Next Article in Journal
TACO: Adversarial Camouflage Optimization on Trucks to Fool Object Detectors
Previous Article in Journal
Generation Z’s Travel Behavior and Climate Change: A Comparative Study for Greece and the UK
 
 
Search for Articles:
Title / Keyword
Author / Affiliation / Email
Journal
Article Type
 
 
Advanced Search
 
Section
Special Issue
Volume
Issue
Number
Page
 
Journals
BDCC
Volume 9
Issue 3
10.3390/bdcc9030071
BDCC-logo
Submit to this Journal Review for this Journal Propose a Special Issue
► Article Menu

Article Menu

share Share announcement Help format_quote Cite question_answer Discuss in SciProfiles
first_page
settings
Order Article Reprints
Font Type:
Arial Georgia Verdana
Font Size:
Aa Aa Aa
Line Spacing:
Column Width:
Background:
Article

Data-Driven Forecasting of CO2 Emissions in Thailand’s Transportation Sector Using Nonlinear Autoregressive Neural Networks

by
Thananya Janhuaton
1,
Supanida Nanthawong
1,
Panuwat Wisutwattanasak
2,
Chinnakrit Banyong
3,
Chamroeun Se
2,
Thanapong Champahom
4,
Vatanavongs Ratanavaraha
1 and
Sajjakaj Jomnonkwao
1,*
1
School of Transportation Engineering, Institute of Engineering, Suranaree University of Technology, Nakhon Ratchasima 30000, Thailand
2
Institute of Research and Development, Suranaree University of Technology, Nakhon Ratchasima 30000, Thailand
3
Program of Industrial and Logistics Management Engineering, Institute of Engineering, Suranaree University of Technology, Nakhon Ratchasima 30000, Thailand
4
Department of Management, Faculty of Business Administration, Rajamangala University of Technology Isan, Nakhon Ratchasima 30000, Thailand
*
Author to whom correspondence should be addressed.
Big Data Cogn. Comput. 2025, 9(3), 71; https://doi.org/10.3390/bdcc9030071
Submission received: 8 February 2025 / Revised: 7 March 2025 / Accepted: 14 March 2025 / Published: 17 March 2025
Browse Figures
Versions Notes

Abstract

Accurately forecasting CO2 emissions in the transportation sector is essential for developing effective mitigation strategies. This study uses an annually spanning dataset from 1993 to 2022 to evaluate the predictive performance of three methods: NAR, NARX, and GA-T2FIS. Among these, NARX-VK, which incorporates vehicle kilometers (VK) and economic variables, demonstrated the highest predictive accuracy, achieving a MAPE of 2.2%, MAE of 1621.449 × 103 tons, and RMSE of 1853.799 × 103 tons. This performance surpasses that of NARX-RG, which relies on registered vehicle data and achieved a MAPE of 3.7%. While GA-T2FIS exhibited slightly lower accuracy than NARX-VK, it demonstrated robust performance in handling uncertainties and nonlinear relationships, achieving a MAPE of 2.6%. Sensitivity analysis indicated that changes in VK significantly influence CO2 emissions. The Green Transition Scenario, assuming a 10% reduction in VK, led to a 4.4% decrease in peak CO2 emissions and a 4.1% reduction in total emissions. Conversely, the High Growth Scenario, modeling a 10% increase in VK, resulted in a 7.2% rise in peak emissions and a 4.1% increase in total emissions.

1. Introduction

Climate change poses a formidable challenge to global sustainability, with the transportation sector being a major source of greenhouse gas emissions, accounting for approximately 23% of total energy-related emissions as of 2019 [1]. Transport continues to rely on oil products for nearly 90.539% of its final energy [2]. According to the International Energy Agency (IEA), CO2 emissions from the transport sector increased by over 250 Mt CO2 in 2022, reaching nearly 8 Gt CO2, which is a 3% rise compared to 2021 [3]. The rise in emissions is attributed to several factors, including increased vehicle ownership, urbanization, and economic growth [4,5]. The transportation sector encompasses various modes, including road, rail, air, and maritime transport. Each mode contributes differently to CO2 emissions, with road transport being the largest contributor, responsible for approximately 70% of total transport emissions [1,6].
In Thailand, the transportation sector is a major contributor to the country’s carbon dioxide (CO2) emissions, mainly due to its reliance on fossil fuels and the increasing demand for mobility. According to the Energy Policy and Planning Office [7], in 2022, the transportation sector accounted for approximately 32.1% of Thailand’s total carbon emissions, second only to the energy production sector, which accounted for 35.5%, as illustrated in Figure 1. The transportation sector’s energy consumption was overwhelmingly dominated by oil, which made up around 97% of its total energy use. Consequently, transport was responsible for an estimated 74% of the CO2 emissions resulting from oil combustion, as depicted in Figure 2. In comparison, the Power Generation, Industry, and Others sectors contributed roughly 0.5%, 12.1%, and 13.2%, respectively. This substantial dependence on fossil fuels highlights the urgent need for sustainable alternatives and effective emission reduction strategies within Thailand’s transportation sector, where accurate CO2 emissions forecasting plays a critical role in policy development, energy optimization, and environmental impact reduction. Several studies have developed models to forecast CO2 emissions from Thailand’s transportation sector [8,9,10].
In addition to its environmental impact, the transportation sector in Thailand faces significant economic challenges due to climate-related hazards. The country experiences an average annual loss of approximately USD 125.78 million (0.01% of GDP) due to hazards impacting its transport infrastructure, primarily roads (78%) and rail (21%) [11]. To address these challenges, Thailand has set climate goals, aiming for carbon neutrality by 2050 and net-zero greenhouse gas emissions by 2065 [12].
Traditional forecasting approaches, including statistical regression methods and time-series models such as AutoRegressive Integrated Moving Average (ARIMA), have been widely applied in emissions modeling. However, traditional forecasting methods often struggle to capture the complex, nonlinear relationships inherent in real-world data.
In contrast, time-series neural networks, specifically those designed for sequential data processing, have demonstrated superior capabilities in capturing temporal dependencies and nonlinear relationships. Among these, Nonlinear AutoRegressive (NAR) and Nonlinear AutoRegressive with Exogenous Inputs (NARX) models effectively integrate time-series analysis with neural network architectures, enabling them to learn from historical patterns while dynamically incorporating external influences. These models have been successfully applied in emissions forecasting, exhibiting enhanced predictive accuracy and adaptability.
However, one key limitation of NAR and NARX models is their inability to explicitly handle uncertainty and imprecise data, which is prevalent in real-world emissions forecasting due to factors such as economic fluctuations, traffic variability, and changing policies. To address this challenge, fuzzy logic-based models have been explored as effective tools for handling uncertainty and imprecise input data in forecasting [13,14].
Therefore, effective energy planning to reduce CO2 emissions in Thailand’s transportation sector is crucial for promoting environmental sustainability. This research conducts a comparative evaluation of Nonlinear AutoRegressive (NAR), Nonlinear AutoRegressive with Exogenous Inputs (NARX), and a Genetic Algorithm-Tuned Type-2 Fuzzy Inference System (GA-T2FIS) to systematically analyze the trade-offs between predictive accuracy and uncertainty handling in emissions forecasting. By integrating time-series neural networks with an uncertainty-aware approach, this study aims to improve forecasting techniques and provide a more flexible framework for emissions prediction under real-world variability. Additionally, this research incorporates economic and vehicle-related factors into predictive models, offering insights that may support transportation energy planning and emissions reduction strategies in Thailand.

