Predictive analytics for environmental impact: Deep learning and AI ensemble strategies in forecasting CO2 emissions

doi:10.21203/rs.3.rs-4081410/v1

Download PDF

Research Article

Predictive analytics for environmental impact: Deep learning and AI ensemble strategies in forecasting CO2 emissions

https://doi.org/10.21203/rs.3.rs-4081410/v1

This work is licensed under a CC BY 4.0 License

Version 1

posted

You are reading this latest preprint version

The forecasting of carbon dioxide (CO₂) emission trends stands as a pivotal step towards achieving environmental sustainability. As countries grapple with the challenge of curbing escalating CO₂ emissions, the significance of accurate forecasting has become increasingly pronounced in recent years. In this study, to unveil the trajectory of CO₂ emissions in Pakistan, forecasting was done through advanced artificial intelligence (AI) driven Artificial Neural Network (ANN), Gated Recurrent Unit (GRU), and Long Short-Term Memory (LSTM) models. Rigorous data preprocessing techniques were applied to historical CO₂ emissions data for Pakistan comprising of 76 points from year 1946 to 2021. Sequences were formulated to capture temporal dependencies, paving the way for model training and validation. The ANN, GRU, and LSTM models were meticulously designed, each bearing unique attributes for time series forecasting. The obtained results yielded valuable insights, epitomized by model evaluations and predictions. The ANN model did really well with a test MAE of 8.111, a test R² of 0.8614 and a test RMSE of 10.25. The GRU model, characterized by a test MAE of 7.936, a test R² of 0.8355 and a test RMSE of 11.25, proved its worth as well. In contrast, the LSTM model demonstrated excellence with a test MAE of 7.941, a test R² of 0.8586 and a test RMSE of 10.45. A novel ensemble approach, combining these three models, yielded a test MAE of 7.876, a test R² of 0.869, and a test RMSE of 10.5043. Further, the models were employed to forecast CO₂ emissions for Pakistan from the year 2022 to 2030. The insights gained from this study not only enhance our understanding of CO₂ emissions trends in Pakistan but also provide valuable guidance for global efforts to adopt cleaner lifestyles and sustainable choices, fostering a healthier planet for all.

CO2 emissions

artificial intelligence

forecasting

sustainability

deep learning

Global warming and environmental changes due to increased greenhouse gas (GHG) concentrations, especially CO₂, are significant concerns. GHGs include methane, nitrous oxides, CO₂, ozone, and water vapor. CO₂ emissions result from various sources, such as fossil fuel combustion, soil erosion, deforestation, and agriculture [1]. The global challenge posed by rising carbon dioxide (CO₂) emissions has spurred extensive research to understand their implications for climate and the environment. The Climate Action Note from the United Nations Environmental Program (UNEP) has highlighted that the global community is currently facing a critical climate emergency [2]. In the context of Pakistan, a country facing unique challenges due to CO₂ emissions, such research becomes particularly relevant. The anthropogenic sources of CO₂ emissions in Pakistan, stemming from rapid urbanization, industrialization, and agricultural activities, have the potential to badly impact local and global climate patterns [3]. According to the report of Eckstein et al. [4], Pakistan has experienced significant climate-related challenges over the past two decades. With a Climate Risk Index (CRI) score of 30.17, Pakistan is among the top 10 countries most affected by climate events from 1998 to 2017. In this context, the Government of Pakistan has implemented a number of initiatives aimed at reducing CO₂ emissions within the country. Pakistan, recognizing the pivotal role of natural elements in shaping and abating climatic challenges, has introduced robust strategies for the restoration of natural capital.

In literature, various methodologies and approaches have been explored to quantify and predict CO₂ emissions, contributing to the understanding of current emission trends in different sectors. Zhu et al. [5] reported a comprehensive selection of carbon emission factors for building materials to estimate retaining soil carbon emissions. Wu et al. [6] introduced a computational method in the context of electricity and fossil fuels to gauge carbon emissions during construction. For green buildings, Zhang et al. [7] introduced a model considering energy consumption, green spaces, and water systems. Wu et al. [8] studied carbon emissions in the Inner Mongolia Autonomous Region's industrial sector, while Ma et al. [9] evaluated carbon emissions from asphalt pavement construction. Kneese et al. [10] analyzed combustion carbon emissions based on material balance. Understanding the factors influencing CO₂ emissions is pivotal for a sustainable future and a greener society. Therefore, researchers have primarily explored the influence of various factors such as industrial structure, population, technological environment, and energy consumption on CO₂ emission [11, 12]. For example, Liu et al. [13] examined carbon emission efficiency for the Yangtze River Economic Belt through a spatial panel measurement model. Gong et al. studied the impact of industrial structure and environmental regulation in the One Belt, One Road (OBOR) region [14].

