflect the model’s prediction accuracy. RMSE is derived from the mean square error (MSE) by taking the square root of the average squared differences between the actual and predicted values [7]:

MSE=1n∑i=1n(yi−yî)2E4

RMSE=MSEE5

Mean absolute percent accuracy (accuracy): Accuracy is computed as the complement of MAPE, representing the percentage of accuracy in predictions relative to observed values [20]:

Accuracy=100%−MAPEE6

Together, these performance indicators provide a comprehensive evaluation of the models, allowing us to gauge both the absolute error magnitude (via RMSE) and percentage accuracy (via MAPE and Accuracy) of the CO₂ emissions predictions.

5. Results and analysis

5.1 Initial Random Forest model

In the initial Random Forest model, we utilized GridSearchCV to optimize the “n_estimators” hyperparameter on the training and validation sets. This tuning process helped to identify the optimal number of trees for improved performance. Based on feature importance analysis, we selected the top six most influential features and created a refined dataset, X_new, containing only these features to enhance model efficiency and reduce computational complexity.

With the optimal hyperparameters from the grid search, we fitted the RandomForestRegressor model using X_new. This approach allowed us to leverage the most relevant features, potentially improving model accuracy and stability by focusing on factors most closely associated with the target variable.

Table 1 presents the accuracy results of the initial Random Forest model in predicting CO₂ emissions. The model demonstrated high predictive accuracy, achieving an accuracy rate of 95.4%. Additionally, the Mean Absolute Percentage Error (MAPE) was 0.046, indicating that the model’s predictions deviated minimally from the actual values. These results suggest that the Random Forest model, with tuned hyperparameters and feature selection, is highly effective in accurately estimating vessel CO₂ emissions.

Evaluation metric for initial model	RMSE	MAPE	Accuracy
Training set	28.271	0.046	95.4%

Table 1.

Initial random forest model accuracy results.

5.2 Feature importance for Random Forest model

Permutation Importance (PIMP) is an effective method for feature selection, particularly valuable in enhancing the interpretability of models by quantifying the significance of each variable [21]. This approach has proven especially useful in evaluating the impact of features within a Random Forest model. By applying permutation importance, we calculate p-values for each feature, derived through a permutation test that normalizes biased measures.

To preserve the relationships between features, the permutation test is applied to the outcome variable rather than directly to the features themselves. This allows us to correct the inherent bias in feature significance estimates produced by Random Forests, providing a more accurate measure of each variable’s contribution to the model’s predictions [21]. The resulting feature importance scores, as shown below, offer valuable insights into the variables that are most influential in predicting CO₂ emissions.

Table 2 presents the feature importance scores derived from the Random Forest model using permutation importance. These scores reflect the relative impact of each feature on predicting CO₂ emissions, with higher values indicating a stronger influence on the model’s predictions.

Feature	Importance
Length	1.121203
grossTonnage	0.086251
Width	0.036428
Draft	0.018683
ship_type_Container ship	0.013443
AgeGroup_6–10	0.005324
AgeGroup_11–15	0.002319
AgeGroup_16–20	0.001167
ship_type_Bulk carrier	0.001013
AgeGroup_0–5	0.000391
ship_type_General cargo ship	0.000074
AgeGroup_21–25	0.000057
ship_type_Oil tanker	0.000051
AgeGroup_6–10	0.000040

Table 2.

The feature importance scores derived from the random forest model.

The analysis of feature importance scores in our Random Forest model identified several key variables that significantly influence CO₂ emissions. The most impactful feature was vessel length (importance score: 1.121203), suggesting that vessel length is a critical factor in emissions prediction. This is followed by gross tonnage (importance score: 0.086251), which indicates that vessel weight also plays a substantial role in determining emissions levels. Vessel width (importance score: 0.036428) emerged as another important feature, highlighting the relevance of overall vessel dimensions in emissions-related predictions. Additionally, draft (importance score: 0.018683), though less influential than the top three features, still showed a notable contribution to CO₂ emissions, underscoring the importance of a vessel’s depth in the water as a factor.

In contrast, certain features exhibited minimal importance scores, suggesting they may not add substantial predictive value to the model. For example, specific ship types (Bulk Carrier, General Cargo Ship, Oil Tanker) demonstrated very low scores, ranging from 0.000074 to 0.001013. This indicates that, apart from Container Ships, vessel type does not strongly impact CO₂ emissions predictions. Similarly, age group classifications (e.g., AgeGroup_11–15, AgeGroup_16–20) yielded low importance scores (0.000057 or below), implying that a vessel’s age may have negligible influence on emissions in this dataset.

