Prediction of Annual Carbon Emissions Based on Carbon Footprints in Various Omani Industries to Draw Reduction Paths with LSTM-GRU Hybrid Model

Abstract

Despite global efforts to address climate change, carbon dioxide (CO₂) emissions are still on the rise. While carbon dioxide is essential for life on Earth, its increasing concentration due to human activities poses severe environmental and health risks. Therefore, accurately and efficiently predicting CO₂ emissions is essential. Hence, this research delves deeply into the prediction of CO₂ emissions by examining various deep learning models utilizing time series data to identify carbon dioxide levels in Oman. First, four important production materials of Oman (oil, gas, cement, and flaring), which have a great impact on CO₂ emissions, were selected. Then, the time series related to the release of CO₂ was collected from 1964 to 2022. After data collection, preprocessing was performed, in which outliers were removed and corrected, and data that had not been measured were completed using interpolation. Then, by dividing the data into two sections, education (1946–2004) and test (2022–2005) and creating scenarios, predictions were made. By creating four scenarios and modeling with two independent GRU and LSTM models and a hybrid LSTM-GRU model, annual carbon was predicted for Oman. The results were evaluated with three criteria: root mean square error (RMSE), mean absolute percentage error (MAPE), and correlation coefficient (r). The evaluations showed that the hybrid LSTM-GRU model with an error of 2.104 tons has the best performance compared to the rest of the models. By identifying key contributors to carbon footprints, these models can guide targeted interventions to reduce emissions. They can highlight the impact of industrial activities on per capita emissions, enabling policymakers to design more effective strategies. Therefore, in order to reduce pollution and increase the productivity of factories, using an advanced hybrid model, it is possible to identify the carbon footprint and make accurate predictions for different countries.

1. Introduction

Recently, there has been a rapid growth in the world’s population along with an accelerated development in industrial activities and consumption habits. The amounts of consumption of metals, plastics, wood, food, and various products have increased significantly [1]. This overall growth has led to an increase in the generation rates of waste, especially municipal solid waste [2,3]. Moreover, the growth rates are expected to further accelerate in the upcoming years [4]. The industrial revolution led to the consumption of a large amount of energy, and these industries depend mainly on nonrenewable sources [5]. The lack of incorporation of renewable resources causes excessive material depletion and exploitation of natural resources [6,7]. Consequently, the increase in the Gross Domestic Product (GDP) and rapid urbanization have been main factors that increase the threats to the environment, causing pollution and carbon emissions [8]. Since 1750, Human activities have been the main reason for the increase in greenhouse gas emissions and concentrations [9,10].

Gas emissions are the main by-product of industrial production [11]. They are generated by burning fossil fuels to achieve the energy required for production. Gas emissions are the direct result of the occurrence of chemical reactions during the combustion of fossil fuels [12]. Hence, developed countries contribute more to the increase in gas emissions in the atmosphere [13]. Moreover, the disposal of waste in landfills also generates gas emissions due to the decomposition process under anaerobic conditions. Landfill gases are composed of around 40% carbon dioxide and 60% methane [14]. It has also been recorded that out of the world’s total carbon dioxide emissions, 2.9% is due to the waste sector, and the percentage is expected to grow in developed countries [15]. The overall greenhouse gas concentration in the atmosphere has been increasing rather rapidly [16]. The National Aeronautics and Space Administration (NASA) has reported an increase of 47% in the concentration of carbon dioxide over the past 170 years. They reported that such an increase would have naturally occurred over the period of 20,000 years [17]. In 2022, the carbon dioxide concentrations were 2%, 7.9%, and 1.5% higher than the concentrations in 2019, 2020, and 2021, respectively [13]. To intensify the issue, it has been reported that the annual death number due to diseases caused by air pollution is approximately 6.5 million [18].

Oman is a country located in the Middle East, known for its diverse economy that heavily relies on the oil and gas industry. Oman is one of the largest oil producers in the region, with oil and gas extraction accounting for a significant portion of the country’s GDP. The country also has a thriving cement industry, with several cement factories spread across the region. These industries have played a crucial role in Oman’s economic development over the years, providing employment opportunities and contributing to the country’s overall growth. However, the oil, gas, and cement-related industries in Oman also pose significant environmental challenges, particularly in terms of their carbon footprint [19,20]. The extraction, refining, and burning of oil and gas are major sources of greenhouse gas emissions, contributing to global climate change. Additionally, the production of cement is a highly energy-intensive process that releases large amounts of carbon dioxide into the atmosphere [21].