2. Literature Review

Forecasting CO2 emissions in the transport sector is critical for mitigating the environmental impacts of transportation, as it enables more informed decision-making and policy development. In the past decade, there has been a growing emphasis on research focused on developing accurate models to predict emissions. This has led to the use of both traditional statistical methods, which have been reliable for many years, and newer machine learning (ML) techniques, which have shown promising results in handling complex and nonlinear data.
In the application of traditional statistical methods for CO2 emission forecasting, Fatima et al. [15] applies Simple Exponential Smoothing (SES) and ARIMA (AutoRegressive Integrated Moving Average) models across multiple sectors in various Asian countries, demonstrating that SES is more effective for nations such as Pakistan and Sri Lanka, while ARIMA performs better for countries like Japan, China, India, Iran, and Singapore. In their analysis of Thailand’s transportation emissions, Ratanavaraha and Jomnonkwao [9] examined multiple forecasting methodologies. Their comprehensive evaluation compared several quantitative approaches, encompassing regression analysis, structural equation modeling, time series forecasting, and trend estimation techniques. Their findings revealed that time series forecasting demonstrated superior performance in capturing emission patterns compared to alternative statistical methods. Additionally, Log-Linear Regression and Path Analysis explore the relationships between independent variables such as GDP, population size, and vehicle registrations and their impact on transportation-related emissions. Traditional statistical models, such as ARIMA, have been widely applied in CO2 emission forecasting. ARIMA works well when emission data follows a linear trend and when the past values of emissions strongly influence future predictions. For example, Rahman and Hasan [16] demonstrated the effectiveness of ARIMA in predicting carbon emissions in Bangladesh, highlighting how the model utilizes historical data to effectively capture the influence of past emissions on future trends. Similarly, ARIMA has been found suitable for forecasting CO2 emissions in various regions, including South Africa, India, China, and Thailand, where it has provided accurate short- to medium-term predictions [17,18,19,20]. The synthesis of these research findings highlights the dual nature of time series forecasting in emissions analysis. While ARIMA provides foundational analytical capabilities, it faces constraints when dealing with complex, nonlinear patterns. ARIMAX, or ARIMA with exogenous variables, represents an advancement in this approach by incorporating external factors and intervention events into the forecasting framework, resulting in enhanced predictive capabilities and more comprehensive emissions modeling [21,22,23]. Sutthichaimethee and Ariyasajjakorn [24] applied the ARIMAX model to forecast CO2 emissions, incorporating exogenous factors such as GDP growth and population growth to enhance predictive accuracy. However, the major limitation of traditional statistical methods lies in their underlying assumptions, such as normality and stationarity of the data, which may not always be held in real-world scenarios. Moreover, these methods are sensitive to issues like multicollinearity, where high correlations between independent variables can distort the results and reduce model reliability.
Machine learning has emerged as a powerful tool for forecasting CO2 emissions, leveraging advanced algorithms to analyze complex datasets and identify patterns that traditional methods may overlook. Various studies have demonstrated the effectiveness of machine learning techniques, including deep learning models, in predicting CO2 emissions across different contexts. Ağbulut [25] shows that Artificial Neural Networks (ANNs), Support Vector Machines (SVMs), and Deep Learning (DL) all provide accurate forecasts for transportation-based CO2 emissions, with SVM emerging as the best-fitting algorithm for Turkey’s transportation sector. Similarly, Sun and Liu [26] demonstrate that Least Squares Support Vector Machine (LSSVM) outperforms models like the logistic model, Artificial Neural Network (ANN), and the Grey Model (GM) in CO2 emissions forecasting in China. Yang et al. [27] employed SVR to predict the consumption of various energy sources, such as gasoline, diesel, natural gas, and crude oil, in Chongqing, China, which subsequently helped estimate CO2 emissions from these sources. Additionally, Zhu et al. [28] applied SVR to forecast CO2 emissions from the Chinese transportation industry, further showcasing the effectiveness of support vector machines in emission forecasting.
In addition to SVMs, other machine learning variants, such as Neural Networks (NNs), have demonstrated high accuracy in CO2 emission forecasting. Faruque et al. [29] conducted a comparative analysis to forecast CO2 emissions in Bangladesh using four types of NNs: Convolutional Neural Network (CNN), CNN–Long Short-Term Memory (CNN-LSTM), Long Short-Term Memory (LSTM), and Dense Neural Network (DNN), with the results showing that DNN outperformed the others. Similarly, Ghalandari et al. [30] employed two types of NNs—Group Method of Data Handling (GMDH) and Multi-Layer Perceptron (MLP)—to estimate CO2 emissions in four European countries (UK, Germany, Italy, and France), finding that MLP was preferred due to its higher accuracy. In contrast, Shabri [31] demonstrated that Lasso-GMDH outperformed the Grey Model (GM), ANN, and GMDH for forecasting CO2 emissions in Malaysia. In Thailand, Salangam [10] research revealed that ANN provided better results than Regression Analysis for forecasting transportation emissions. Similarly, Janhuaton, Ratanavaraha and Jomnonkwao [8] applied ANN to forecast CO2 emissions from Thailand’s transportation sector, showcasing its advantages over traditional methods such as ARIMAX.
The comparative analysis conducted by Tawiah et al. [32] revealed the superiority of nonlinear approaches in emissions forecasting. Their research demonstrated that NAR-based modeling significantly outperformed conventional methodologies, including standard time series techniques and traditional neural network architectures. This comprehensive evaluation, which examined multiple forecasting frameworks ranging from basic statistical methods to advanced analytical tools, established the advantages of nonlinear modeling for predicting CO2 emissions in Pakistan’s context. Similarly, Xu et al. [33] found that the NARX model outperformed both the ANN model and the Linear Regression model in forecasting CO2 emissions in China, demonstrating its superior ability to capture complex nonlinear relationships and feedback mechanisms. Table 1 provides a summary of studies related to CO2 emission forecasting.
Recent studies have shown that Fuzzy logic has been applied in emissions forecasting as an alternative to traditional statistical and machine learning models. For instance, fuzzy logic has been successfully applied in urban air quality forecasting, where it effectively captures the inherent uncertainties associated with environmental data
Beyond traditional statistical and machine learning models, fuzzy logic systems have been explored as effective tools for handling uncertainty in forecasting models. Type-2 Fuzzy Inference Systems (T2FIS) extend Type-1 Fuzzy Systems by incorporating an additional degree of uncertainty through the Footprint of Uncertainty (FOU) [43,44]. Research has demonstrated that T2FIS outperforms Type-1 systems in time-series forecasting, due to its ability to accommodate higher levels of uncertainty [45].
However, one major challenge in implementing fuzzy systems is the manual selection of fuzzy membership functions and rule sets, which can lead to suboptimal model performance. To overcome these challenges, researchers have integrated optimization techniques such as Genetic Algorithms (GAs) to tune fuzzy parameters automatically, leading to the development of GA-T2FIS models.

3. Materials and Methods

3.1. Data Collection

In the context of forecasting CO2 emissions in Thailand’s transportation sector, key variables such as population, GDP, vehicle kilometers traveled (VK), and registered vehicles have been consistently utilized. Ratanavaraha and Jomnonkwao [9], Salangam [10], and Janhuaton, Ratanavaraha and Jomnonkwao [8] used population and GDP as primary indicators, where population reflects transportation demand and GDP captures economic activity influencing transportation-related emissions. VK, used by Janhuaton, Ratanavaraha and Jomnonkwao [8], is divided into passenger vehicles, trucks (freight), and motorcycles, representing transport activity levels, while registered vehicles, utilized by Ratanavaraha and Jomnonkwao [9] and Salangam [10] provide insight into the vehicle fleet’s composition and its impact on emissions. Building on these foundations, this study utilizes a comprehensive dataset spanning from 1993 to 2022, offering a 30-year annual view sourced from official national statistics, transportation databases, and economic reports, ensuring the data’s reliability and relevance for the analysis. The trend in the data is illustrated in Figure 3. The dataset was split into training and testing sets to evaluate the model’s performance. Following standard practice, 80% of the data was allocated to the training set and 20% to the testing set.