Based on these findings, we recommend focusing on the most influential features, such as Length, Gross Tonnage, Width, and Draft, for CO₂ emissions prediction. Conversely, features, such as specific ship types (excluding Container Ships) and age groups, may be deprioritized or ignored in future modeling efforts, given their minimal impact. This approach can streamline the model, improving computational efficiency while maintaining predictive accuracy. These insights provide a data-driven foundation for prioritizing design and operational factors that contribute most to emissions, offering valuable guidance or emissions-reduction strategies within the shipping industry.

6. Discussion and conclusion

The findings of this study underscore the importance of vessel dimensions and design characteristics, such as length, gross tonnage, and width, as key determinants of CO₂ emissions in the maritime sector. These insights suggest that optimizing vessel design could lead to significant reductions in emissions. For instance, longer, more streamlined vessels have been shown to experience reduced hydrodynamic drag, which in turn lowers fuel consumption. This finding aligns with prior research that advocates for “slender” vessel designs as an emissions-reduction strategy. Therefore, future vessel designs could incorporate these structural adjustments to enhance fuel efficiency and reduce CO₂ output.

For existing vessels, retrofitting options, such as hull shape modifications and the installation of bulbous bows optimized for different speeds, could provide immediate emissions benefits. Operational optimization also emerged as a critical factor, particularly through speed and draft management. Speed optimization, or “slow steaming,” has consistently demonstrated its potential to reduce fuel consumption, and the results of this study highlight its continued relevance. Maritime companies should consider implementing speed optimization protocols across their fleets, particularly on longer voyages, to maximize fuel efficiency. Similarly, draft management, tailored to cargo load and route conditions, can help reduce resistance, leading to lower emissions.

The high predictive accuracy of the Random Forest model used in this study indicates that advanced machine learning techniques hold promise for real-time decision-making in emissions reduction. Integrating predictive analytics into onboard systems could allow vessel operators to make data-driven decisions that optimize speed, route, and other operational factors, ultimately minimizing CO₂ emissions. This approach could be particularly valuable in voyage planning, where models can suggest optimal speeds and routes that balance emissions reduction with timely delivery. By leveraging predictive insights, shipping companies could develop dynamic, adaptive strategies to maintain efficiency in various operational scenarios.

Together, these insights and proposed actions provide a multifaceted approach to reducing CO₂ emissions in the maritime sector. By focusing on both immediate operational adjustments and longer-term investments in sustainable practices, the maritime industry can make meaningful progress toward environmental sustainability while enhancing operational efficiency and remaining competitive in an increasingly eco-conscious global market.

7. Future work

We propose the implementation of digital platforms for tracking and monitoring vessel emissions data to allow companies to set measurable targets and assess progress toward achieving their CO₂ emission goals. Such transparency will not only enhance a company’s reputation but can also align with the growing market demand for sustainability in maritime operations. In the long term, companies should consider investing in green technologies, including wind-assisted propulsion systems, hybrid engines, and solar power integration, which offer sustainable alternatives to traditional fuel-reliant propulsion. We strongly suggest collaborations with industry stakeholders to shape policies and standards for sustainable practices to accelerate industry-wide adoption of effective CO₂ emissions-reduction strategies.

Conflict of interest

The authors declare no conflict of interest.