All countries are facing challenges in reducing the harmful effects of climate change [22]. The issues have been reported on a global scale and are not limited to a particular area [23,24]. Carbon emissions have been causing natural disasters like droughts, healthcare issues, and flash floods [25]. In Oman, the effects of climate change have been observed in the increase in the frequency of tropical cyclones during the last decade [26,27,28]. Gulf counties have relatively high carbon intensities. The growth rate of greenhouse gas emissions in Oman has been recorded as 10% per year over the period of 15 years from 2000 to 2015. Records show that greenhouse gas emissions in Oman grew from 15,654 Gg in 2000 to 65,913 Gg in 2015. This growth rate of gas emissions is higher than the population growth rate and the GDP growth rate per year [4]. Additionally, the increase in the solid waste amounts has caused a high demand for landfill areas in Oman [29]. A total of 3.7% of the total gas emissions in the country are produced by the waste sector according to 2015 records [30]. In order to prevent natural disasters caused by climate change, it is crucial to work towards stabilizing gas emissions and allowing the ecosystems to adapt [31].

The United Nations Climate Change Conference COP21 has set a goal of preventing the increase in global temperature and reducing the overall emissions by 50% in 2050 [13]. Since then, countries have been attempting to limit gas emissions, improve energy efficiency, and utilize sustainable systems to mitigate the climate change issue [32]. The Omani government has also committed to taking urgent actions and decisions in line with the United Nations sustainable development goals (SDGs) [33]. A goal of achieving 10% of renewable energy by 2050 has been set. Furthermore, Oman has committed to reducing its greenhouse gas emissions by 2% by the year 2030 [30]. However, these actions have the potential of being more effective by adopting technology-based solutions that include the prediction of carbon emissions and drawing reduction paths accordingly. Prediction of the amounts of carbon emissions that will be produced is vital in order to reduce their harmful impacts [34]. Reliable measurements based on the current carbon footprint are a prerequisite for accuracy for setting reduction paths and policies [35]. Such predictions aid decision-makers towards more accurate and effective mitigation plans and avoid failures and obstacles [36,37]. Currently, machine learning, statistical models, and neural networks have been popular approaches to predicting carbon emissions due to the requirements for immediate and accurate predictions [38]. Prediction models work by considering a considerable amount of historical data and past trends, which include the carbon footprint, to produce an estimate for the future [39].

Various approaches have been developed, each leveraging different methodologies and data sources to forecast CO₂ emissions. A study utilized multiplicative regression to predict CO₂ emissions across 117 countries, incorporating 12 socioeconomic indicators such as GDP per capita, energy consumption, renewable energy share, and human development index (HDI) [40]. This approach is robust but struggles with non-linear relationships between predictors and emissions. The combination of Long Short-Term Memory (LSTM) networks and Principal Component Analysis (PCA) has been utilized for predicting CO₂ levels in China with impressive precision [41]. One study introduced a dynamic, multilayer artificial neural network model to forecast emissions in 17 key emitting countries, achieving a remarkable 96% accuracy [42]. Using GDP, urban population ratio, and trade openness as predictors, the model provides valuable insights into emission trends. The findings suggest that major emitters like China, India, and Iran will see continued increases in emissions, whereas nations such as the USA, Japan, and the UK are projected to steadily reduce theirs. Another research forecasted India’s carbon dioxide emissions for the upcoming ten years by analyzing univariate time-series data from 1980 to 2019. It utilized six distinct analytical models: ARIMA, SARIMAX, Holt-Winters, linear regression, random forest, and LSTM [43].

Considering Figure 1, which shows per capita CO₂ emissions for several important countries in the world, it can be seen that Oman has experienced an increase in per capita CO₂ emissions in recent years. This information can be found in the Energy and Environment section (https://ourworldindata.org/ (accessed on 5 December 2023)). Hence, this study aims to motivate the prediction models to predict the annual carbon emissions based on carbon footprints in various Omani industries to draw reduction paths. Measuring the amounts of carbon emissions will represent a key component in the formulation of effective policies for the mitigation of climate change and global warming, as well as the promotion of sustainable development. Therefore, the involvement of advanced and hybrid models can be very effective.