3.2. Model Development

All model developments were implemented using MATLAB. The software was utilized for data preprocessing, model training, and performance evaluation.

3.2.1. NAR

The Autoregressive (AR) model is a traditional linear approach that predicts future values based on past observations, assuming a linear relationship between them. The mathematical formulation of the AR model presented by
y t = ϕ 1 y t − 1 + ϕ 2 y t − 2 + … + ϕ n y t − n + ε t
where y t represents the predicted value at time t , ϕ n is the autoregressive coefficients that determine the influence of past values, n is the number of time delays (lags), and ε t is the error term [46,47].
While the AR model is effective for linear time-series patterns, many real-world datasets exhibit nonlinear behaviors that AR fails to capture. To overcome this limitation, the Nonlinear Auto-Regressive (NAR) neural network model is an extension of the traditional autoregressive (AR) model by introducing a nonlinear function approximated by the neural network f ⋅ , which allows for more flexibility in capturing complex dependencies in the data [32,48,49,50]. Making it a popular choice for such predictive tasks [51]. The NAR model leverages the power of neural networks to capture nonlinear relationships within the data, which traditional linear AR models cannot adequately address. The mathematical formulation of the NAR model can be expressed as
y t = f y t − 1 ,   y t − 2 ,   y t − 3 , … ,   y t − n + ε t
The NAR model employed in this study is developed to forecast CO2 emissions based solely on historical CO2 emission data, without the inclusion of exogenous variables. As depicted in Figure 4, the network architecture consists of a neural network with a single hidden layer. To enhance the model’s predictive accuracy, hyperparameter optimization was performed using grid search. The search space included critical hyperparameters such as the number of neurons in the hidden layer and the number of time delays (lags), ranging from 1 to 10 for both parameters. The optimal configuration was determined by minimizing the Mean Squared Error (MSE) on the test set. This process yielded a model with 4 neurons in the hidden layer and 7 lags, achieving an MSE of 2,097,728.016.
To assess the adequacy of the NAR model, researchers examined the autocorrelation function (ACF) of the residuals. The ACF plot of the residuals is shown in Figure 5. The ACF plot indicates that most autocorrelations fall within acceptable limits, suggesting that the residuals exhibit minimal autocorrelation and validating the model’s effectiveness.

3.2.2. NARX

The NAR model has been continuously enhanced in its performance, leading to the development of new forms such as the Nonlinear Auto-Regressive with Exogenous Input (NARX) model. The NARX model is a time series model used to predict the future values of a target variable based on its past values and other relevant exogenous variables. It extends the traditional Auto-Regressive with Exogenous Input (ARX) model, which assumes a linear relationship between input and output variables, by incorporating nonlinear capabilities to capture more complex relationships between the variables [52]. The mathematical formulation of the NARX model can be expressed as
y t = f x t − 1 ,   x t − 2 , … ,   x t − m ,   y t − 1 ,   y t − 2 , … ,   y t − n + ε t
where x t denotes the current and past values of the external input (exogenous variables), m is the number of input delays, and ε t is the error term [33,53,54].
In this study, two distinct NARX models were developed to explore the impact of different sets of exogenous variables on CO2 emission forecasts. The network architecture of both models is illustrated in Figure 6.
  • NARX—VK includes economic variables (population and GDP) and vehicle kilometers (VK) data for motorcycles, passenger vehicles, and trucks. The hyperparameter tuning for this model was performed using grid search within a range of 1 to 10 for key parameters, including input delay, feedback delay, and the number of neurons in the hidden layer. The optimal model configuration, selected by minimizing the Mean Squared Error (MSE), consisted of 10 neurons in the hidden layer, an input delay of 6, and a feedback delay of 5, yielding an MSE of 810,401.020.
  • NARX—RG comprises economic variables (population and GDP) and registered vehicle data categorized into small, medium, and large vehicles. Similarly, grid search was utilized to tune the model’s hyperparameters, with the search space set between 1 and 10. The optimal model configuration, determined by minimizing the MSE, featured a hidden layer with 2 neurons, an input delay of 6, and a feedback delay of 2, achieving an MSE of 8,916,623.766.
The ACF plots of both models as depicted in Figure 7 suggest that most residual autocorrelations are within the confidence bounds, indicating that NARX-VK and NARX-RG model effectively capture nonlinear dependencies. The NAR and NARX models utilized in this study were enhanced with Bayesian Regularization during the training process by using the grid search method to search for the best set of hyperparameters. The ACF plots show that all models adequately capture the underlying patterns in the data.

3.2.3. Type-2 Fuzzy Inference System (T2FIS) with Genetic Algorithm (GA)

The Type-2 Fuzzy Inference System (T2FIS) is an extension of the standard fuzzy logic system that allows for better handling of uncertainties in the dataset [55]. In this study, a type-2 Sugeno fuzzy inference system model was employed, Type-2 Sugeno systems, only the input membership functions are Type-2 fuzzy sets, allowing them to handle uncertainty more effectively. The output membership functions remain the same as in a Type-1 Sugeno system, taking either a constant value or a linear function of the input values [56].
The T2FIS model employs the same dataset as NARX—VK, integrating vehicle kilometers (VK) data along with economic variables such as population and GDP. Unlike manually defined rule bases, GA was used to optimize the fuzzy rules, ensuring adaptability and improved performance. GA was also leveraged for hyperparameter tuning, specifically to refine the rule set by minimizing the Root Mean Squared Error (RMSE), ensuring optimal performance in handling nonlinear dependencies. GA operates through an iterative process of selection, crossover, and mutation to find the best set of fuzzy rules. In this study, tournament selection was used as the selection mechanism to choose the fittest individuals for reproduction. The population size was set to 200 to ensure a diverse set of solutions during the evolutionary process. The crossover rate was set at 80% probability, allowing for significant genetic diversity among offspring. The fitness function was defined as the Root Mean Squared Error (RMSE), which guided the optimization process by evaluating how well the generated rules minimized prediction errors.
The structure of the trained GA-T2FIS model is illustrated in Figure 8, which shows the structure of the trained FIS, which contains the 201 learned rules, 5 input variables (VK-Passenger, VK-Freight, VK-Motorcycle, Population, and GDP), and the CO2 emission output. The final RMSE achieved by the trained T2FIS model was 769.676, demonstrating its effectiveness in capturing the nonlinear relationships in the dataset.

3.3. Comparative Analysis Framework

This study compares the NAR model and the NARX model to evaluate their predictive accuracy and suitability for forecasting CO2 emissions in Thailand’s transportation sector. The predictive performance of NAR and NARX frameworks was assessed using Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), and Mean Absolute Percentage Error (MAPE).
The interpretation of MAPE follows established guidelines, which categorize the forecasting accuracy into four levels [25,57,58]:
  • MAPE ≤ 10%: Indicates high prediction accuracy.
  • MAPE > 10% and ≤20%: Indicates good prediction accuracy.
  • MAPE > 20% and ≤50%: Indicates reasonable prediction accuracy.
  • MAPE > 50%: Indicates inaccurate prediction accuracy.

3.4. Sensitivity Analysis Framework

In this study, the sensitivity analysis scenario focuses on the factors most relevant to CO2 emission dynamics, using variables from the best-performing NARX model. This scenario assesses the impact of transportation-related policies, providing valuable insights for stakeholders’ decision-making. To evaluate the influence of these factors on CO2 emissions, the baseline scenario is adjusted by varying the change rate of each factor by −10% and +10%, respectively [28], while other factors remain at baseline levels. This approach allows for a quantitative assessment of the impact of different factors on CO2 emissions in Thailand’s transportation sector.

