Application of Artificial Intelligence to Predict CO₂ Emissions: Critical Step towards Sustainable Environment

Abstract

Prediction of carbon dioxide (CO₂) emissions is a critical step towards a sustainable environment. In any country, increasing the amount of CO₂ emissions is an indicator of the increase in environmental pollution. In this regard, the current study applied three powerful and effective artificial intelligence tools, namely, a feed-forward neural network (FFNN), an adaptive network-based fuzzy inference system (ANFIS) and long short-term memory (LSTM), to forecast the yearly amount of CO₂ emissions in Saudi Arabia up to the year 2030. The data were collected from the “Our World in Data” website, which offers the measurements of the CO₂ emissions from the years 1936 to 2020 for every country on the globe. However, this study is only concerned with the data related to Saudi Arabia. Due to some missing data, this study considered only the measurements in the years from 1954 to 2020. The 67 data samples were divided into 2 subsets for training and testing with the optimal ratio of 70:30, respectively. The effect of different input combinations on prediction accuracy was also studied. The inputs were combined to form six different groups to predict the next value of the CO₂ emissions from the past values. The group of inputs that contained the past value in addition to the year as a temporal index was found to be the best one. For all the models, the performance accuracies were assessed using the root mean squared errors (RMSEs) and the coefficient of determination (R²). Every model was trained until the smallest RMSE of the testing data was reached throughout the entire training run. For the FFNN, ANFIS and LSTM, the averages of the RMSEs were 19.78, 20.89505 and 15.42295, respectively, while the averages of the R² were found to be 0.990985, 0.98875 and 0.9945, respectively. Every model was applied individually to forecast the next value of the CO₂ emission. To benefit from the powers of the three artificial intelligence (AI) tools, the final forecasted value was considered the average (ensemble) value of the three models’ outputs. To assess the forecasting accuracy, the ensemble was validated with a new measurement for the year 2021, and the calculated percentage error was found to be 6.8675% with an accuracy of 93.1325%, which implies that the model is highly accurate. Moreover, the resulting forecasting curve of the ensembled models showed that the rate of CO₂ emissions in Saudi Arabia is expected to decrease from 9.4976 million tonnes per year based on the period 1954–2020 to 6.1707 million tonnes per year in the period 2020–2030. Therefore, the finding of this work could possibly help the policymakers in Saudi Arabia to take the correct and wise decisions regarding this issue not only for the near future but also for the far future.

1. Introduction

Saudi Arabia is the biggest exporter of fossil fuels in the world. Controlling CO₂ emissions will help in improving the quality of life all over the world. Therefore, every country has to keep an eye on the CO₂ emissions within its region. To do this, the data have to be recorded from sensors, and a qualitative prediction model has to be built. Higher prediction values from this model are alarm signs for expected harmful situations that could inform and help the authorities to take suitable actions.

Despite the fact that the increase in CO₂ emissions is harmful to human beings and the environment, an adequate amount of CO₂ is an essential requirement for plants and agriculture. According to records, the United Nations announced in the COP27 conference that was held in Sharm Elsheikh, Egypt in November 2022 that CO₂ emissions have increased by 1% in 2022 to reach 37.5 billion tonnes [1]. This big number is an alarm to the whole world that the global warming problem will continue, and hence, facing it is a must. Moreover, the promises have to be transferred to real actions on the ground [2]. To be on the right track to reducing CO₂ emissions, two strategies have to be considered seriously. The first is to decrease the utility of CO₂ emission fuels by increasing the use of other alternatives, such as renewable and green energy sources. The energy obtained from photovoltaic (PV) cells is an example of a renewable energy source. In fact, research is growing rapidly in this direction [3,4]. The second is to increase green agricultural areas. Fortunately, Saudi Arabia is currently adopting two routes to positively participate in decreasing the global warming catastrophe. To help authorities in taking the proper decision, a trustable forecasting tool is required.

The need for data forecasting can be encountered in many engineering applications. Therefore, it obtained great interest from many researchers of different disciplines and directions. To build a forecasting model, there are many tools and techniques to do so. However, building a robust model that can correctly predict future values from past values is a challenging task. In this regard, artificial intelligence (AI) and machine learning (ML) techniques occupy a substantial and competitive position among other tools. They are considered the best-nominated approaches that can perfectly handle data modelling and forecasting tasks [5,6,7,8,9].

The applications of artificial intelligence (AI) and deep learning (DL) tools are growing from day to day. These applications can be categorized into two main fields: classification and prediction. The applications of classifying tools include the fingerprint classifier, voice recognition, image, etc. However, from the prediction tool perspective, applications might include modelling and forecasting of time-series data, traffic, stock, weather, etc. The reliability of a predicting model comes from the appropriate selection of the model’s type, and then a successful training phase is implemented. There are many AI techniques to be used as predictors, such as artificial neural networks (ANNs) [10], adaptive network-based fuzzy inference systems (ANFISs) [11], long short-term memory (LSTM) [12], regression support vector machines (RSVMs) [13], autoregressive integrated moving averages (ARIMAs) [14], etc.

Similarly, AI techniques can be applied to forecast CO₂ emissions based on the measured time-series data. Vlachas et al. used LSTM as a forecasting model to forecast time-series of a high-dimensional, chaotic system. The results have been compared with those obtained by the Gaussian processes (GPs) and showed that LSTM performed better than the GP [15]. Amarpuri et. al. applied a deep learning hybrid model based on a convolution neural network (CNN) and long short-term memory network (LSTM), namely, (CNN-LSTM), to predict CO₂ levels in India in the year 2020. This study aimed to provide an estimated value of the CO₂ emissions in India, as the country had an agreement to reduce the carbon dioxide levels to 30–35% of the level in the year 2005 according to the Paris Agreement [16]. Kumari and Singh used three statistical models, two linear regression models and a deep learning model to predict CO₂ emissions in India for the next ten years based on the data from 1980 to 2019 [14]. The statistical models included the autoregressive integrated moving average (ARIMA) model, the seasonal autoregressive integrated moving average with exogenous factors (SARIMAX) model, and the Holt–Winters model. The machine learning models included linear regression and random forest techniques, while the deep learning-based model included long short-term memory (LSTM). They concluded that LSTM outperformed the other five models, which produced an RMSE value of 60.635. Meng and Noman applied four machine learning (ML) SARIMA models, namely, SARIMAX, to predict total CO₂ emissions, taking into account the effect of COVID-19 pandemic circumstances in the reduction of the emission value [17]. The study included a comparison between the four prediction models to suggest the effective one based on the lowest value of mean absolute percentage error (MAPE). The authors considered the forecasting periods as near future prediction, future prediction and far future prediction for the years 2022 to 2027, 2022 to 2054 and 2022 to 2072, respectively. They showed that the post-COVID-19 model forecasts CO₂ emission in a reasonable behaviour. Birjandi et. al. applied an artificial neural network (ANN) model of eleven neurons in the hidden layer with two different activation functions, normalized RBF and tansig, to predict CO₂ in four Southeastern countries: Malaysia, Indonesia, Singapore and Vietnam [18]. They concluded that the ANN model with the normalized RBF produced the highest correlation value with the measured data.