2. Materials and Methods

2.1. Study Area and Data Required

One of the main factors driving human-induced climate change is the rapid yearly growth of atmospheric carbon dioxide, making it crucial to decrease this rate for impact mitigation. Over the past 42 years, Oman’s yearly CO₂ emissions have significantly risen as a result of the consumption of crude oil and natural gas. The sudden uptick in emissions poses a serious challenge for Oman as it aims to transition to a low-carbon economy. Therefore, an attempt is made to study the country of Oman, which has a high volume of production in the oil and gas industries (Figure 2).

The economy and fiscal policies of Oman are closely linked to its vast reserves of oil and gas, which have played a pivotal role in driving the country’s economic development over the past few decades. Over the last few decades, Oman’s yearly CO₂ emissions have seen significant growth as per the search findings, largely influenced by the utilization of oil and gas, as well as cement manufacturing. Oman’s CO₂ emissions are dominated by the oil and gas sector (including energy production and fugitive emissions), which accounts for over 90% of total greenhouse gas emissions, while cement production contributes around 5%. Transportation represents the second-largest CO₂-specific source at 12–18%, depending on the year and methodology, with smaller contributions from agriculture, waste, and land-use changes [44]. The heavy reliance on fossil fuel extraction and energy-intensive industries underscores the urgency of decarbonizing these sectors to meet Oman’s net-zero targets.

Oman, with a population of about 4.5 million and heavy reliance on oil and gas (accounting for around 70% of its exports), emits approximately 20.7 million tonnes of CO₂ annually, resulting in a moderate per capita emission of about 4.6 tonnes. Compared to similar oil-dependent, low-population countries like Qatar, Kuwait, and Brunei, Oman’s total and per capita emissions are significantly lower—Qatar and Kuwait emit over 100 million tonnes annually with per capita emissions exceeding 24 and 35 tonnes, respectively, largely due to larger production scales and energy-intensive industries such as LNG and refining [44]. Brunei, with a smaller population and production scale, emits less overall but has higher per capita emissions (19.4 tonnes). Oman’s relatively lower emissions reflect its smaller oil production and ongoing efforts toward energy diversification and clean energy targets, highlighting how production scale and energy policy critically influence emissions among oil-dependent economies.

Preprocessing is a crucial step in preparing raw, unorganized data by resolving errors, filling missing values, eliminating noise, and correcting inconsistencies. It improves the quality of data, boosts the performance of machine learning models through methods like normalization and dimensionality reduction, minimizes computational effort, and enhances interpretability to support well-informed decisions.

In this study, in order to predict and evaluate annual carbon, the amount of carbon production of four important industries of Oman, namely oil, gas, cement, and flaring, was used. For this purpose, the data required for this operation was collected from the site (https://ourworldindata.org/ (accessed on 5 December 2023)) from 1964 to 2022. The statistical characteristics of the data used are shown in

Table 1. Statistical characteristics of carbon data from 1964 to 2022 [44].

Industry	Statistical Characteristics	CO2 (Tonnes per Person)
Annual	Min	0.01
	Max	17.79
	Mean	9.11
	Zero number	0
	Variance	29.67
	Skewness	0.03
	Count	59
Oil	Min	0
	Max	6.96
	Mean	2.27
	Zero number	1
	Variance	2.14
	Skewness	0.53
	Count	59
Gas	Min	0
	Max	12.42
	Mean	4.78
	Zero number	15
	Variance	19.64
	Skewness	0.54
	Count	59
Cement	Min	0
	Max	0.59
	Mean	0.22
	Zero number	21
	Variance	0.04
	Skewness	0.24
	Count	59
Flaring	Min	0
	Max	8.72
	Mean	1.81
	Zero number	7
	Variance	3.55
	Skewness	2.38
	Count	59

. After data collection, preprocessing was performed, in which outliers were removed and corrected, and data that had not been measured were completed using interpolation. In order to model and predict the amount of annual carbon, the data were divided into two sections: 70% training (2004–1964) and 30% test (2022–2005). Figure 3 shows Oman’s per capita CO₂ emissions from oil, gas, cement, and flaring from 1964 to 2022. This information can be found in the Energy and Environment section (https://ourworldindata.org/ (accessed on 5 December 2023)).

Several time series prediction methods have been applied to predict CO₂ emissions over time, utilizing both traditional statistical models and advanced machine learning approaches. Autoregressive Integrated Moving Average (ARIMA) and its variants, such as SARIMAX (Seasonal ARIMA with exogenous factors), have been widely used for CO₂ emission prediction. These models are effective for smooth and linear time series but struggle with non-linear and stochastic variations typical in real-world CO₂ data [41,45]. Models like Holt-Winters exponential smoothing have been applied to handle seasonality and trends in CO₂ emissions. However, their reliance on strong assumptions about data properties limits their adaptability to complex datasets [45]. Support Vector Machines (SVMs) have demonstrated utility in predicting CO₂ emissions, particularly when combined with additional features like climate or economic indicators. However, they require careful tuning and may not inherently handle temporal dependencies. Hence, hybrid models can somewhat improve the limitations and challenges of traditional methods.