References

1. Spinelli F, Mancini S, Vitiello L, Bilandi RN, De Carlini M. Shipping decarbonization: An overview of the different stern hydrodynamic energy saving devices. Journal of Marine Science and Engineering. 2022;10(5):574. DOI: 10.3390/jmse10050574
2. Mallouppas G, Yfantis E. Decarbonization in shipping industry: A review of research, technology development, and innovation proposals. Journal of Marine Science and Engineering. 2021;9:415, 10.3390/jmse9040415
3. Nadery A, Ghassemi H, Chybowski L. The effect of the PSS configuration on the hydrodynamic performance of the KP505 propeller behind the KCS. Ocean Engineering. 2021;234:109310, 10.1016/j.oceaneng.2021.109310
4. Reis MS, Rendall R, Palumbo B, Lepore A, Capezza C. Predicting ships’ CO₂ emissions using feature-oriented methods. Applied Stochastic Models in Business and Industry. 2019;36(1):110-123. DOI: 10.1002/asmb.2477
5. Jeon M, Noh Y, Shin Y, Lim OK, Lee I, Cho D. Prediction of ship fuel consumption by using an artificial neural network. Journal of Mechanical Science and Technology. 2018;32(12):5785-5796. DOI: 10.1007/s12206-018-1126-4
6. Bal Beşikçi E, Arslan O, Turan O, Ölçer A. An artificial neural network-based decision support system for energy efficient ship operations. Computers & Operations Research. 2016;66:393, 10.1016/j.cor.2015.04.004-401
7. Bassam AM, Phillips AB, Turnock SR, Wilson PA. Ship speed prediction based on machine learning for efficient shipping operation. Ocean Engineering. 2022;245:1-11. DOI: 10.1016/j.oceaneng.2021.110449
8. Tay ZY, Hadi J, Chow F, Loh DJ, Konovessis D. Big data analytics and machine learning of harbour craft vessels to achieve fuel efficiency: A review. Journal of Marine Science and Engineering. 2021;9(12):1351. DOI: 10.3390/jmse9121351
9. Dos Santos Ferreira R, Padilha De Lima JV, Caprace JD. Comparative analysis of machine learning prediction models of container ships propulsion power. Ocean Engineering. 2022;255:111439, 10.1016/j.oceaneng.2022.111439
10. Uyanık T, Yalman Y, Kalenderli Ö, Arslanoğlu Y, Terriche Y, Chun-Lien S, et al. Data-driven approach for estimating power and fuel consumption of ship: A case of container vessel. Mathematics. 2022;10:1-21. DOI: 10.3390/math10224167
11. Lindstad DE, Eskeland GS. Low carbon maritime transport: How speed, size and slenderness amount to substantial capital energy substitution. Transportation Research Part D Transport and Environment. 2015;41(3):244-256. DOI: 10.1016/j.trd.2015.10.006
12. THETIS-MRV. Available from: https://mrv.emsa.europa.eu/#public/eumrv
13. Al-Shehari T, Alsowail R. An insider data leakage detection using one-hot encoding, synthetic minority oversampling and machine learning techniques. Entropy. 2021;23(10):1258. DOI: 10.3390/e23101258
14. Breiman L. Random forests. Machine Learning. 2001;45(1):5-32. DOI: 10.1023/a:1010933404324
15. Couronné R, Probst P, Boulesteix A. Random forest versus logistic regression: A large-scale benchmark experiment. BMC Bioinformatics. 2018;19(1):1-14. DOI: 10.1186/s12859-018-2264-5
16. Lee T, Ullah A, Wang R. Bootstrap Aggregating and Random Forest. Springer EBooks; 2020. pp. 389-429. DOI: 10.1007/978-3-030-31150-6_13
17. Great Learning Team. An Introduction to GridSearchCV | What is Grid Search | Great Learning. Great Learning Blog: Free Resources What Matters to Shape Your Career!. 2022. Available from: https://www.mygreatlearning.com/blog/gridsearchcv/#:~:text=GridSearchCV%20is%20a%20technique%20for,the%20parameters%20must%20be%20entered
18. Panapakidis IP, Sourtzi V, Dagoumas A. Forecasting the fuel consumption of passenger ships with a combination of shallow and deep learning. Electronics. 2020;9(5):776. DOI: 10.3390/electronics9050776
19. Khair U, Fahmi H, Hakim SA, Rahim R. Forecasting error calculation with mean absolute deviation and mean absolute percentage error. Journal of Physics: Conference Series. 2017;930:012002. DOI: 10.1088/1742-6596/930/1/012002
20. Singh S. Two Sides of the MAPE Coin. Supply Chain Link Blog - Arkieva. 2019. Available from: https://blog.arkieva.com/two-sides-of-the-mape-coin/
21. Altmann A, Tolosi L, Sander O, Lengauer T. Permutation importance: A corrected feature importance measure. Computer Applications in the Biosciences. 2010;26(10):1340-1347. DOI: 10.1093/bioinformatics/btq134

[1] 1. Spinelli F, Mancini S, Vitiello L, Bilandi RN, De Carlini M. Shipping decarbonization: An overview of the different stern hydrodynamic energy saving devices. Journal of Marine Science and Engineering. 2022;10(5):574. DOI: 10.3390/jmse10050574

[2] 2. Mallouppas G, Yfantis E. Decarbonization in shipping industry: A review of research, technology development, and innovation proposals. Journal of Marine Science and Engineering. 2021;9:415, 10.3390/jmse9040415

[3] 3. Nadery A, Ghassemi H, Chybowski L. The effect of the PSS configuration on the hydrodynamic performance of the KP505 propeller behind the KCS. Ocean Engineering. 2021;234:109310, 10.1016/j.oceaneng.2021.109310