Researchers are promoted to apply AI techniques in the forecasting of CO₂ emissions applications. Examples of reputative AI tools are the FFNN, ANFIS and LSTM. Many recent studies used the FFNN [19,20], ANFIS [21,22] and LSTM [14,23,24] techniques to build a forecasting model of CO₂ emissions. For the FFNN modelling approach, Mutascu used the single-layer, 20-neuron feed-forward artificial neural network approach to predict CO₂ emissions in the United States of America [25]. The model was built based on the annual data collected during the years ranging from 1984 to 2020. The simulation results showed that the forecasting accuracy was 92.42%, and the forecasting error percentage was 7.58%. The RMSE and the R² were used as the statistical assessment of the modelling accuracy. Moreover, Wang et. al. combined several machine learning tools in a hybrid procedure to forecast CO₂ emissions in China [26]. They applied four two-stage procedures. Support vector regression (SVR) with an ANN was one of the used procedures. Regarding the ANFIS modelling tool, Cansiz et. al. applied some AI techniques that included an ANN and ANFIS to predict the CO₂ emissions from the transportation sector [21]. Their results showed that the ANN model was the best. For the LSTM technique, Kumari and Singh conducted a comparison study between some ML tools, including LSTM, to predict CO₂ emissions in India [14]. They concluded that the LSTM model was the best to predict CO₂ emissions, as it produced the lowest RMSE of 60.635 among the other techniques.

Particularly, some researchers applied several AI tools to forecast the CO₂ emissions in Saudi Arabia. Hamieh et. al. conducted a study to quantify the CO₂ emissions from industrial facilities in Saudi Arabia [27]. They studied the CO₂ emissions from six industrial entities, such as electricity generation, desalination, oil refining, cement, petrochemicals and iron and steel, which are responsible for more than 70% of the country’s total emissions. This was to register the details about the emission source locations, rates and characteristics. Alam and AlArjani compared CO₂ emission forecasting in Gulf countries based on an annual basis [28]. They used statistical tools such as autoregressive integrated moving averages (ARIMAs), artificial neural networks (ANNs) and Holt–Winters exponential smoothing (HWES) as the forecasting models. Their conclusion was to use the ANN model for forecasting the CO₂ in Gulf countries, as it produced the best results relative to the ARIMA and the HWES models. Alajmi studied the modelling of carbon emissions and electricity generation in SA [29]. This study used a structural time-series model (STSM) and logarithmic mean Divisia index (LMDI) for estimating the long-run elasticities. The results showed that three controlling variables such as gross domestic product (GDP), electricity generation and population are significantly affecting carbon dioxide (CO₂) emissions. Furthermore, the findings from utilizing the logarithmic mean division index (LMDI) analysis showed that there was an increase of 1377.56 million tonnes in CO₂ emissions from the three controlling factors between 1980 and 2017 in Saudi Arabia. Habadi and Tsokos built a statistical-based time-series model to forecast carbon dioxide in the atmosphere in the Middle East and the atmospheric temperature in Saudi Arabia. The model is of a seasonal autoregressive integrated moving average (seasonal-ARIMA) type, and the data are based on the monthly temperature from 1970–2015 [30]. Althobaiti and Shabri performed a comparison between the nonlinear grey Bernoulli model NGBM (1,1) and the GM (1,1) model for predicting the yearly CO₂ emissions in Saudi Arabia from 2017 to 2021 based on the data from 1970 to 2016 [31]. The study concluded that the NGBM (1,1) produced better results based on the mean absolute percentage error (MAPE), which was below 10%. Alkhathlan and Javid studied and analyzed the underlying energy demand trend (UEDT) for the total carbon emissions and carbon emissions from the domestic transport sector in Saudi Arabia using a structural time-series technique [32]. The considered data was in the period from 1971 to 2013. The study demonstrated that there is a positive correlation between the growth in real income and CO₂ emissions.

From the aforementioned literature review, it can be noticed that some researchers have been interested in CO₂ emissions either in Saudi Arabia or the Middle East area. In the following paragraph, a brief description of each study is presented but from a critical viewpoint.

In [27], the work presents a database for the volumetric concentrations of CO₂ emissions from all major stationary industrial sources in Saudi Arabia. However, the paper does not show any modelling or forecasting approach. In [28], the authors developed a comparative study between three prediction models, namely, ARIMA, ANN and Holt–Winters exponential smoothing. They used 55 data samples that represent the recordings during the years 1960:2014. Their goal was to build the best model for forecasting CO₂ emissions in the annual range of 2015:2025. The paper showed that the ANN was the best, but the paper does not show the value of the training/testing ratio to assess the model’s reliability. In [29], the authors built a structural time series model (STSM) to study the effect of gross domestic product (GDP), electricity generation and population on CO₂ emissions in Saudi Arabia between 1980 and 2017. The paper does not show any data-forecasting strategy. In [30], the authors used a statistical forecasting technique, namely, the autoregressive integrated moving average (ARIMA), to predict the atmospheric CO₂ in the entire Middle East between 1996–2015. The forecasting range was applied monthly through only the years 2014:2015. In [31], the study used a statistical method, namely, the nonlinear grey Bernoulli model, to forecast CO₂ in Saudi Arabia. The model was built using a dataset in the range of 1970:2016, and the forecasting period represents only the annual range between 2017:2021.

Despite all the abovementioned studies being related to CO₂ emissions, none of them has used machine learning or artificial intelligence tools to model CO₂ emissions in Saudi Arabia. Moreover, none of them has applied the obtained model to forecast the annual emissions between 2020 to 2030. Almost all of these studies have used statistical methods in their work. Furthermore, all of them have used an entire set of data to build a model. However, in our proposed work, three ensembled machine learning tools, such as FFNN, ANFIS and LSTM, have been used to perform the forecasting task. In addition, during the modelling procedure, the optimal training/testing ratio of 70/30 has been applied [33].