2.2. An Overview of GRU and LSTM

Predictive analytics in this study involves merging gated recurrent unit (GRU) with long short-term memory (LSTM) techniques. Particularly for time series analysis, they have proven to be both highly effective and widely favored by researchers. Both GRU and LSTM, as RNN variants, are specifically crafted to effectively manage sequences of data while mitigating the issue of vanishing gradients often seen in traditional RNN models [46]. The differential feature between LSTM and GRU is the isolation of cell state and hidden state by the former, whereas the latter integrates them into one hidden state. LSTM excels at capturing long-term dependencies, whereas GRU is known for its efficiency [47]. Hence, there are differences between these two models, which can be combined to achieve the desired results. The components and main layers of the LSTM-GRU model are shown in Figure 4.

In this model, different layers such as fully connected, batch normalization, and Reshape were used to increase the accuracy of the model. A densely connected layer, also called a fully connected layer, is a neural network layer where each neuron is linked to all neurons in the previous layer. Through its dense connectivity, the layer can grasp sophisticated connections between input and output features. Batch normalization is a cutting-edge method utilized in the realm of deep learning to expedite the process of training neural networks. The process involves standardizing the input data for each layer, which helps maintain a consistent distribution of these inputs and lessens any internal covariate shift that could arise during the training phase [48]. The Reshape layer is used to change the shape of the input data without changing its content. The Reshape layer maintains the original data, simply altering its arrangement. The Reshape layer’s purpose involves decreasing the input data dimensionality, a practice that can counteract overfitting and enhance the model’s computational efficiency, particularly within complex neural networks. Figure 5 shows the LSTM-GRU model structure and layers used in this study for modeling in two forms: a summary graphic and a node graph plot.

Setting the parameters for deep learning models includes choosing hyperparameters—such as the number of layers, neurons, learning rate, and batch size—using methods like manual tuning or automated approaches like grid search, random search, or Bayesian optimization [48]. Meanwhile, model parameters, such as weights, are adjusted during the training phase through optimization techniques like Adam or SGD. In this study, a manual method was used to determine the appropriate parameters, so that by changing the parameters, an optimal and desirable model could be obtained.

Table 2. Hyperparameters are used to model LSTM, GRU, and LSTM-GRU models.

Type of Parameters	Values/Layer
Network Type	Feed-forward propagation
Data Division	Train (70%) Test (30%)
Number of Hidden layers (Neurons)	10–45
Batch Size	34–210
learning Function	0.01–0.027
Activation Function	Ramp, Logistic
Normalization Function	Batch Normalization
Training function	Adam

shows the hyperparameters used to model LSTM, GRU, and LSTM-GRU models.

2.2.1. Gated Recurrent Unit (GRU)

The GRU was created to combat the challenge of gradients disappearing or skyrocketing [49]. Introduced in 2014, GRUs offer a simpler approach compared to LSTM networks for recurrent neural networks. GRUs employ a gating system to selectively modify the hidden state during every time step, enabling them to efficiently capture patterns in sequential data. GRU’s primary function is to address the issue of disappearing gradients that commonly arise in traditional recurrent neural networks [50]. GRU and LSTM are sometimes seen as interchangeable because they both can deliver exceptional results in specific scenarios. The three key components of sigmoid layers in GRU are the update gate, reset gate, and tanh layer. To overcome vanishing gradient issues, GRU uses the update gate and reset gate in conjunction with making output decisions. The following formula is used to compute the update gate z_t for timestep t [50].

$$z_t=\sigma (w_z·[h_{(t-1)},x_t\left]\right)$$

(1)

The weight of x_t and h _{(t − 1)} is used in the calculation by multiplying them together and then adding the result. Following the calculation, a sigmoid activation function is implemented to normalize the output between 0 and 1. The equation for calculating the reset gate r_t at timestep t is as follows:

$$r_t=\sigma (w_r·[h_{(t-1)},x_t\left]\right)$$

(2)

After multiplying x_t and h _{(t − 1)} by their respective weights, the result is combined through calculation. The reset gate is instrumental in assisting the model to decide which past information should be neglected. Determining the current memory content. The process begins by introducing a fresh memory component that activates the reset gate to retain the pertinent information from previous experiences.

$$h_t=tanh(w·[r_t\ast h_{(t-1)},x_t\left]\right)$$

(3)

Ultimately, the unit is required to determine the h_t vector that stores details about the current unit and then transmit it down to the network. This can be calculated using the mathematical equation:

$$h_t=(1-z_t)\ast h_{(t-1)}+(z_t\ast h_t)$$

(4)

When the calculation shows that the vector z_t is near zero, a significant portion of the current information will be disregarded as it is not useful for making predictions [50].