4. Result and Discussion

4.1. Analysis Comparison

Table 2 presents a performance comparison of the forecasting models based on Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), and Mean Absolute Percentage Error (MAPE). The NAR model, which relies solely on past CO2 emissions, demonstrates the weakest performance, with an MAE of 5500.069, RMSE of 6339.937, and MAPE of 7.2%. indicating its limited ability to capture CO2 emission trends.
Incorporating external variables, the NARX-RG model significantly improves prediction accuracy, reducing errors to an MAE of 2636.467, RMSE of 3833.286, and MAPE of 3.7%. The NARX-VK model achieves the highest accuracy, with an MAE of 1621.449, an RMSE of 1853.799, and a MAPE of 2.2%, demonstrating that vehicle kilometers are a strong predictor of CO2 emissions in Thailand’s transportation sector.
The GA-T2FIS model achieves an MAE of 1962.723, an RMSE of 2442.957, and a MAPE of 2.6%. While its accuracy is slightly lower than that of NARX-VK, GA-T2FIS provides smoother trend forecasting with reduced sensitivity to sudden variations, as depicted in Figure 9. This stability is attributed to its fuzzy logic framework and genetic algorithm optimization, which enhance its ability to handle uncertainties and nonlinear relationships effectively.
Despite its simplicity, the NAR model serves as a useful baseline due to its straightforward implementation and lack of reliance on external variables. However, its lower accuracy underscores the advantages of incorporating additional predictors in CO2 emission forecasting. While NAR is less effective than models incorporating external variables, it remains a valuable starting point for time-series forecasting, especially in cases where external data are unavailable or unreliable.
Among all models, NARX-VK and GA-T2FIS closely follow actual CO2 emission trends, while the NAR model exhibits the largest deviations. The NARX-RG model provides moderate improvements over NAR, but the significant reduction in MAPE for both NARX-VK and GA-T2FIS highlights their superior forecasting capabilities. The GA-T2FIS model, benefiting from genetic algorithm optimization, generates smooth trend predictions with minimal sharp fluctuations. Ultimately, GA-T2FIS offers a reliable forecasting approach by balancing accuracy and generalization, whereas NARX-VK stands out as the most precise model.

4.2. Sensitivity Analysis

For the preparation of input data for predicting CO2 emissions using the NARX-VK model, the input variables are forecasted using a Nonlinear Autoregressive (NAR) model. Hyperparameter tuning was conducted for both the number of neurons in the hidden layer and the number of time delays (lags), with a search range of 1–10 for both parameters. The model’s structure and accuracy, as measured on the test set, are shown in Table 3. The MAPE (Mean Absolute Percentage Error) values indicate high accuracy across all models, confirming their reliability for further analysis.
The target year for this study is set based on the Nationally Determined Contribution (NDC) goal to reduce greenhouse gas emissions by 2030. To examine the influence of VK variables on CO2 dynamics, a sensitivity analysis was conducted, as shown in
In the Baseline Scenario, the input variables predicted in Table 3 are used to estimate CO2 emissions. The Green Transition Scenario involves the aggressive adoption of electric vehicles (EVs) and improvements in public transportation, reducing the need for private vehicle use while increasing the demand for cleaner transport. The accelerated development pathway envisions intensified economic expansion coupled with rapid urbanization trends, driving increased transportation needs. This scenario anticipates the measured implementation of sustainable technologies, contributing to elevated mobility patterns. For both this intensive growth projection and the environmental transition scenario, the analysis applies symmetrical adjustments of positive and negative 10% to baseline values across all vehicle kilometer parameters, encompassing passenger vehicles, freight transport, and motorcycles.
The results presented in Table 4 highlight the impact of different scenarios on CO2 emissions in the transportation sector for the period from 2023 to 2030. The Baseline Scenario shows a peak CO2 emission value of 80,529.515 × 103 tons, with cumulative emissions from 2023 to 2030 totaling 589,619.215 × 103 tons. This scenario assumes no significant changes in policy or vehicle usage patterns beyond the current trajectory, serving as a reference point for comparison.
In the Green Transition Scenario, where VK-passenger, VK-freight, and VK-motorcycle are reduced by 10% compared to the baseline, the peak CO2 emission value decreases to 76,967.345 × 103 tons, reflecting a 4.4% reduction. The total cumulative emissions decline to 565,217.995 × 103 tons, marking a 4.1% reduction from the baseline. This scenario underscores the importance of adopting green transportation technologies, such as electric vehicles (EVs), improved public transportation, and stricter emission regulations, demonstrating that sustainable policies can significantly lower the environmental impact of the transportation sector.
Conversely, the High Growth Scenario, which assumes a 10% increase in VK variables due to rapid economic expansion and urbanization, results in peak emissions rising to 86,363.707 × 103 tons, an increase of 7.2% over the baseline. The total cumulative emissions reach 613,683.051 × 103 tons, representing a 4.1% increase. This scenario highlights the potential negative consequences of unchecked economic growth, where increased transport demand drives CO2 emissions higher. Without the timely adoption of green technologies and policy interventions, emissions will continue to rise, emphasizing the need for proactive measures to balance economic growth and environmental sustainability.

5. Implications and Recommendations

The results from the NARX model and sensitivity analysis highlight the significant role of travel demand, as indicated by the vehicle kilometer variable, influencing CO2 dynamics. Implementing sustainable transportation policies is a key strategy to mitigate the impact of travel activities on CO2 emissions. This section explores approaches to achieve this reduction.
Stakeholder involvement is crucial in the planning and implementation of sustainable transport initiatives. Engaging a wide range of stakeholders—including government agencies, local communities, and private sector actors—ensures that diverse perspectives are considered, leading to more effective and inclusive transportation solutions [59,60]. This collaborative approach is vital for addressing the complex challenges posed by urban transportation systems, which must balance economic, social, and environmental objectives [61].
One of the primary approaches to fostering sustainable transportation is the promotion of modal interchange, which encourages the use of various sustainable transport modes such as walking, cycling, and public transport. This strategy not only enhances mobility but also reduces reliance on fossil fuel-based transportation, thereby decreasing CO2 emissions [62]. The effectiveness of such policies hinges on the seamless integration of different transport modes, allowing for efficient transitions that can significantly lower the carbon footprint associated with travel [63].
The adoption of innovative technologies also plays a significant role in enhancing the sustainability of transportation systems. Intelligent Transport Systems (ITSs) can optimize traffic management and improve operational efficiency, thereby reducing emissions associated with congestion and inefficient travel patterns [64]. Additionally, the use of alternative fuels such as electric vehicles is gaining traction as a means to decrease the carbon intensity of transportation activities [65]. These technological advancements, combined with policy measures such as fuel taxes and incentives for sustainable practices, can contribute to reducing the transportation sector’s overall environmental impact [66].
Furthermore, the concept of sustainable urban mobility is gaining prominence, particularly in the context of smart cities. The reliance on private vehicles contributes significantly to urban air pollution, greenhouse gas emissions, and traffic congestion, which adversely affect public health and urban livability [67,68]. Urban transportation planning must prioritize reducing the use of traditional vehicles, especially cars, in favor of more sustainable options that alleviate environmental pressures [69]. This shift is essential for creating urban environments that are not only more livable but also resilient to the impacts of climate change [70].