In conclusion, up to the knowledge of the authors, this is the first use of an ensemble composed of three AI and ML tools, such as FFNN, ANFIS and LSTM, to forecast CO₂ emissions in Saudi Arabia for the period from 2020 to 2030. Accordingly, the authors believe that the proposed strategy is considered novel.

The work contribution can be highlighted as follows:

• Forecasting CO₂ emissions in Saudi Arabia for the period 2020–2030, which has not been studied before.
• Forecasting models have been built using three artificial intelligence tools, namely, FFNN, ANFIS and LSTM.
• The dataset was obtained from a verified repository to ensure the obtained models are trustable [34].
• The optimal train/test ratio of 70/30 has been applied to avoid overfitting [33].

Table 1. Key comparisons between previous studies and our proposed work.

Reference	Publication Year	Model	Model Type	Modelling Range (Years)	No. of Samples	Forecasting Range (Years)	Train/Test Ratio
[27]	2022	-	-	-		-	-
[28]	2021	ARIMA, ANN and Holt–Winters Exponential Smoothing	Statistical and ANN	1960:2014	55	2015:2025	100/0
[29]	2022	STSM	Statistical	1980:2017	38	-	100/0
[30]	2017	ARIMA	Statistical	1996–2015	20	2014:2015	100/0
[31]	2022	Nonlinear Grey Bernoulli Model	Statistical	1970:2016	47	2017:2021	100/0
Proposed methodology		FFNN, ANFIS and LSTM	AI and ML	1954:2020	67	2020:2030	70/30

summarizes the key comparisons between the previous studies and our proposed work. From the table, it can be concluded that the work proposed in this research paper is novel and unique.

Fortunately, the CO₂ emission resulting from all contributors in Saudi Arabia is only 674.9 million tonnes. However, the emissions resulting from the entire world reached a value of 37.5 billion tonnes. This implies that Saudi Arabia has contributed only 1.8%. Despite this small value of contribution, the country is adopting a strategy to reduce these emissions gradually to reach net zero emissions by 2060. From this perspective, this paper presents a study to forecast CO₂ emissions using three powerful artificial intelligence tools, such as FFNN, ANFIS and LSTM, for the coming years up to 2030.

In conclusion, FFNN, ANFIS and LSTM are three machine learning tools that have proven their efficiency in data modelling and forecasting. Therefore, they are adopted in this study to build the forecasting model of CO₂ emissions in Saudi Arabia up to 2030.

The rest of the paper is organized as follows: The structure, function and mechanism of the three AI models are presented in Section 2. In Section 3, the data preparation is illustrated. However, in Section 4, the modelling phase is introduced in the results and discussion. In Section 5, the proposed ensemble model is presented. Finally, this work’s conclusions are summarized in Section 6.

2. Artificial Intelligence Models

In this study, three competitive AI modelling tools have been applied to predict CO₂ emissions in Saudi Arabia. In the following paragraphs, a general idea of the FFNN, ANFIS and LSTM AI tools will be presented in the following subsections.

2.1. FFNN Model

In 1943, McCulloch and Pitts proposed the mathematical model of the artificial neuron called the perceptron, which is a single arithmetic unit. This perceptron tries to exactly mimic the signal processing procedure of the natural neuron. Particularly, the human brain is composed of billions of these tiny neural cells. These cells are linked together to compose a network to be able to execute a certain complex task. The interconnection between neurons is accomplished through the cell’s dendrites and synapses. The dendrite is a receiving input terminal, while the synapse is a transmitting output terminal. For a neuron to receive signals from its neighbouring cells, the dendrites are connected to the synapses of these cells to receive a signal from every individual preceding it. The strength of each received signal is mainly based on the synaptic strength between the neuron and its exciting cell. This strength is mathematically modelled by a weight value. All the weighted inputs are summed together in the cell’s soma, and the output is triggered to the cell’s synapse through the axion as soon as it reaches or exceeds a certain critical value. Therefore, the output is generated by feeding the summed weighted inputs through an activation function. This output is propagated to the next cell to process the input signals with the same procedure to end up with a final output to the network. Modelling these actions is a little bit simple even it if can be used to model very complicated problems. The primary perceptron unit (artificial neuron) is shown in Figure 1a [35]. However, the mathematical representation of signal processing in the perceptron from receiving the inputs to transmitting its output is illustrated in Equations (1) and (2) [36]:

$$u_k={\sum }_{j=1}^m{\omega }_{k_j}{x}_j$$

(1)

$$y_k=\varnothing (u_k+b_k)$$

(2)

where

$u_k$: the sum of the weighted inputs;

$y_k$: the output of the k^th perceptron;

${\omega }_{k_j}$: the j^th synaptic weight sent to the k^th preceding neuron;

$x_j$: the j^th input from the preceding neuron;

$b_k$: the bias of the k^th neuron which is a constant;

$\varnothing (.)$: the activation function.

Despite the simplicity of the mathematical model of the perceptron, which is the primitive model of the artificial neuron, it can efficiently model highly nonlinear engineering problems. This satisfaction comes from the principle of “Divide and Conquer”. This principle deals with a complex problem by dividing it into several minor problems and attacks each problem separately. By adopting this strategy, combining simple processing units in a network can definitely deal with highly complicated problems efficiently. In particular, a multi-layer perceptron artificial neural network (MLPANN), known in the AI field as the ANN, is an example of this strategy. In an MLPANN, every layer is composed of many single units, and the outputs of every layer are transmitted to the next layer and so on until the final output is calculated at the last layer. The first and last layers are assigned to the inputs and outputs, respectively, while the in-between layers are hidden. It is worth mentioning that the number of hidden layers and the number of their units are problem-specific, and they can be determined by a trial-and-error technique to identify the optimal numbers. The feed-forward neural network (FFNN) is one of the ANN architectures, and it is considered in this study. The FFNN layout is shown in Figure 1b [37]. Generally speaking, the most important issue in dealing with the ANN is the determination of the best set of synaptic weights. Usually, this issue is considered an optimization problem. Therefore, the network is trained using a set of input–output experimental data to obtain the required weights’ vector that satisfies the minimum mean squared errors (MSE) between the experimental output and the network’s output. This learning procedure is called a supervised-learning technique. The most common analytical algorithm that is used to train the FFNN is back-propagation (BP). However, the ANN can also be trained by integrating other recent and modern optimization algorithms, such as GA, PSO, ACO, etc.