2.2.2. Long Short-Term Memory (LSTM)

The LSTM is a different type of recurrent neural network that is capable of being trained through an optimization algorithm, such as gradient descent, on a specific training sequence. Adjusting all the weights of the LSTM network is necessary to match the error rate derivations for each weight. This memory cell is controlled by three gates [51]. The input gate determines which new information to store in the memory cell from both the current input and the previous hidden state. The Forget Gate determines which details will be eliminated from the previous memory cell state. The Output Gate determines which information should be output from the present input, past hidden state, and memory cell. The formula to calculate the Forget Gate is stated as follows:

$$p_s=\sigma (W_{p[h_{s-1},x_s]}+b_p)$$

(5)

At time s, the prior hidden state is denoted as h_s−1, while the weight matrix is represented by W_p and the input is denoted as x_s; additionally, the bias vector is b_p and the Forget Gate activation vector is simply referred to as p. By utilizing the mathematical equation provided, we can determine the effectiveness of the input gate [48].

$$q_s=\sigma (W_{q[h_s-1,x_s]}+b_s)$$

(6)

$$v_s=tanh(W_{v[h_s-1,x_s]}+b_v)$$

(7)

When considering time s within this process are variables including input gate activation vector q at that moment which leads to weight matrix W_q impacting previous hidden state h_s−1 and current input x_s alongside bias vectors b_s resulting in a new set of values for cell state stored in vs. followed by two additional variables: weight matrix W_v and bias vector b_v. The output gate is responsible for generating the final output based on the altered cell state. The following equation represents the output gate in mathematical terms [49].

$$f_s=tanh(W_{f[h_s-1,x_s]}+b_f)$$

(8)

$$h s=f_s\times tanh\left(v_s\right)$$

(9)

3. Performance Evaluation

By applying the coefficient of determination (R²), root mean square (RMSE), and mean absolute percentage error (MAPE) measures, the efficacy of annual carbon prediction models was evaluated accordingly. The statistical measure R² (or r²) shows the amount of variability in the dependent variable that can be predicted by the independent variable(s) within a regression model. Also, evaluating the goodness-of-fit between a regression model and observed data can be effectively achieved using the coefficient of determination. One common statistical measure used to quantify the magnitude of a varying quantity is the root mean square, which calculates this by taking the square root of the average squared values [52]. Its significance in science and engineering is evident through its diverse applications. Utilized widely in regression scenarios and for model evaluation, MAPE offers an intuitive view into relative error that sets it apart from other loss functions [53]. It measures the average magnitude of error produced by a model, or how far off predictions are on average.

$$R^2={({\displaystyle \frac{{\sum }_{i=1}^N\left(X_{oi}-\overline{X_o}\right)\left(X_{pi}-\overline{X_p}\right)}{\sqrt{{\sum }_{i=1}^N{\left(X_{oi}-\overline{X_o}\right)}^2·{\sum }_{i=1}^N{\left(X_{pi}-\overline{X_p}\right)}^2}}})}^2$$

(10)

$$RMSE=\sqrt{{{\displaystyle \frac{1}{N}}{\sum }_{i=1}^N(X_{pi}-X_{oi})}^2}$$

(11)

$$MAPE=100\times {\displaystyle \frac{1}{N}}{\sum }_{i=1}^N\left|{\displaystyle \frac{X_{o,i}-X_{p,i}}{X_{o,i}}}\right|$$

(12)

In this context, $X_{pi}$ refers to the estimated data and $X_{oi}$ refers to the observed data, with N being the number of available data points. The average values for estimated and observed data are given by $\overline{X_p}$ and $\overline{X_o}$.