6. Conclusions

In the context of this study, the NAR model provided baseline performance, demonstrating that relying solely on past CO2 emissions limits predictive accuracy. In contrast, the NARX models, especially the NARX–VK model, outperformed the NAR model by incorporating vehicle kilometers and economic variables. This underscores the importance of exogenous inputs in capturing the broader factors that drive CO2 emissions. Notably, the NARX–VK model also showed greater accuracy than the ANN model from the previous study [8]. Despite its improved accuracy, the added complexity and data requirements of NARX models necessitate careful consideration in both model design and data selection to mitigate potential limitations such as poor data quality.
Additionally, the GA-T2FIS model demonstrated a strong capability in forecasting CO2 emissions by leveraging a rule-based structure optimized through a Genetic Algorithm (GA). While its accuracy was slightly lower than NARX-VK, it exhibited greater robustness in handling uncertainties and nonlinear relationships. The stability of GA-T2FIS highlights its potential as an alternative approach when balancing prediction accuracy and model interpretability.
The study reveals that different scenarios have substantial effects on CO2 emissions. The Green Transition Scenario shows a promising pathway to reducing emissions through technological advancements and policy measures. In contrast, the High Growth Scenario poses a risk of higher emissions if such measures are not implemented swiftly. The sensitivity analysis indicates that reducing VK values, potentially through improvements in public transportation and the adoption of electric vehicles (EVs), can significantly decrease emissions, reinforcing the case for green policies. These findings underscore the importance of aggressive, sustainable transportation policies to meet the NDC targets for 2030. Without these interventions, as shown in the High Growth Scenario, emissions are likely to increase, undermining efforts to reduce greenhouse gases. Therefore, it is crucial to focus on effectively defining the technological direction and ensuring alignment between policies and technologies to facilitate the transition toward a sustainable, low-emission transportation system. Achieving these objectives requires active stakeholder involvement in the planning and implementation of sustainable transport initiatives. The implementation of sustainable transport initiatives requires coordinated participation across multiple sectors. By fostering collaboration between public authorities, community organizations, and industry stakeholders, transportation planning benefits from comprehensive expertise and diverse viewpoints, ultimately generating more robust and equitable mobility strategies. This collaborative approach is vital for addressing the complex challenges posed by urban transportation systems, which must balance economic, social, and environmental objectives.
In conclusion, incorporating environmental issues into transportation activities requires a multifaceted approach that includes promoting modal interchange, engaging stakeholders, leveraging innovative technologies, and rethinking urban mobility strategies. By prioritizing sustainability in transportation planning and operations, it is possible to significantly reduce CO2 emissions and enhance the overall quality of urban life.

7. Limitations and Future Work

This study acknowledges several limitations that affect model performance and applicability. First, data availability and quality constraints present a challenge, as historical CO2 emission data for Thailand is limited, making it difficult to develop models with high precision. Second, while the proposed models demonstrate strong forecasting capabilities, real-world implementation remains challenging because accurate CO2 emission forecasting also requires reliable input forecasts. Additionally, this study does not comprehensively assess the socio-economic trade-offs associated with proposed mitigation strategies, necessitating further research to evaluate their feasibility, costs, and broader impacts.
Future research will focus on enhancing the T2FIS model to improve both forecasting accuracy and adaptability. One potential avenue for improvement is the incorporation of additional exogenous factors, such as infrastructure development, policy interventions, and fuel pricing trends, which may provide a more comprehensive representation of the variables influencing CO2 emissions. Additionally, the integration of fuzzy time series concepts, as demonstrated in previous research [71] will be explored to further improve predictive performance and uncertainty handling in dynamic environments.

Author Contributions

Conceptualization, T.J. and S.J.; methodology, T.J.; software, V.R. and T.C.; validation, P.W., C.S. and T.C.; formal analysis, T.J. and S.N.; investigation, T.C. and C.B.; resources, S.J. and T.C.; data curation, T.J., C.S. and C.B.; writing—original draft preparation, T.J. and S.N.; writing—review and editing, T.C., P.W. and S.J.; visualization, S.J.; supervision, V.R. and S.J.; project administration, S.J.; funding acquisition, S.J. All authors have read and agreed to the published version of the manuscript.

Funding

This research was funded by the Suranaree University of Technology (grant number: IRD7-704-65-12-23).

Data Availability Statement

The data that support the findings of this study are available from the corresponding author upon reasonable request.

Acknowledgments

This research benefited significantly from the cooperation of multiple organizations in providing access to crucial datasets including EPPO, World Bank, Bank of Thailand, and Department of Highways. The authors acknowledge the valuable support of these institutions in facilitating the successful completion of this study.

Conflicts of Interest

The authors declare no conflicts of interest.