2.2. ANFIS Model

The concept of fuzzy sets first appeared in the mid-sixties of the last century by Lotfi Zadeh [38]. He extended and generalized the classical binary set that is composed of only two values {0, 1} to accept elements with multi-degree memberships in the range [0 1]. This new concept needed new types of logic and algebra similar to Boolean logic and Boolean algebra. Therefore, fuzzy logic (FL) and fuzzy mathematics emerged. Nevertheless, this concept is extended to fuzzy numbers, fuzzy inference, fuzzy rules, etc. The work with the fuzzy set is analogical to the real-life way of thinking. In a binary set, the sharp threshold is applied to the membership degree of its instances. With this threshold, any element either belongs (takes a value of 1) or does not belong (takes a value of 0) to the set. However, in fuzzy sets, partial membership is possible. This partial degree of membership is very natural to match our real-life style. In this regard, the representation of an event might take a value of 0.3 or 0.8 depending on its degree of belonging to such a set through predefined membership functions (MFs). Therefore, partial belonging to multiple sets at the same time is valid in the concept of fuzzy sets.

The development of fuzzy systems first emerged by applying fuzzy logic in engineering applications in the mid-seventies of the last century by Ebrahim Mamdani. Professor Mamdani proposed the Mamdani-type fuzzy rule that takes the form as shown in Equation (3) [39]:

IF a is A and b is B THEN c is C

(3)

where a and b are the input variables; c is the output variable and A, B and C are the MFs of a, b and c, respectively.

The second keystone in the development of fuzzy systems is the Sugeno-type fuzzy rule which was proposed by Takagi, Sugeno and Kang (TSK) and takes the form as in Equation (4) [39]:

IF a is A and b is B THEN c = f(a, b)

(4)

where a and b are the input variables; c is the output variable; A and B are the MFs of inputs a and b, respectively, and f(a, b) is a mathematical function of the inputs.

In 1993, Kang proposed the adaptive network-based fuzzy inference system, ANFIS [40]. The ANFIS is a product of the collaboration between the FL and the ANN to mimic the fuzzification, inference engine and defuzzification processes. Consequently, the ANFIS is emulated in a networked structure as the same as the ANN, as shown in Figure 2. The fuzzification and the defuzzification processes are responsible for converting the crisp values of the variables to fuzzy values and vice versa, respectively. The inference engine is responsible for producing the final output by inferring the output of each individual rule through a Min operator and then aggregating these outputs to deliver the final single output through a Max operator. This inference technique is called the Min–Max procedure. The optimal parameters of the MFs and the weights of the connected network are determined through a training algorithm. The ANFIS is usually trained through a supervised learning procedure that uses a hybrid algorithm including the gradient descent (GD) and the least squares estimate (LSE). However, other modern optimization techniques can also be applied to ANFIS training.

2.3. LSTM Model

Long short-term memory (LSTM) is one of the artificial intelligence and deep learning tools that proved its efficiency in prediction models. It was first proposed by Hochreiter and Schmidhuber in 1997 [41]. It can be considered a higher version of the classical recurrent neural network (RNN). However, LSTM overcomes one of the main drawbacks of using the RNN, which is known as the vanishing gradient problem [42]. The RNN is a variant of neural networks that takes the past value of the output to be truncated with the present inputs vector to produce the future value of the output. It suffers from the premature convergence problem when the optimizer gets stuck in the local optima during the training process. This occurs when the gradient is getting smaller and smaller and hence produces an unchanged weight vector. LSTM sorted out this problem by taking into consideration the significant data samples even if it has been seen for a long time. Therefore, the working mechanism of the LSTM network is composed of three gates: the forget gate, the input gate and the output gate. Despite the first gate being called the forget gate, it is actually responsible for memorising significant data from the past. The structure of the LSTM network (cell) with its operational gates is shown in Figure 3.

Equations (5)–(10) describe the working operations of the LSTM cell as follows [43]:

$$f_k=\sigma \left(x_k{U}^f+h_{k-1}{W}^f\right)$$

(5)

$$i_k=\sigma \left(x_k{U}^i+h_{k-1}{W}^i\right)$$

(6)

$$o_k=\sigma \left(x_k{U}^o+h_{k-1}{W}^o\right)$$

(7)

$${\tilde{C}}_k=\mathit{tanh}\left(x_k{U}^g+h_{k-1}{W}^g\right)$$

(8)

$$C_k=f_k{C}_{k-1}+i_k{\tilde{C}}_k$$

(9)

$$h_k=\mathit{tanh}\left(C_k\right)\ast o_k$$

(10)

where $f_k,i_k$ and $o_k$ are the responses of the forget, input and output gates, respectively; k is a time index; $x_k$ and $y_k$ are the network’s input and output vectors, respectively; $U^n$ and $W^n$ are the weight vectors associated with the input and output vectors of the gate n, respectively; $C_k$ denotes the network’s state; $\sigma$ is the sigma function that maps its inputs to the range [0 1], as in Equation (11) and $\mathrm{t}\mathrm{a}\mathrm{n}\mathrm{h}$ is the hyperbolic tangent function that maps its inputs to the range [−1 1], as in Equation (12).

$$\sigma (x)=\frac{1}{1+e^{-x}}$$

(11)

$$tanh(x)=\frac{e^x-e^{-x}}{e^x+e^{-x}}$$

(12)

3. Data Preparation

In this study, the data were downloaded from the “Our World in Data” website [34]. This data repository recorded many data with different categories for every country in the world. However, this study is only interested in the data for the CO₂ emissions in Saudi Arabia. The available data are in the range from 1936 to 2020, but the two years 1951 and 1952 are missing. Therefore, this study considered only the data in the range from 1953 to 2020, which ends up with 67 data samples. Figure 4 shows the plot of the collected CO₂ emissions in Saudi Arabia within the considered yearly range. However, without effective monitoring and controlling of the tremendous increase, it will negatively contribute to the global warming problem not only in the Middle East region but also all over the world. It is clear from the figure that since 1966, the CO₂ emissions in Saudi Arabia were increasing rapidly, and this matches with the third industrial revolution which started in 1950 by emerging the digital world, internet, computer programming and computer networks, etc.