4. Results

In this section, we discuss the evaluation of the used models and the correlation of the used data. Following the successful training of the models and verifying the results—ensuring no overfitting or underfitting—the testing phase commenced. Figure 6 shows the correlation of four important production materials in Oman that are closely related to pollution. According to this figure, it can be seen that the three materials, gas, cement, and oil, are more related to annual pollution in Oman, with correlations of 0.89, 0.91, and 0.73, respectively. On the other hand, in this research, four different scenarios (M1, M2, …, M4) were used to model annual carbon. The scenarios define the inputs utilized for model training: the first scenario incorporates oil, the second combines oil and gas, the third expands to include oil, gas, and cement, while the fourth integrates oil, gas, cement, and flaring. The classification of applied scenarios is given in

Table 3. Different combinations of inputs used to estimate the annual carbon in the studied model.

Scenario	Input				Output
M1	Oil				Annual CO2
M2	Oil	Gas			Annual CO2
M3	Oil	Gas	Cement		Annual CO2
M4	Oil	Gas	Cement	Flaring	Annual CO2

.

According to

Table 4. Evaluation parameters of the studied models in the test period.

Parameter	Testing
	Model	R	RMSE (Tonnes per Person)	MAPE (Tonnes per Person)
Annual CO2	LSTM-GRU-M1	0.881	5.129	31.034
	LSTM-GRU-M2	0.891	3.259	19.603
	LSTM-GRU-M3	0.978	2.104	11.876
	LSTM-GRU-M4	0.934	2.777	16.487
	LSTM-M1	0.859	6.393	38.748
	LSTM-M2	0.881	4.615	27.064
	LSTM-M3	0.889	3.845	23.232
	LSTM-M4	0.886	3.994	24.032
	GRU-M1	0.852	5.926	35.644
	GRU-M2	0.887	3.972	24.062
	GRU-M3	0.909	3.285	19.069
	GRU-M4	0.926	3.043	16.867

, which shows the evaluation criteria for the three GRU, LSTM, and LSTM-GRU models, it can be seen that the GRU-M4 (fourth scenario) model with an error of 3.04, a correlation of 0.926 and an average absolute percentage of error of 16.86 compared to the LSTM-M3 (third scenario) model with an error of 3.845, correlation 0.889 and average absolute percentage of error 23.23 has a high performance. On the other hand, the LSTM-GRU-M3 (third scenario) combined model, with an error of 2.104, a correlation of 0.978, and an average absolute error percentage of 11.87, has high accuracy compared to the two independent LSTM and GRU models.

LSTM and GRU models offer significant advantages for CO₂ emission prediction by effectively capturing complex temporal dependencies and non-linear patterns in time series data, often outperforming traditional statistical and simpler machine learning models in accuracy. GRUs provide a more computationally efficient alternative to LSTMs with faster training times while maintaining strong performance. However, these deep learning models require substantial computational resources, careful hyperparameter tuning, and large datasets to avoid overfitting [49,51]. Additionally, their “black box” nature limits interpretability compared to more transparent models. Overall, LSTM-GRU architectures strike a powerful but resource-intensive balance between predictive capability and complexity, making them well-suited for detailed emission prediction when sufficient data and computational capacity are available.

Research on predicting greenhouse gas emissions from U.S. industrial sources demonstrated that LSTM outperformed GRU in computational efficiency and accuracy, especially for small-sample time-series datasets, while GRU’s simpler architecture made it easier to train and handle long-term dependencies effectively [54]. Another study integrated Bi-LSTM with GRU, leveraging Bi-LSTM’s bidirectional processing to capture historical and future trends, while GRU enhanced memory efficiency and reduced overfitting through fine-tuning mechanisms [55]. These findings collectively emphasize the adaptability of LSTM-GRU models across contexts, with LSTM excelling in data-rich environments and GRU offering computational simplicity for smaller datasets or faster predictions.

The scatter and bar graphs of GRU, LSTM, and LSTM-GRU models are shown in Figure 7. According to the scatter diagram, the LSTM-M3 model with a coefficient of determination of 0.79 has more dispersion than the GRU-M4 model with a coefficient of determination of 0.857. Also, the LSTM-GRU-M3 model has a high reliability with less dispersion and a high coefficient of determination of 0.956.

On the other hand, according to the bar graph of the actual and forecast values, the LSTM model has suffered from underfitting in the years 2007 to 2014 and from 2016 to 2022. Also, the GRU-M4 model from 2005 to 2018 has suffered from a slight lack of underfitting in the predictions. The hybrid model LSTM-GRU-M3, unlike the two models GRU and LSTM, has a good forecast in most years.