Recently, AI and ML methodologies have been successfully employed by researchers to predict CO₂ emissions from various anthropogenic sources. Masini et al. [15] explored supervised ML and high-dimensional models for time series forecasting, and Crespo Cuaresma et al. [16] investigated univariate models such as Auto Regressive for time series problems. Furthermore, Elsworth et al. [17] proposed an LSTM model addressing limitations such as sensitivity of hyperparameters for any time series model. In a similar way, artificial intelligence (AI) methodologies can be employed to predict CO₂ emissions using time-series data. For instance, Amarpuri [18] predicted India's CO₂ emission levels using a deep learning hybrid approach (LSTM-CNN). Kumari and Singh employed a trio of statistical approaches, consisting of couple of regression models and a modern deep learning model, to foresee CO₂ emissions trend in the region of India over the coming decade [19]. Their findings concluded the superior performance of the LSTM model in CO₂ emissions prediction, outperforming the other techniques with the minimum Root Mean Square Error (RMSE) of 60.635. In recent times, several studies have employed the FFNS and ANFIS for prediction of CO₂ emission [20–22]. For instance, Mutascu et al. considered a single-layer, 20-neuron feed-forward ANN to forecast CO₂ trend within the United States of America (USA) [23] using FFNN modeling approach. Alam et al. concluded the superiority of ANN model in forecasting CO₂ trends across Gulf countries through employing analytical methodologies encompassing autoregressive integrated moving averages (ARIMAs), artificial neural networks (ANNs), and Holt–Winters exponential smoothing (HWES) [24] Zuo et al. [25] proposed an integrated LSTM-STRIPAT model for CO₂ emission prediction in China. Nasser et al. [26] proposed three robust artificial intelligence techniques, specifically the feed-forward neural network (FFNN), adaptive network-based fuzzy inference system (ANFIS), and long short-term memory (LSTM) for the prediction of CO₂ emissions in Saudi Arbia.

Knowing the capabilities of the above models, it is imperative to explore their comparative effectiveness in predicting CO₂ emissions in Pakistan. This study aims to address the challenges posed by CO₂ emissions in Pakistan comprehensively by evaluating and comparing the predictive capabilities of ANN, GRU, and LSTM models. Furthermore, this study explores the potential of an ensemble approach to enhance predictive accuracy to gain results near possible to reality. This research aims to provide valuable insights into the future trajectory of CO₂ emissions in Pakistan, contributing to informed policy formulation and sustainable development.

This section presents the methodology employed for forecasting carbon dioxide (CO₂) emissions using three specific artificial intelligence techniques: Artificial Neural Network (ANN), Gated Recurrent Unit (GRU), and Long Short-Term Memory (LSTM). This process involves data pre-processing and an explanation of the machine learning models' architectures.

2.1. Data Pre-Processing

Data preprocessing is an important step in any data-driven research to ensure that the raw data is transformed into a structured and suitable format for the required analysis. In this study, the historical carbon dioxide (CO₂) emissions data for Pakistan is sourced from the "Our World in Data" repository, covering the years from 1946 to 2021 which becomes 76 data points in total [27]. The data was subjected to several preprocessing steps to enhance its quality and relevance for further analysis. To ensure uniformity and ease of analysis, the CO₂ emissions values were transformed from their original units to million tonnes. This unit transformation was carried out by dividing the emissions values by 1,000,000. As duplicates, if present, could lead to biased analysis and erroneous results, so, to ensure data integrity and accuracy, a check for duplicate records within the dataset was applied. However, in the current study, no duplicates were found within the dataset, affirming the reliability of the collected data. To ensure that the input data is within a consistent numerical range, the CO₂ emissions values were scaled using the Min-Max scaling technique which is commonly used in scaling datasets for normalization purposes. Min-Max scaling transformed the emissions values to a range between 0 and 1, preserving the data's relative relationships.

The preprocessed dataset was separated into training and testing parts to facilitate model training, and evaluation. The split ratio of 80% training data and 20% testing data was used. To capture temporal patterns and relationships in the CO₂ emissions data, sequences of emissions values were generated. The sequence length was set to one, implying that each emissions value is considered in relation to its immediate predecessor. These datasets, containing scaled emissions values organized into sequences, formed the basis for constructing and training artificial neural network (ANN), Gated Recurrent Unit (GRU), and Long Short-Term Memory (LSTM) models for predicting the CO₂ emissions in Pakistan.

A data visualization step was done to gain a comprehensive understanding of the original CO₂ emissions data trends in Pakistan. The scatter plot (Fig. 1) with a blue line depicts the patterns and fluctuations in CO₂ emissions over the years. The \(x\)-axis of the graph shows the year values while the \(y\)-axis represents the CO₂ emission values in million tonnes.

2.2. Machine Learning Models

In this section, the chosen models, namely the Artificial Neural Network (ANN), Gated Recurrent Unit (GRU), and Long Short-Term Memory (LSTM), are introduced, followed by the procedural details of model training.