This study addressed the selection of the data samples randomly for training and testing. In other words, to select random samples from the entire range for the training phase, the remaining samples were reserved for the testing phase. The ratio between the training and testing samples was set to 70:30 for all models. On the other hand, this study investigated the best input combination that produced the best performance.

Table 2. The description of input combinations examined in this study.

Gr #	Input Combination	Description
1	Present	CO2(t)
2	Present and change	CO2(t) + ΔCO2(t)
3	Present and two past values	CO2(t) + CO2(t − 1) + CO2(t − 2)
4	Year and present	t + CO2(t)
5	Year and present and change	t + CO2(t) + ΔCO2(t)
6	Year and present and two past values	t + CO2(t) + CO2(t − 1) + CO2(t − 2)

illustrates the six groups of the input combination that have been examined in this study to predict the next value of the CO₂ emission (CO₂(t + 1)) in Saudi Arabia, where t is an index to the year.

The data are now ready for the modelling procedure. Therefore, in the following section, the methodologies of building the considered FFNN, ANFIS and LSTM models to forecast CO₂ emissions in Saudi Arabia are described.

4. Results and Discussion

The available 67 data points were reformulated to be in an input–output data matrix form. For predicting the next sample from the present sample, the data matrix was prepared to have 66 data records. These records were divided into 46 samples for training, and the remaining 20 samples were held for testing with the optimal ratio of 70:30, respectively. The code was built using MATLAB R2022 in an intel CORE i7, 8 MB RAM computer, and the findings of the current work are presented in the following subsections. The obtained results from the addressed AI models will be presented in the following subsections.

4.1. FFNN

The study of predicting CO₂ in Saudi Arabia included the feed-forward neural network (FFNN) as one of the artificial intelligence tools that proved its efficiency in many engineering applications [44]. Despite the FFNN’s success in pattern recognition applications in the field of data classification, it also produces a completive result in case of data regression problems. Therefore, the FFNN was nominated in this study as the first AI tool. Before starting the modelling phase, the input–output data were normalized to be in the range [0 1] by dividing the whole data set by the maximum value of the input and output vectors. This normalization procedure is beneficial for the learning phase because it unifies the inputs’ ranges to the learning algorithm while looking for the optimal weights’ vector. For a fair comparison, the FFNN model was trained with the same optimal training-to-testing ratio of 70:30. The structure of the network included n-input and single output architecture with one hidden layer that contained ten perceptrons (units). The number of inputs, n, was dependent on the input combination (number of inputs in the group) to be studied. The network was trained with the backpropagation (BP) algorithm for 10 epochs until the lowest RMSE value for the testing data was reached.

Table 3. The RMSE and the coefficient of determination for the training and testing data of the FFNN model.

		RMSE			R2
Gr #	Inputs	Training	Testing	Average	Training	Testing	Average
1	Present	27.033	19.404	23.2185	0.98394	0.98956	0.98675
2	Present & Change	28.798	18.855	23.8265	0.97928	0.99208	0.98568
3	Two past values & Present	38.331	18.303	28.317	0.96089	0.99524	0.978065
4	Year & Present	25.015	14.545	19.78	0.98494	0.99703	0.990985
5	Year & Present & Change	31.858	16.869	24.3635	0.97764	0.97581	0.976725
6	Year & two past values & Present	22.947	19.481	21.214	0.98574	0.99289	0.989315

presents the resulting RMSE and R² as comparison statistical measures for the training and testing phases.

The table illustrates that the input combinations in groups (Gr. #1, Gr. #4 and Gr. #6) are the best three groups that produced the minimum average RMSE and the maximum R² values. Figure 5 presents the plots of the resulting outputs of the FFNN.

The prediction accuracy plot shown in Figure 5b demonstrates the good modelling phase where the model’s outputs are very close to the measured data samples for both training and testing subsets. These prediction accuracies are reinforced by the regression curve shown in Figure 5c, which shows that the FFNN has been trained correctly, and its output has more than a 99% correlation value with the measurements. Figure 5d illustrates also the convergence curves during the training phase, and it shows that the network is well-trained after 10 epochs. To show the entire performance of the FFNN model, the model predictions of the whole data set were plotted versus the experimental data points, as shown in Figure 6. The curves in Figure 6 show that the predicted output of the FFNN is very close to the measured data and tracks it well, which indicates the success of the training phase.

4.2. ANFIS

The ANFIS model parameters were assigned as the Gaussian-shape, the subtractive clustering (SC) and the weighted average (Wavg) for the fuzzification, rules generation and defuzzification processes, respectively. For obtaining a smoother prediction curve, the Gaussian-shape is a very suitable MF, as it gives a fine transition from one predicted point to the next. On the contrary, the other triangular or trapezoidal shapes produce jumps in the predictions. On the other hand, the SC technique produced the optimal as well as the smallest number of rules that can effectively handle the trend of the dataset. The ANFIS structure used in this work is the Sugeno-type fuzzy model. In this type, each rule considers the consequence (output) as a linear function of the antecedent (inputs). Accordingly, the defuzzification method aggregates the outputs by obtaining the weighted average of each rule to end up with a final single-valued output. To build a robust model, the process of continuous training was conducted for the ANFIS model until a stopping criterion was met. In this study, the training process stopped when the testing data’s root mean squared error (RMSE) reached the minimum value. The statistical markers were calculated in both phases to assess the performance of the resulting model in each phase.

Table 4. The RMSE and the coefficient of determination for the training and testing data of the fuzzy model.

		RMSE			R2			# Rules
Gr #	Inputs	Training	Testing	Average	Training	Testing	Average
1	Present	30.1601	13.1190	21.63955	0.9768	0.9962	0.9865	3
2	Present and change	30.6822	14.9410	22.8116	0.9773	0.9934	0.98535	4
3	Present and two past values	28.1576	14.9611	21.55935	0.9787	0.9955	0.9871	4
4	Year and present	28.0677	13.7224	20.89505	0.9821	0.9954	0.98875	3
5	Year and present and change	23.4659	16.7586	20.11225	0.9868	0.9939	0.99035	5
6	Year and present and two past values	21.3019	15.4643	18.3831	0.9894	0.9954	0.9924	4

illustrates the values of the root mean squared error (RMSE), coefficient of determination (R²) and number of generated fuzzy rules (#Rules). The first metric (RMSE) measures the accuracy of the model’s predictions, while the latter (R²) measures the tracking capability. The model is said to be better when its predictions give the lowest RMSE and the biggest R² values for both groups of datasets (training and testing). R² is defined as the squared of the correlation coefficient, which implies that its value is in the range [0 1] where 0 and 1 indicate no correlation and full correlation between the predicted and actual signals, respectively. By referring to Table 4 and according to the list of the lowest three average values and the list of the highest three coefficients of determination, it can be noticed that there are three input combinations shared between the two lists. The first combination is the year and the present value (Gr. #4), the second combination is the year and present value and its rate of change (Gr. #5), while the third combination is the year with the present value and the two past values (Gr. #6).