According to Figure 8, which shows the over bar plot for three models, the LSTM-M3 model has less similarity than the actual values, and the difference in statistical parameters, such as average, minimum, and maximum, is very different. On the other hand, the GRU-M4 model with medium similarity to real values performs better than the LSTM-M3 model. However, the LSTM-GRU-M3 model with high similarity and low difference in statistical values has acceptable accuracy and has predicted annual carbon values for Oman with high accuracy. Based on the correlation map illustrated in Figure 6, it is evident that the LSTM-GRU-M3 model—featuring oil, gas, and cement as its three inputs—outperforms other scenarios in terms of accuracy and overall performance.

The numerical performance indicators of the LSTM-GRU model, such as RMSE, MAPE, and R², demonstrate high accuracy in predicting CO₂ emissions, which is essential for developing effective emission reduction strategies. This level of accuracy supports practical applications like identifying emission hotspots and evaluating mitigation policies. However, the interpretation of these metrics in Oman’s context requires alignment with its specific emission reduction targets. The LSTM-GRU model can assist by predicting emissions from critical sectors like oil, gas, and cement production, enabling policymakers to monitor progress and optimize interventions. However, the practical utility of such forecasts also depends on their integration into policy frameworks and real-time decision-making systems.

Oman’s CO₂ emission prediction models are closely tied to the country’s quantified emission reduction targets and sectoral strategies. By 2030, Oman aims to cut greenhouse gas emissions by 21% from 2020 levels, which equates to a reduction of about 29 million tonnes of CO₂ equivalent from a projected 2030 baseline of 137 million tonnes of CO₂, bringing emissions down to roughly 108 million tonnes of CO₂. According to the model’s forecast for 2020 (14.33) and 2022 (14.55), in tonnes per person, this value should decrease by about 1.5 percent to reach the level of 2020.

5. Discussion

Annual carbon prediction with hybrid models is an essential tool in predicting future carbon emissions and understanding the impact of human activities on the environment. By combining multiple predicting techniques such as time series analysis, machine learning algorithms, and statistical methods, hybrid models can provide more accurate and reliable predictions. Findings from the evaluation criteria and graphic diagrams are consistent with previous research [56,57]. The combination of the LSTM model with other models has led to favorable accuracy in numerous research studies across different fields. Zhao et al. [58] focused on the Hetao Plain in northern China, emphasizing the rising risk of groundwater vulnerability due to climate anomalies and human activities. By integrating Convolutional Neural Networks (CNN), Long Short-Term Memory (LSTM), and a self-attention mechanism, the research improves spatio-temporal predictions of groundwater vulnerability. Kong et al. [59] enhanced sLSTM, an NLP model, to address its limitations in time series forecasting (TSF) tasks. Introducing P-sLSTM, which incorporates patching and channel independence, the model achieves state-of-the-art results.

Due to the importance of the topic, many studies have been performed in the field of predicting annual carbon emissions with deep and hybrid learning models. Iftikhar et al. [13] conducted a comprehensive analysis to predict CO₂ emissions in Pakistan by examining various combinations of regression and time series techniques. The results showed that the combined model used with minimum error (RMSE = 0.047, MAPE = 3.762) had an acceptable performance, and the anticipated final optimal hybrid combination forecasting model indicates that Pakistan’s per capita CO₂ emissions will reach 1.130 metric tons by the year 2030. Therefore, Zhao et al. [60] explored the integration of a regression model (MIDAS) and BP neural network (MIDAS-BP model) to predict carbon dioxide emissions within the United States. The findings of this investigation indicated that utilizing the MIDAS-BP model, which boasts a 90% accuracy rate, is effective in predicting carbon dioxide emissions for both short-term and long-term scenarios. Also, a sophisticated automated integrated moving average model with external factors (ARIMAX) was designed by Shakiru et al. [61] for the prediction of CO₂ emissions in Tanzania. ARIMAX was created through the combination of a multiple linear regression model (ARIMA-MLR) and an ARIMA-Quantile regression model (ARIMA-QR) to forecast upper and lower quantiles. Both ARIMA-MLR and ARIMA-QR models have shown superior prediction accuracy compared to the traditional ARIMA model, with lower mean absolute percentage error (MAPE) and root mean square error (RMSE). ARIMA, widely used for CO₂ emissions prediction, is effective in modeling linear trends but struggles with non-linear dynamics, which hybrid ARIMA models address by integrating exogenous variables or combining with machine learning techniques for improved accuracy. Studies have demonstrated that machine learning models like random forest and deep learning methods like LSTMs outperform statistical models like ARIMA in capturing complex patterns and dependencies in emissions data, particularly when applying ensemble techniques or hybrid architectures.