References

  1. Jaramillo, P.; Kahn Ribeiro, S.; Newman, P.; Dhar, S.; Diemuodeke, O.E.; Kajino, T.; Lee, D.S.; Nugroho, S.B.; Ou, X.; Hammer Strømman, A.; et al. Transport. In Climate Change 2022: Mitigation of Climate Change. Contribution of Working Group III to the Sixth Assessment Report of the Intergovernmental Panel on Climate Change; Shukla, P.R., Skea, J., Slade, R., Al Khourdajie, A., van Diemen, R., McCollum, D., Pathak, M., Fradera, R., Some, P., Vyas, P., et al., Eds.; Cambridge University Press: Cambridge, UK; New York, NY, USA, 2022. [Google Scholar]
  2. International Energy Agency. Energy Consumption in Transport by Fuel in the Net Zero Scenario, 1975–2030. Available online: https://www.iea.org/data-and-statistics/charts/energy-consumption-in-transport-by-fuel-in-the-net-zero-scenario-1975-2030 (accessed on 25 September 2024).
  3. International Energy Agency. Tracking Clean Energy Progress 2023: Transport. Available online: https://www.iea.org/reports/tracking-clean-energy-progress-2023 (accessed on 25 September 2024).
  4. Funsho Idowu, O.; Adedoyin, F.F.; Bekun, F.V.; Balsalobre-Lorente, D. Modeling the environmental implications of car ownership and energy consumption in the UK: Evidence from NARDL model. Int. J. Sustain. Transp. 2022, 16, 1097–1109. [Google Scholar] [CrossRef]
  5. Praopun, A.; Poon, T. System dynamics model of CO2 emissions from urban transportation in Chiang Mai city. GEOMATE J. 2020, 18, 209–216. [Google Scholar]
  6. Umer, M.; Abas, N.; Rauf, S.; Saleem, M.S.; Dilshad, S. GHG emissions estimation and assessment of Pakistan’s power sector: A roadmap towards low carbon future. Results Eng. 2024, 22, 102354. [Google Scholar] [CrossRef]
  7. Energy Policy and Planning Office. CO2 Emission by Sector and Energy Type. Available online: https://www.eppo.go.th/index.php/th/en-energystatistics/co2-statistic (accessed on 25 September 2024).
  8. Janhuaton, T.; Ratanavaraha, V.; Jomnonkwao, S. Forecasting Thailand’s Transportation CO2 Emissions: A Comparison among Artificial Intelligent Models. Forecasting 2024, 6, 462–484. [Google Scholar] [CrossRef]
  9. Ratanavaraha, V.; Jomnonkwao, S. Trends in Thailand CO2 emissions in the transportation sector and Policy Mitigation. Transp. Policy 2015, 41, 136–146. [Google Scholar] [CrossRef]
  10. Salangam, S. Towards a predictor for CO2 emission usingregression analysis and an artificial neural network. J. Ind. Technol. Suan SunandhaRajabhat Univ. 2022, 10, 54–65. [Google Scholar]
  11. ATO. Transport and Climate Profile: Thailand. Available online: https://asiantransportobservatory.org/documents/219/Thailand-transport-and-climate-policy.pdf (accessed on 31 December 2024).
  12. ONEP. Long-Term Low Greenhouse Gas Emission Development Strategy; Office of Natural Resources and Environmental Policy and Planning: Bangkok, Thailand, 2022.
  13. Umoh, U.; Umoeka, I.; Ntekop, M.; Babalola, E. Interval type-2 fuzzy neural networks for short-term electric load forecasting: A comparative study. Int. J. Soft Comput. (IJSC) 2018, 9, 1–20. [Google Scholar] [CrossRef]
  14. Mai, D.S.; Bui, K.-T.T.; Van Doan, C. Application of Interval Type-2 Fuzzy Logic System and Ant Colony Optimization for Hydropower Dams Displacement Forecasting. Int. J. Fuzzy Syst. 2023, 25, 2052–2066. [Google Scholar] [CrossRef]
  15. Fatima, S.; Saad, S.; Zia, S.; Hussain, E.; Fraz, T.; Khan, M.S. Forecasting Carbon Dioxide Emission of Asian Countries Using ARIMA and Simple Exponential Smoothing Models. Int. J. Econ. Environ. Geol. 2019, 10, 64–69. [Google Scholar] [CrossRef]
  16. Rahman, A.; Hasan, M.M. Modeling and forecasting of carbon dioxide emissions in Bangladesh using Autoregressive Integrated Moving Average (ARIMA) models. Open J. Stat. 2017, 7, 560–566. [Google Scholar] [CrossRef]
  17. Kamoljitprapa, P.; Sookkhee, S. Forecasting models for carbon dioxide emissions in major economic sectors of Thailand. J. Phys. Conf. Ser. 2022, 2346, 012001. [Google Scholar] [CrossRef]
  18. Kour, M. Modelling and forecasting of carbon-dioxide emissions in South Africa by using ARIMA model. Int. J. Environ. Sci. Technol. 2023, 20, 11267–11274. [Google Scholar] [CrossRef]
  19. Ning, L.; Pei, L.; Li, F. Forecast of China’s Carbon Emissions Based on ARIMA Method. Discret. Dyn. Nat. Soc. 2021, 2021, 1441942. [Google Scholar] [CrossRef]
  20. Li, P.; Yang, H.; Wu, H.; Wang, Y.; Su, H.; Zheng, T.; Zhu, F.; Zhang, G.; Han, Y. Deep learning model for solar and wind energy forecasting considering Northwest China as an example. Results Eng. 2024, 24, 102939. [Google Scholar] [CrossRef]
  21. Kongcharoen, C.; Kruangpradit, T. Autoregressive integrated moving average with explanatory variable (ARIMAX) model for Thailand export. In Proceedings of the 33rd International Symposium on Forecasting, Seoul, Republic of Korea, 23–26 June 2013; pp. 1–8. [Google Scholar]
  22. Ling, A.; Darmesah, G.; Chong, K.; Ho, C. Application of ARIMAX model to forecast weekly cocoa black pod disease incidence. Math. Stat. 2019, 7, 29–40. [Google Scholar]
  23. Peter, Ď.; Silvia, P. ARIMA vs. ARIMAX–which approach is better to analyze and forecast macroeconomic time series. In Proceedings of the 30th International Conference Mathematical Methods in Economics, Karviná, Czech Republic, 11–13 September 2012; pp. 136–140. [Google Scholar]
  24. Sutthichaimethee, P.; Ariyasajjakorn, D. Forecast of Carbon Dioxide Emissions from Energy Consumption in Industry Sectors in Thailand. Environ. Clim. Technol. 2018, 22, 107–117. [Google Scholar] [CrossRef]
  25. Ağbulut, Ü. Forecasting of transportation-related energy demand and CO2 emissions in Turkey with different machine learning algorithms. Sustain. Prod. Consum. 2022, 29, 141–157. [Google Scholar] [CrossRef]
  26. Sun, W.; Liu, M. Prediction and analysis of the three major industries and residential consumption CO2 emissions based on least squares support vector machine in China. J. Clean. Prod. 2016, 122, 144–153. [Google Scholar] [CrossRef]
  27. Yang, S.; Wang, Y.; Ao, W.; Bai, Y.; Li, C. Prediction and analysis of CO2 emission in Chongqing for the protection of environment and public health. Int. J. Environ. Res. Public Health 2018, 15, 530. [Google Scholar] [CrossRef]
  28. Zhu, C.; Wang, M.; Du, W. Prediction on Peak Values of Carbon Dioxide Emissions from the Chinese Transportation Industry Based on the SVR Model and Scenario Analysis. J. Adv. Transp. 2020, 2020, 8848149. [Google Scholar] [CrossRef]
  29. Faruque, M.O.; Rabby, M.A.J.; Hossain, M.A.; Islam, M.R.; Rashid, M.M.U.; Muyeen, S.M. A comparative analysis to forecast carbon dioxide emissions. Energy Rep. 2022, 8, 8046–8060. [Google Scholar] [CrossRef]
  30. Ghalandari, M.; Forootan Fard, H.; Komeili Birjandi, A.; Mahariq, I. Energy-related carbon dioxide emission forecasting of four European countries by employing data-driven methods. J. Therm. Anal. Calorim. 2021, 144, 1999–2008. [Google Scholar] [CrossRef]
  31. Shabri, A. Forecasting the annual carbon dioxide emissions of Malaysia using Lasso-GMDH neural network-based. In Proceedings of the 2022 IEEE 12th Symposium on Computer Applications & Industrial Electronics (ISCAIE), Penang, Malaysia, 21–22 May 2022; pp. 123–127. [Google Scholar]
  32. Tawiah, K.; Daniyal, M.; Qureshi, M. Pakistan CO2 Emission Modelling and Forecasting: A Linear and Nonlinear Time Series Approach. J. Environ. Public Health 2023, 2023, 5903362. [Google Scholar] [CrossRef] [PubMed]
  33. Xu, G.; Schwarz, P.; Yang, H. Determining China’s CO2 emissions peak with a dynamic nonlinear artificial neural network approach and scenario analysis. Energy Policy 2019, 128, 752–762. [Google Scholar] [CrossRef]
  34. Amarpuri, L.; Yadav, N.; Kumar, G.; Agrawal, S. Prediction of CO2 emissions using deep learning hybrid approach: A Case Study in Indian Context. In Proceedings of the 2019 Twelfth International Conference on Contemporary Computing (IC3), Noida, India, 8–10 August 2019; pp. 1–6. [Google Scholar]
  35. Nyoni, T.; Mutongi, C. Modeling and forecasting carbon dioxide emissions in China using Autoregressive Integrated Moving Average (ARIMA) models. EPRA Int. J. Multidiscip. Res. 2019, 5, 215–224. [Google Scholar]
  36. Dansawad, N. Forecasting Carbon Dioxide Emissions (CO2) from Industry Sector in Thailand. Rajamangala Univ. Technol. Srivijaya Res. J. 2023, 15, 408–422. [Google Scholar]
  37. Wen, T.; Liu, Y.; Bai, Y.h.; Liu, H. Modeling and forecasting CO2 emissions in China and its regions using a novel ARIMA-LSTM model. Heliyon 2023, 9, e21241. [Google Scholar] [CrossRef]
  38. Kumari, S.; Singh, S.K. Machine learning-based time series models for effective CO2 emission prediction in India. Environ. Sci. Pollut. Res. 2023, 30, 116601–116616. [Google Scholar] [CrossRef]
  39. Emami Javanmard, M.; Tang, Y.; Wang, Z.; Tontiwachwuthikul, P. Forecast energy demand, CO2 emissions and energy resource impacts for the transportation sector. Appl. Energy 2023, 338, 120830. [Google Scholar] [CrossRef]
  40. Ji, T.; Li, K.; Sun, Q.; Duan, Z. Urban transport emission prediction analysis through machine learning and deep learning techniques. Transp. Res. Part D Transp. Environ. 2024, 135, 104389. [Google Scholar] [CrossRef]
  41. Karseewong, S.; Boonlha, K. Prediction of carbon dioxide emission from energy consumption in Thailand with SARIMA-ANN-REG model. Srinakharinwirot Univ. J. Sci. Technol. 2024, 16, 251667. [Google Scholar]
  42. Junsiri, C.; Sutthichaimethee, P.; Phong-a-ran, N. Modeling CO2 Emission Forecasting in Energy Consumption of the Industrial Building Sector under Sustainability Policy in Thailand: Enhancing the LISREL-LGM Model. Forecasting 2024, 6, 485–501. [Google Scholar] [CrossRef]
  43. Wagner, C.; Hagras, H. Uncertainty and Type-2 Fuzzy Sets and Systems. In Proceedings of the 2010 UK Workshop on Computational Intelligence (UKCI), Colchester, UK, 8–10 September 2010; pp. 1–5. [Google Scholar] [CrossRef]
  44. Paul, S. Chatter Mitigation in Milling Process Using Discrete Time Sliding Mode Control with Type 2-Fuzzy Logic System. Appl. Sci. 2019, 9, 4380. [Google Scholar] [CrossRef]