In the randomly selected sets of training and testing samples, the fuzzy modelling succeeded in obtaining the better model. This is because the subtractive clustering algorithm used the entire range of the input samples. On the contrary, when using sequential data sets for training and testing, the fuzzy model builds its rules based on a part of the entire range, and hence, it is difficult for it to predict data far from this trained range. It can be noticed from Table 4 that in fuzzy modelling, adding the time index to a group of inputs keeps the temporal feature to the data and hence enhances the values of average prediction and the R². This is clear when comparing the average and R² values obtained from (Gr. #1, Gr. #4) and (Gr. #2, Gr. #5) with those obtained from (Gr. #3, Gr. #6). The fuzzy model structure and the membership functions of the input combination Gr. #4 are shown in Figure 7a,b, respectively. The training and testing predictions of the ANFIS model are plotted against the first and second inputs and are shown in Figure 8a,b, respectively. However, the training and testing prediction errors plotted against the first and second inputs are shown in Figure 8c,d, respectively. The model predictions are noticed to be well-matched with the measured samples for both training and testing subsets. This is supported by the prediction accuracy curve plotted in Figure 9a. The 3D surface shown in Figure 9b is also generated to illustrate the dependency and behaviour of the output on the variation of the inputs. For more inspection of the ANFIS model predictions, the predictions in both training and testing subsets were plotted in 3D view, as shown in Figure 10a,b, respectively.

To show the entire performance of the ANFIS model, the model predictions of the whole data set were plotted versus the experimental data points, as shown in Figure 11.

The curves in Figure 11 show that the predicted output of the ANFIS is very close to the measured data and tracks it efficiently, which indicates a successful training phase.

4.3. LSTM

Before the training procedure starts, the data is standardised with its mean and standard deviation as follows:

$$y\left(t\right)=\frac{x\left(t\right)-\mu }{\sigma }$$

(13)

where $x\left(t\right)$ and $y\left(t\right)$ are the data sample and its standardized value at time t, respectively. $\mu$ and $\sigma$ are the mean and standard deviation values of the training dataset.

Table 5. The RMSE and the coefficient of determination for the training and testing data of the LSTM model.

		RMSE			R2
Gr #	Inputs	Training	Testing	Average	Training	Testing	Average
1	Present	22.5454	25.1811	23.86325	0.9888	0.9856	0.9872
2	Present and change	15.376	21.7467	18.56135	0.9948	0.9911	0.99295
3	Present and two past values	18.4012	22.5109	20.45605	0.992	0.9914	0.9917
4	Year and present	17.2924	13.5535	15.42295	0.9919	0.9971	0.9945
5	Year and present and change	14.712	19.6651	17.18855	0.9957	0.9688	0.98225
6	Year and present and two past values	18.6523	23.0988	20.87555	0.9922	0.9859	0.98905

introduces the resulting RMSE and the R² values obtained through the training and testing phases for the LSTM model. By referring to Table 5 and according to the list of the lowest three average values and the list of the highest three coefficients of determination, the three best groups are Gr. #2, Gr. #4 and Gr. #6.

Figure 12 illustrates the plots of the predicted samples using LSTM against the observed (real) data for both training and testing samples of Gr. #4. Figure 13 presents the errors between the LSTM predictions and the real data of the randomly selected subsets for the same input group. The lowest training and testing RMSE values resulting from the modelling phase were found to be 17.2924 and 13.5535, respectively, which occurred at Gr. #4.

To show the entire performance of the LSTM model, the model predictions of the whole data set were plotted versus the experimental data points, as shown in Figure 14. The curves in Figure 14 show that the predicted output of the LSTM model is very close to the measured data and tracks it well, which implies that the training phase was successful.

Referring to Table 3, Table 4 and Table 5, it can be concluded that input combination #4 is common in the best results list for the three LSTM, ANFIS and FFNN models. Therefore, the investigation study will be focused only on Gr. #4. Figure 15 shows the predictions of every individual model of the three considered AI models and their forecasting curves until the year 2030. Based on the FFNN forecasting curve, the CO₂ emissions will remain constant with a slight change for the next ten years. However, the ANFIS forecasting model predicts that the emissions will increase for the next ten years but at a smaller rate than in the previous period. On the other hand, the LSTM expects that the curve will decrease for a while and then start to increase again.

For Gr. #4, the percentage of the mean absolute error (MAPE) to the average of the testing data for the FFNN, ANFIS and LSTM models were calculated, and their values were found to be 2.939%, 2.82% and 10.6881%, respectively.

5. Ensembled Model

Based on the “no free lunch” principle, a single model cannot have the power of two united models. In this study, the overall decision was considered with the ensemble of the three AI models’ responses to benefit from the collaboration of these models. This ensemble is represented by the average value of their outputs to end up with the final single decision, as shown in Figure 16. For the entire data set, the RMSE and the R² values were calculated for each individual model as well as their averages.

Table 6. The RMSE and R² of the considered models’ responses and their average.

RMSE				R2
FFNN	ANFIS	LSTM	AVG	FFNN	ANFIS	LSTM	AVG
26.275	24.62	32.834	22.298	0.98502	0.98544	0.97458	0.98932

presents the statistical markers of the RMSE and the R² for the three considered AI models and their average values. It can be noticed that the RMSE of the ensemble model is the lowest among any individual model. Furthermore, the resulting output of the ensembled model is more correlated with the measured data than the output of any individual model based on the highest R² value. The experimental data and the prediction values of the three models and their ensemble from 1954–2030 are tabulated in

Table A1. Original and AI models’ predictions data for the CO₂ emissions in Saudi Arabia from 1954–2030.