[4] 4. Reis MS, Rendall R, Palumbo B, Lepore A, Capezza C. Predicting ships’ CO₂ emissions using feature-oriented methods. Applied Stochastic Models in Business and Industry. 2019;36(1):110-123. DOI: 10.1002/asmb.2477

[5] 5. Jeon M, Noh Y, Shin Y, Lim OK, Lee I, Cho D. Prediction of ship fuel consumption by using an artificial neural network. Journal of Mechanical Science and Technology. 2018;32(12):5785-5796. DOI: 10.1007/s12206-018-1126-4

[6] 6. Bal Beşikçi E, Arslan O, Turan O, Ölçer A. An artificial neural network-based decision support system for energy efficient ship operations. Computers & Operations Research. 2016;66:393, 10.1016/j.cor.2015.04.004-401

[7] 7. Bassam AM, Phillips AB, Turnock SR, Wilson PA. Ship speed prediction based on machine learning for efficient shipping operation. Ocean Engineering. 2022;245:1-11. DOI: 10.1016/j.oceaneng.2021.110449

[8] 8. Tay ZY, Hadi J, Chow F, Loh DJ, Konovessis D. Big data analytics and machine learning of harbour craft vessels to achieve fuel efficiency: A review. Journal of Marine Science and Engineering. 2021;9(12):1351. DOI: 10.3390/jmse9121351

[9] 9. Dos Santos Ferreira R, Padilha De Lima JV, Caprace JD. Comparative analysis of machine learning prediction models of container ships propulsion power. Ocean Engineering. 2022;255:111439, 10.1016/j.oceaneng.2022.111439

[10] 10. Uyanık T, Yalman Y, Kalenderli Ö, Arslanoğlu Y, Terriche Y, Chun-Lien S, et al. Data-driven approach for estimating power and fuel consumption of ship: A case of container vessel. Mathematics. 2022;10:1-21. DOI: 10.3390/math10224167

[11] 11. Lindstad DE, Eskeland GS. Low carbon maritime transport: How speed, size and slenderness amount to substantial capital energy substitution. Transportation Research Part D Transport and Environment. 2015;41(3):244-256. DOI: 10.1016/j.trd.2015.10.006

[12] 12. THETIS-MRV. Available from: https://mrv.emsa.europa.eu/#public/eumrv

[13] 13. Al-Shehari T, Alsowail R. An insider data leakage detection using one-hot encoding, synthetic minority oversampling and machine learning techniques. Entropy. 2021;23(10):1258. DOI: 10.3390/e23101258

[14] 14. Breiman L. Random forests. Machine Learning. 2001;45(1):5-32. DOI: 10.1023/a:1010933404324

[15] 15. Couronné R, Probst P, Boulesteix A. Random forest versus logistic regression: A large-scale benchmark experiment. BMC Bioinformatics. 2018;19(1):1-14. DOI: 10.1186/s12859-018-2264-5

[16] 16. Lee T, Ullah A, Wang R. Bootstrap Aggregating and Random Forest. Springer EBooks; 2020. pp. 389-429. DOI: 10.1007/978-3-030-31150-6_13

[17] 17. Great Learning Team. An Introduction to GridSearchCV | What is Grid Search | Great Learning. Great Learning Blog: Free Resources What Matters to Shape Your Career!. 2022. Available from: https://www.mygreatlearning.com/blog/gridsearchcv/#:~:text=GridSearchCV%20is%20a%20technique%20for,the%20parameters%20must%20be%20entered

[18] 18. Panapakidis IP, Sourtzi V, Dagoumas A. Forecasting the fuel consumption of passenger ships with a combination of shallow and deep learning. Electronics. 2020;9(5):776. DOI: 10.3390/electronics9050776

[19] 19. Khair U, Fahmi H, Hakim SA, Rahim R. Forecasting error calculation with mean absolute deviation and mean absolute percentage error. Journal of Physics: Conference Series. 2017;930:012002. DOI: 10.1088/1742-6596/930/1/012002

[20] 20. Singh S. Two Sides of the MAPE Coin. Supply Chain Link Blog - Arkieva. 2019. Available from: https://blog.arkieva.com/two-sides-of-the-mape-coin/

[21] 21. Altmann A, Tolosi L, Sander O, Lengauer T. Permutation importance: A corrected feature importance measure. Computer Applications in the Biosciences. 2010;26(10):1340-1347. DOI: 10.1093/bioinformatics/btq134