  45. Aladi, J.H.; Wagner, C.; Garibaldi, J.M. Type-1 or interval type-2 fuzzy logic systems—On the relationship of the amount of uncertainty and FOU size. In Proceedings of the 2014 IEEE International Conference on Fuzzy Systems (FUZZ-IEEE), Beijing, China, 6–11 July 2014; pp. 2360–2367. [Google Scholar]
  46. Kotu, V.; Deshpande, B. Chapter 12—Time Series Forecasting. In Data Science, 2nd ed.; Kotu, V., Deshpande, B., Eds.; Morgan Kaufmann: Cambridge, MA, USA, 2019; pp. 395–445. [Google Scholar]
  47. Chodakowska, E.; Nazarko, J.; Nazarko, Ł. ARIMA Models in Electrical Load Forecasting and Their Robustness to Noise. Energies 2021, 14, 7952. [Google Scholar] [CrossRef]
  48. Benrhmach, G.; Namir, K.; Namir, A.; Bouyaghroumni, J. Nonlinear Autoregressive Neural Network and Extended Kalman Filters for Prediction of Financial Time Series. J. Appl. Math. 2020, 2020, 5057801. [Google Scholar] [CrossRef]
  49. Tealab, A.; Hefny, H.; Badr, A. Forecasting of nonlinear time series using ANN. Future Comput. Inform. J. 2017, 2, 39–47. [Google Scholar] [CrossRef]
  50. Wang, Y.; Jia, R.; Dai, F.; Ye, Y. Traffic Flow Prediction Method Based on Seasonal Characteristics and SARIMA-NAR Model. Appl. Sci. 2022, 12, 2190. [Google Scholar] [CrossRef]
  51. Chen, S.; Billings, S.A. Neural networks for nonlinear dynamic system modelling and identification. Int. J. Control 1992, 56, 319–346. [Google Scholar] [CrossRef]
  52. Ljung, L. System Identification: Theory for the User; Prentice Hall PTR: Upper Saddle River, NJ, USA, 1999. [Google Scholar]
  53. Khan, N.A.; Sulaiman, M.; Tavera Romero, C.A.; Alarfaj, F.K. Theoretical Analysis on Absorption of Carbon Dioxide (CO2) into Solutions of Phenyl Glycidyl Ether (PGE) Using Nonlinear Autoregressive Exogenous Neural Networks. Molecules 2021, 26, 6041. [Google Scholar] [CrossRef]
  54. Mohebbi, M.R.; Karimi Jashni, A.; Dehghani, M.; Hadad, K. Short-Term prediction of carbon monoxide concentration using artificial neural network (NARX) without traffic data: Case study: Shiraz City. Iran. J. Sci. Technol. Trans. Civ. Eng. 2019, 43, 533–540. [Google Scholar] [CrossRef]
  55. Chao, F.; Zhou, D.; Lin, C.M.; Yang, L.; Zhou, C.; Shang, C. Type-2 Fuzzy Hybrid Controller Network for Robotic Systems. IEEE Trans. Cybern. 2020, 50, 3778–3792. [Google Scholar] [CrossRef] [PubMed]
  56. MathWorks. Type-2 Fuzzy Inference Systems. Available online: https://www.mathworks.com/help/fuzzy/type-2-fuzzy-inference-systems.html (accessed on 31 December 2024).
  57. Bakay, M.S.; Ağbulut, Ü. Electricity production based forecasting of greenhouse gas emissions in Turkey with deep learning, support vector machine and artificial neural network algorithms. J. Clean. Prod. 2021, 285, 125324. [Google Scholar] [CrossRef]
  58. Emang, D.; Shitan, M.; Abd Ghani, A.N.; Noor, K.M. Forecasting with univariate time series models: A case of export demand for peninsular Malaysia’s moulding and chipboard. J. Sustain. Dev. 2010, 3, 157. [Google Scholar] [CrossRef]
  59. Barfod, M.B. Supporting Sustainable Transport Appraisals Using Stakeholder Involvement and McDa. Transport 2018, 33, 1052–1066. [Google Scholar] [CrossRef]
  60. Szymańska, E.; Panfiluk, E.; Kiryluk, H. Innovative Solutions for the Development of Sustainable Transport and Improvement of the Tourist Accessibility of Peripheral Areas: The Case of the Białowieża Forest Region. Sustainability 2021, 13, 2381. [Google Scholar] [CrossRef]
  61. Schmale, J.; Schneidemesser, E.v.; Dörrie, A. An Integrated Assessment Method for Sustainable Transport System Planning in a Middle Sized German City. Sustainability 2015, 7, 1329–1354. [Google Scholar] [CrossRef]
  62. Conticelli, E.; Gobbi, G.; Rosas, P.I.S.; Tondelli, S. Assessing the Performance of Modal Interchange for Ensuring Seamless and Sustainable Mobility in European Cities. Sustainability 2021, 13, 1001. [Google Scholar] [CrossRef]
  63. Etukudoh, E.A. A Review of Sustainable Transportation Solutions: Innovations, Challenges, and Future Directions. World J. Adv. Res. Rev. 2024, 21, 1440–1452. [Google Scholar] [CrossRef]
  64. Grant-Muller, S.; Usher, M.D. Intelligent Transport Systems: The Propensity for Environmental and Economic Benefits. Technol. Forecast. Soc. Change 2014, 82, 149–166. [Google Scholar] [CrossRef]
  65. Yu, A.; Wei, Y.; Chen, W.; Peng, N.; Peng, L. Life cycle environmental impacts and carbon emissions: A case study of electric and gasoline vehicles in China. Transp. Res. Part D Transp. Environ. 2018, 65, 409–420. [Google Scholar] [CrossRef]
  66. López, E.; Monzon, A.; Pfaffenbichler, P. Assessment of Energy Efficiency and Sustainability Scenarios in the Transport System. Eur. Transp. Res. Rev. 2011, 4, 47–56. [Google Scholar] [CrossRef]
  67. Deweerdt, T.; Fabre, A. The Role of Land Use Planning in Urban Transport to Mitigate Climate Change: A Literature Review. Adv. Environ. Eng. Res. 2022, 3, 033. [Google Scholar] [CrossRef]
  68. Christian, D.A.; Bachtiar, A.; Candi, C. Analysis of Health-Based Transportation System for Health Transformation in DKI Jakarta. J. Soc. Res. 2023, 2, 4103–4112. [Google Scholar] [CrossRef]
  69. Mavļutova, I.; Atstāja, D.; Grasis, J.; Kuzmina, J.; Uvarova, I.; Roga, D. Urban Transportation Concept and Sustainable Urban Mobility in Smart Cities: A Review. Energies 2023, 16, 3585. [Google Scholar] [CrossRef]
  70. Lopez-Carreiro, I.; Monzon, A.; Lopez, E. MaaS Implications in the Smart City: A Multi-Stakeholder Approach. Sustainability 2023, 15, 832. [Google Scholar] [CrossRef]
  71. Bajestani, N.S.; Kamyad, A.V.; Esfahani, E.N.; Zare, A. Nephropathy forecasting in diabetic patients using a GA-based type-2 fuzzy regression model. Biocybern. Biomed. Eng. 2017, 37, 281–289. [Google Scholar] [CrossRef]
Bdcc 09 00071 g001
Figure 1. Thailand’s CO2 emission by sector in 2022.
Figure 1. Thailand’s CO2 emission by sector in 2022.
Bdcc 09 00071 g001
Bdcc 09 00071 g002
Figure 2. Thailand’s CO2 emission by oil consumption in 2022.
Figure 2. Thailand’s CO2 emission by oil consumption in 2022.
Bdcc 09 00071 g002
Bdcc 09 00071 g003
Figure 3. Normalized input and output variable.
Figure 3. Normalized input and output variable.
Bdcc 09 00071 g003
Bdcc 09 00071 g004
Figure 4. Network diagrams of NAR model.
Figure 4. Network diagrams of NAR model.
Bdcc 09 00071 g004
Bdcc 09 00071 g005
Figure 5. ACF plots of NAR model.
Figure 5. ACF plots of NAR model.
Bdcc 09 00071 g005
Bdcc 09 00071 g006
Figure 6. (a) Network diagrams of NARX—VK Model and (b) network diagrams of NARX—RG Model.
Figure 6. (a) Network diagrams of NARX—VK Model and (b) network diagrams of NARX—RG Model.
Bdcc 09 00071 g006
Bdcc 09 00071 g007
Figure 7. (a) ACF plots of NARX—VK and (b) ACF plots of NARX—RG Model.
Figure 7. (a) ACF plots of NARX—VK and (b) ACF plots of NARX—RG Model.
Bdcc 09 00071 g007
Bdcc 09 00071 g008
Figure 8. Structure of T2FIS.
Figure 8. Structure of T2FIS.
Bdcc 09 00071 g008
Bdcc 09 00071 g009
Figure 9. Comparison of actual CO2 emissions with forecasts on the test set.
Figure 9. Comparison of actual CO2 emissions with forecasts on the test set.
Bdcc 09 00071 g009
Table 1. Summary of CO2 emission forecasting studies.
Table 1. Summary of CO2 emission forecasting studies.
Author(s)YearCountryVariables ConsideredMethodologyScetorPeriodTime Step
Amarpuri et al. [34]2019IndiaHistorical emission recordCNN-LSTMAll1960–20172018–2020
Nyoni and Mutongi [35]2019ChinaHistorical emission recordARIMAAll1960–20142015–2024
Dansawad [36]2021ThailandHistorical emission record from the industrial sector in ThailandMoving Average, Trend Analysis, Single/Double/Triple Exponential Smoothing, DecompositionIndustrial2017–20203 months
Wen et al. [37]2023ChinaPopulation size, Urbanization rate, Consumption level, Primary industry, Secondary industry, Tertiary industry, Total import and export value, Energy consumption structure, Road mileage, Railway mileage, Passenger, Freight, Energy intensity, Tech progressARIMA, LSTM, ARIMA-LSTMAll1997–20172018–2025
Kumari and Singh [38]2023IndiaHistorical emission recordARIMA, SARIMAX, Holt–Winters, Random forest, Linear Regression, LSTMAll1980–20192020–2030
Emami Javanmard et al. [39]2023 GDP, Population, Number of Passengers, Load Volume in Rail Transport, Energy Types (Oil, Gas, Electricity, Renewable)AR, ARIMA, ARFIMA, SARIMA, GARCH, SVR, Grey Model, MIDAS, WOATransportation1990–20192020–2048
Ji et al. [40]2024ChinaPopulation,
Vehicle kilometers, GDP per capita, and Annual increase (Year)		ANN,
SVM,
Deep Learning		Transportation2009–20222025–2050
Karseewong and Boonlha [41]2024ThailandHistorical emission recordSARIMA-ANN-REGEnergy2022–20235 months
Junsiri et al. [42]2024ThailandGDP, Urbanization rate, Industrial structure,
net exports,
Indirect foreign investment,
Foreign tourists,
Industrial building rate,
Employment,
Health and illness, Social security, Consumer protection, Energy Consumption, Energy intensity		LISREL-LGMIndustrial1990–20232024–2033
Janhuaton, Ratanavaraha and Jomnonkwao [8]2024ThailandGDP, population, vehicle kilometerANN, SVR, ARIMAXTransportation1993–20222023–2037
Table 2. Performance comparison of different models for CO2 emission forecasting.
Table 2. Performance comparison of different models for CO2 emission forecasting.
ModelEvaluation Metric
MAE (103 Tons)RSME (103 Tons)MAPE (%)
NAR5500.0696339.9377.2
NARX—RG2636.4673833.2863.7
NARX—VK1621.4491853.7992.2
GA-T2FIS1962.7232442.9572.6
Table 3. Predictive accuracy analysis: NAR model performance for key input parameters.
Table 3. Predictive accuracy analysis: NAR model performance for key input parameters.
VariableParameterMAPE
Feedback DelayHidden Layer Size
VK—passenger8104.6%
VK—freight412.5%
VK—motorcycle712.5%
GDP753.2%
Population540.1%
Table 4. CO2 emission scenarios at different change rates (2023–2030).
Table 4. CO2 emission scenarios at different change rates (2023–2030).
ScenarioChange Rate
(%)		Peak Value (103 Tons)Change in the
Peak Value (%)		Total CO2 Emission
from 2023 to 2030		Change in Total
Volume (%)
Baseline080,529.515-589,619.215-
Green Transition−1076,967.345−4.4%565,217.995−4.1%
High Growth1086,363.7077.2%613,683.0514.1%
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.