Original Data			Predicted CO2(t + 1)
Inputs		Output	AI Models’ Predictions
t(Year)	CO2(t)	CO2(t + 1)	FFNN	ANFIS	LSTM	AVERAGE
1954	1.238	2.114	12.50611	−5.08987	−2.01652	1.799911
1955	2.114	1.63	12.679	−2.75545	−1.02799	2.965184
1956	1.63	2.099	12.89792	−1.4773	−9.73667	0.561316
1957	2.099	1.997	13.2161	0.542316	−10.5423	1.072047
1958	1.997	1.854	13.63345	2.119091	−8.34127	2.470423
1959	1.854	2.674	14.19032	3.665127	−5.0274	4.276015
1960	2.674	3.568	14.95435	5.961256	−0.62737	6.762744
1961	3.568	6.25	15.96297	8.316796	4.074516	9.451427
1962	6.25	6.939	17.34324	12.06533	9.549446	12.986
1963	6.939	7.041	19.06351	14.2668	14.49436	15.94156
1964	7.041	4.216	21.25589	16.01483	19.63215	18.96763
1965	4.216	6.407	23.88634	15.48812	24.05496	21.14314
1966	6.407	25.485	27.48476	18.87039	30.33186	25.56233
1967	25.485	29.079	33.15059	35.42178	41.40516	36.65918
1968	29.079	35.271	38.87615	39.9241	46.89222	41.89749
1969	35.271	45.251	45.87024	46.47104	54.6504	48.99723
1970	45.251	59.757	54.38627	56.00684	63.27654	57.88989
1971	59.757	70.282	64.68987	69.13748	73.45562	69.09432
1972	70.282	95.054	75.49803	79.21233	82.22738	78.97924
1973	95.054	98.702	89.86426	100.6304	96.3252	95.60662
1974	98.702	83.261	100.8192	105.461	102.1858	102.822
1975	83.261	101.46	104.6741	95.16006	102.069	100.6344
1976	101.46	118.07	118.4744	111.7087	115.6099	115.2643
1977	118.07	115.025	131.8265	127.1765	127.5863	128.8631
1978	115.025	138.005	135.3067	126.9572	132.2368	131.5002
1979	138.005	169.24	151.2927	147.8069	147.6745	148.9247
1980	169.24	175.305	172.9232	175.0902	165.833	171.2822
1981	175.305	157.893	180.114	182.1811	172.8314	178.3755
1982	157.893	160.837	168.4436	170.9286	169.5612	169.6445
1983	160.837	155.512	169.7221	175.6413	177.098	174.1538
1984	155.512	172.421	163.3358	173.8981	183.0917	173.4419
1985	172.421	204.604	175.3805	189.1726	200.2302	188.2611
1986	204.604	190.447	206.4798	214.9121	224.8404	215.4108
1987	190.447	202.252	187.7666	206.7261	227.6711	207.3879
1988	202.252	203.424	198.4083	217.1242	241.2667	218.9331
1989	203.424	208.497	196.9009	219.8459	251.2149	222.6539
1990	208.497	288.546	200.3989	225.2428	262.3358	229.3259
1991	288.546	316.175	294.0228	281.6761	305.9605	293.8864
1992	316.175	340.642	303.7727	302.1944	318.6143	308.1938
1993	340.642	328.139	302.1878	320.8028	308.4464	310.479
1994	328.139	262.421	292.5973	314.3748	270.5121	292.4948
1995	262.421	258.275	249.5026	271.1697	215.8058	245.4927
1996	258.275	215.807	246.4183	270.1953	212.6085	243.074
1997	215.807	207.24	227.3038	242.8541	214.6285	228.2622
1998	207.24	225.986	239.6036	238.7858	234.3291	237.5728
1999	225.986	296.364	265.2283	253.5679	265.5935	261.4632
2000	296.364	296.588	293.3954	305.8328	313.5931	304.2738
2001	296.588	325.698	312.3551	308.6746	328.8756	316.6351
2002	325.698	326.517	334.3511	334.7226	341.3091	336.7943
2003	326.517	394.585	360.4692	339.2313	343.7582	347.8196
2004	394.585	395.855	379.8552	405.7337	377.9214	387.8368
2005	395.855	431.292	408.6755	413.6198	399.3748	407.2234
2006	431.292	386.507	429.8014	452.2525	432.9277	438.3272
2007	386.507	432.336	472.5769	419.0285	441.094	444.2331
2008	432.336	465.844	491.8752	468.4847	479.0602	479.8067
2009	465.844	517.716	510.1533	500.8319	526.1445	512.3766
2010	517.716	497.659	514.9955	538.7184	573.7942	542.5027
2011	497.659	563.18	551.6828	535.1195	587.5349	558.1124
2012	563.18	540.805	546.4432	575.9426	600.6204	574.3354
2013	540.805	609.206	580.0512	572.1777	593.1379	581.7889
2014	609.206	674.878	571.6647	611.921	591.4464	591.6774
2015	674.878	639.056	561.8672	649.3612	593.2609	601.4964
2016	639.056	639.378	598.6502	639.9262	585.4526	608.0096
2017	639.378	621.953	613.3382	647.1517	578.1209	612.8703
2018	621.953	622.413	627.1504	646.1793	572.3751	615.2349
2019	622.413	625.508	634.0576	653.4732	569.892	619.1409
2020	625.508	672.38	638.4638	674.8708	565.2777	626.204
2021	-	-	642.2474	691.7801	563.9153	632.6476
2022	-	-	645.0641	706.6142	566.0118	639.23
2023	-	-	647.1015	720.4899	569.6484	645.7466
2024	-	-	648.4952	733.9226	573.9244	652.1141
2025	-	-	649.3532	747.15	578.4414	658.3149
2026	-	-	649.7565	760.2819	583.0188	664.3524
2027	-	-	649.762	773.3691	587.5727	670.2346
2028	-	-	649.4067	786.4351	592.0653	675.9691
2029	-	-	648.712	799.491	596.4827	681.5619
2030	-	-	647.6882	812.542	600.8256	687.0186

in Appendix A.

Figure 17 shows the predictions of the three AI models and their forecasting curves until the year 2030. The figure shows the average value of the ensembled model. From the average value curve, it can be noticed that with the current strategy, the CO₂ emissions are expected to be increased but at a slight rate. From the available measurements, the increment in CO₂ emissions was 9.4976 million tonnes/year in the period from 1953 to 2020. However, according to results from the AI models, it is expected that the increment will be only 6.1707 million tonnes/year in the period from 2020 to 2030. This implies that the increment rate decreases by 35.03%. This supports the strategy that the Saudi Arabian authorities adopted in order to decrease CO₂ emissions to reach net zero by 2060. Eventually, the findings of this study and the predictions of the artificial intelligence models suggest that keeping the rate decrement within 35% every 10 years would help in reaching net zero by 2060 or before.