To adapt CO₂ emission prediction models from Oman’s oil, gas, and cement sectors to other regions or industries, it is essential to modify key input variables and data types to reflect the new context’s unique characteristics. For different regions, this includes adjusting the energy mix, incorporating coal and renewables, regional climate factors, industrial structure, and policy frameworks. For other industries such as power generation or transportation, variables like fuel types, technology efficiency, fleet composition, and activity levels must be tailored accordingly. Additionally, data resolution (temporal and spatial) and feature engineering should be adapted to capture local emission drivers accurately. This targeted customization ensures the model remains robust and relevant, enabling precise emission forecasts and effective policy guidance in diverse settings.

According to the research performed to predict annual carbon emissions, we can point out the importance of hybrid models in this field. Hybrid approaches help mitigate biases often present in purely dynamic models. By incorporating statistical techniques, they can adjust for systematic errors without the need for extensive bias correction processes. Therefore, hybrid models excel in managing non-stationary data, which is common in real-world scenarios. Their ability to integrate both linear and non-linear components allows them to adapt better to changing patterns over time compared to traditional models that may struggle with such variability.

6. Conclusions

Integrating machine learning with carbon footprinting delivers more accurate, timely, and actionable insights for reducing emissions. These capabilities are essential for organizations aiming to meet sustainability targets and adapt to evolving climate policies, making ML a cornerstone of modern carbon management strategies. Thus, this study comprehensively predicts and evaluates the annual CO₂ emissions by considering the time series of important oil, cement, gas, and flaring productions in Oman. Therefore, two separate GRU and LSTM models were used alongside an integrated LSTM-GRU model, all of which are based on machine learning methods. The results were analyzed using three evaluation criteria: MAPE, RMSE, r, and graphical charts. Research findings indicated that the hybrid LSTM-GRU model achieved a remarkable accuracy of 94% in the third scenario, which incorporated data from oil, gas, and cement. In comparison, the standalone LSTM model within the same scenario demonstrated an accuracy of 82%. Meanwhile, the independent GRU model in the fourth scenario, which utilizes oil, gas, cement, and flaring inputs, attained an accuracy of 85%.

Despite their advantages, hybrid models face challenges. (i) The complexity of model integration, which requires careful selection of components and parameters. (ii) The need for substantial computational resources to process large datasets effectively. (iii) Ensuring that hybrid models remain operationally viable within existing prediction frameworks. Future research is likely to focus on refining these models further, enhancing their scalability, and improving their operational applicability across diverse fields. As computational techniques advance and datasets grow larger, the potential for hybrid models in prediction will continue to expand significantly. Also, in future studies, incorporating external factors such as economic growth rates, technological advancements, policy changes, and climate variability can significantly enhance the accuracy and relevance of CO₂ emission prediction models. These factors influence energy demand, industrial activity, and emission intensities, and their inclusion allows models to better capture dynamic real-world conditions.

The study focused on predicting CO₂ emissions in Oman, offering valuable insights tailored to the nation’s oil-centric economy and arid climate. However, its applicability to other countries is constrained by dependence on sector-specific variables—like energy-intensive industries—and localized environmental factors, coupled with challenges related to computational complexity and reliance on robust data systems. To ensure these models can be adapted for nations with diverse economic landscapes (e.g., agriculture-based or service-driven economies) and climates, suggested measures include: customizing models to integrate region-specific emission drivers, employing hybrid strategies to balance computational efficiency with prediction accuracy (e.g., merging machine learning algorithms with traditional statistical approaches), advancing real-time data infrastructure, and aligning emissions forecasts with localized policy frameworks for practical implementation.

To adapt the LSTM-GRU model for other regions or industries, key modifications include incorporating region-specific data (e.g., energy policies, industrial practices) or industry-specific variables (e.g., fertilizer use for agriculture, fleet data for transportation), adjusting feature engineering and hyperparameters to reflect local dynamics, and enhancing scalability through transfer learning or dynamic input handling for incomplete datasets. Data preprocessing must account for regional variability, while evaluation metrics should align with localized goals (e.g., CO₂ intensity per production unit). Integration with local monitoring systems (e.g., IoT sensors) ensures real-time applicability, enabling the model to maintain accuracy across diverse contexts while addressing unique emissions drivers and regulatory frameworks.