Share and Cite

MDPI and ACS Style

Janhuaton, T.; Nanthawong, S.; Wisutwattanasak, P.; Banyong, C.; Se, C.; Champahom, T.; Ratanavaraha, V.; Jomnonkwao, S. Data-Driven Forecasting of CO2 Emissions in Thailand’s Transportation Sector Using Nonlinear Autoregressive Neural Networks. Big Data Cogn. Comput. 2025, 9, 71. https://doi.org/10.3390/bdcc9030071

AMA Style

Janhuaton T, Nanthawong S, Wisutwattanasak P, Banyong C, Se C, Champahom T, Ratanavaraha V, Jomnonkwao S. Data-Driven Forecasting of CO2 Emissions in Thailand’s Transportation Sector Using Nonlinear Autoregressive Neural Networks. Big Data and Cognitive Computing. 2025; 9(3):71. https://doi.org/10.3390/bdcc9030071

Chicago/Turabian Style

Janhuaton, Thananya, Supanida Nanthawong, Panuwat Wisutwattanasak, Chinnakrit Banyong, Chamroeun Se, Thanapong Champahom, Vatanavongs Ratanavaraha, and Sajjakaj Jomnonkwao. 2025. "Data-Driven Forecasting of CO2 Emissions in Thailand’s Transportation Sector Using Nonlinear Autoregressive Neural Networks" Big Data and Cognitive Computing 9, no. 3: 71. https://doi.org/10.3390/bdcc9030071

APA Style

Janhuaton, T., Nanthawong, S., Wisutwattanasak, P., Banyong, C., Se, C., Champahom, T., Ratanavaraha, V., & Jomnonkwao, S. (2025). Data-Driven Forecasting of CO2 Emissions in Thailand’s Transportation Sector Using Nonlinear Autoregressive Neural Networks. Big Data and Cognitive Computing, 9(3), 71. https://doi.org/10.3390/bdcc9030071

Article Metrics

Back to TopTop