To obtain a robust model, the possibility of the occurrence of overfitting has to be avoided. Thus, two strategies have been adopted in our research. The first one is to use an appropriate ratio between the training and testing samples to perform the training phase properly. In system modelling and machine learning fields, the ratio of 80:20 is commonly used [45]. In this regard, Joseph proposed a criterion for calculating the optimal training/testing splitting ratio, R_s, based on the number of parameters, n. Accordingly, this ratio is calculated as shown in Equation (14) [33]:

$$R_s=\sqrt{n}:1$$

(14)

Therefore, by following this criterion and with the number of input parameters $n=2$, the optimal training/testing ratio should be 70/30, which was adopted in our study. The second strategy is to assess the model predictions during the testing phase via the values of the statistical markers of RMSE and R². Small and big RMSE and R² values, respectively, indicate a successful training phase. Subsequently, the obtained model is reliable and can produce trustable predictions.

The models’ predictions have been validated with an unseen record of CO₂ emissions in Saudi Arabia. This measurement has recently been added to the database repository to express emissions in the year 2021, which has a value of 672.38 tonnes. The measured, predicted and average values of this validation record when applying the FFNN, ANFIS and LSTM models are illustrated in

Table 7. The measured, predicted and average values of the validating data obtained by the considered AI models.

Original Data (Measured)			Predicted CO2(t + 1)
Inputs		Output	AI Models’ Predictions
t(Year)	CO2(t)	CO2(t + 1)	FFNN	ANFIS	LSTM	Ensemble
2020	625.508	672.38	638.4638	674.8708	565.2777	626.204

. Moreover, the prediction error percentages obtained by these models, including their ensemble, are introduced in

Table 8. The prediction errors of the validation record obtained by the considered AI models.

Measured	Prediction Error (%)
CO2(t + 1)	FFNN	ANFIS	LSTM	Ensemble
672.38	5.0442	0.3704	15.9288	6.8675

. From Table 8, it can be noticed that the ANFIS model produced the best prediction accuracy with the lowest error value of 0.3704%. On the other hand, the LSTM model had the highest prediction error value that reached 15.9288%. However, the FFNN presented a moderate error value of 638.4638%. Consequently, the percentage error value of the model’s ensemble was only 6.8675%, which implies that a forecasting accuracy of 93.1325% has been gained. These obtained high-accuracy predictions proved the reliability of the adopted forecasting strategy of using artificial intelligence tools as a forecasting ensemble. By comparing our findings with the other previous works, it is noticed that the results obtained by our suggested ensemble are better than those obtained by a recent study of forecasting CO₂ emissions in the USA. This study produced a forecasting accuracy and prediction error of 92.42% and 7.58%, respectively [25]. However, our proposed methodology produced accuracy and error values of 93.1325% and 6.8675%, respectively. Furthermore, our results are compared with the one obtained in [14]. The LSTM model in [14] produced prediction RMSE and R² values of 60.635 and 0.990, respectively, while our LSTM model delivers values of 15.42295 and 0.9945, respectively. This indicates that our suggested strategy is outstanding and promising. Usually, a forecasting model is classified according to its resulting percentage error as inaccurate, reasonable, good or highly accurate when the percentage error is greater than 50%, 20–50%, 10–20% and less than 10%, respectively [46]. Therefore, the proposed FFNN and ANFIS can be classified as highly accurate models, while the LSTM model can be considered good. In addition, as our proposed ensemble produced a validating error of 6.8675%, it can certainly be classified as a highly accurate forecasting model.

In conclusion, the main contribution of this work can be summarized in the following points, which makes it novel:

• An ensemble of three AI models, namely, FFNN, ANFIS and LSTM, has been built to forecast CO₂ emissions in Saudi Arabia during the period from 2020 to 2030.
• The optimal train/test ratio of 70:30 was used throughout the training phase [33].
• The ensemble produced a forecasting error percentage of 6.8675% for the validating sample.
• The forecasting models are classified as highly accurate based on the formula in [46].

6. Conclusions

Tracking carbon dioxide (CO₂) emissions obtained major interest all over the world in order to effectively monitor global warming. Therefore, this work presents a study of using artificial intelligence (AI) tools to forecast CO₂ emissions in Saudi Arabia, the biggest fossil fuels exporter in the world, through the period from the year 2020 to 2030. Three powerful AI prediction models have been applied in this study. These models include the feed-forward neural network (FFNN), the adaptive network-based fuzzy inference system (ANFIS) and long short-term memory (LSTM). The annual measurements of CO₂ emissions cover the period from 1936 to 2020 but with some missing values. The dataset was filtered to cover only the period from 1954 to 2020. For all models, the available records were divided randomly with the optimal ratio of 70:30 into two groups for training and testing, respectively. Every model was trained until the lowest root mean squared error (RMSE) of the testing data was obtained. The study also investigated the best input combination that produces the best performance. The predicting models’ performances have been assessed based on the smallest RMSE and the biggest coefficient of determination (R²) values. The group containing the past value and the temporal annual index year was found to be the best combination. The FFNN, ANFIS and LSTM produced average RMSEs of 19.78, 20.89505 and 15.42295, respectively, while the averages of R² were found to be 0.990985, 0.98875 and 0.9945, respectively. The final predicted output was considered as the ensembled (average) value of the outputs resulting from the three AI models. The ensemble was validated with a new measurement of the year 2021, and the calculated percentage error was found to be 6.8675% with an accuracy of 93.1325%, which implies that the model is highly accurate. According to the ensemble forecasting values, the rate of CO₂ emissions is expected to decrease in Saudi Arabia from 9.4976 million tonnes per year in the period 1954:2020 to 6.1707 million tonnes per year in the period 2020–2030. This implies that the increment rate will decrease by 35.03%. Accordingly, this result will definitely encourage and help policymakers to take the right action not only for the near future but also for the far future. The findings of this study suggest that keeping the decrement rate within 35% every 10 years will help in reaching net zero CO₂ emissions in Saudi Arabia by 2060 or even before. In future work, another different ML tool, such as a regression support vector machine (RSVM), can be studied and included in the ensemble to increase the entire model’s prediction accuracy. Furthermore, the model’s predictability can be enhanced by restraining these models with any new annual measurements that might be added to the database.