2.2.1. Artificial Neural Networks (ANN)

The Artificial Neural Network (ANN) is a popular deep learning model inspired by the human brain's neural network structure. It consists of interconnected layers of artificial neurons that process and transform input data to produce the desired output. The ANN model is adept at capturing complex relationships within data, making it suitable for time-series forecasting tasks [28].

In this study, the ANN model was implemented for forecasting the CO₂ emission in Pakistan. Before feeding data into the ANN, it was reshaped. It was shaped into a structure that the ANN can understand. The ANN model consisted of four layers: the input layer, two hidden layers with 64 and 32 neurons respectively, and the output layer. To make sure that the ANN model learned effectively, dropout layers were added between dense layers. These are like a break for the ANN, helping it focus better. The activation function used is 'ReLU' (Rectified Linear Unit), which helps the model capture complex relationships in the data. After building the model, it was compiled by specifying the loss function and optimizer. In presented case, the loss function is 'mean_squared_error', which calculates the difference between estimated and real values. The optimizer 'Adam' adjusts the model's weights to minimize the loss. Next, the ANN model was trained using the training data. Early stopping was set up to prevent overfitting by monitoring validation loss. The model was trained over 50 epochs with a batch size of 1. In this way, ANN model underwent the training process with the aim of learning the patterns in CO₂ emissions data and making accurate predictions for the future. Here is the summary of the model’s architecture (Table 1.).

Table 1

ANN model summary
Layer (type)	Output Shape	Param #
Dense	(None, 64)	128
Dropout	(None, 64)	0
Dense_1	(None, 32)	2,080
Dropout_1	(None, 32)	0
Dense_2	(None, 1)	33
Training Data	80%
Validation Data	20%
Loss Function	mean_squared_error
Optimizer	Adam
Activation Function	ReLU
Epochs	50
Batch Size	1
Stopping Criteria	Early stopping
Performance Metrics	R², RMSE, MAE

2.2.2. Gated Recurrent Unit (GRU)

The Gated Recurrent Unit (GRU) is a type of recurrent neural network (RNN) architecture that has gained prominence in time series forecasting due to its ability to handle sequential data efficiently [29, 30]. Unlike traditional RNNs, GRUs incorporate gating mechanisms that enable them to capture both short and long-term dependencies in sequences. This makes them particularly suited for modeling complex temporal relationships present in time series data, such as the CO₂ emissions dataset considered in this study.

In present study, we have harnessed the power of the GRU architecture to predict CO₂ emissions in Pakistan's time-series data. The GRU model starts with a GRU layer of 64 units, followed by a dropout layer for regularization. The return sequences parameter is set to True, allowing the layer to return sequences of outputs for each time step. The subsequent GRU layer with 32 units further refines the learned representations. Finally, a dense layer is added to produce the output. The model is compiled using the mean squared error loss function and the Adam optimizer. To prevent overfitting, early stopping is applied with a patience of 10 epochs and the restoration of best weights. The training process iterates over 50 epochs with a batch size of 1, allowing the model to progressively refine its internal representations based on the input data. The GRU model's multi-layer architecture and its sequential processing make it adept at capturing intricate patterns in time-series data. By leveraging the temporal relationships within CO₂ emissions data, our GRU-based approach offers the potential to yield accurate predictions. Here’s GRU model summary (Table 2.).

Table 2

GRU model summary
Layer (type)	Output Shape	Param #
GRU	(None, 1, 64)	12,864
Dropout	(None, 1, 64)	0
GRU_1	(None, 32)	9,408
Dense	(None, 1)	33
Training Data	80%
Validation Data	20%
Loss Function	mean_squared_error
Optimizer	Adam
Epochs	50
Batch Size	1
Return Sequence	True
Stopping Criteria	Early stopping
Performance Metrics	R², RMSE, MAE

2.2.3. Long Short-Term Memory (LSTM)

The Long Short-Term Memory (LSTM) is a specialized type of recurrent neural network (RNN) architecture designed to capture and learn long-range dependencies in sequential data [31]. This makes LSTM models particularly well-suited for time series forecasting, where patterns can span across multiple time steps. LSTMs incorporate memory cells that can store and retrieve information over extended periods, making them effective in understanding the intricate temporal relationships within time series data, such as the CO₂ emissions dataset under investigation.

The architecture of the LSTM model employed in this study consisted of two LSTM layers, each containing 264 units. The first LSTM layer returns sequences of outputs for each time step due to the “return_sequences = True” parameter. A dropout layer with a dropout rate of 0.2 is introduced for regularization purposes. The second LSTM layer enhances the learned representations further. Subsequently, two fully connected dense layers are employed to refine and produce the final output. The model is compiled with the mean squared error loss and the Adam optimizer. The LSTM model is trained using the training data and validated using the testing data. To prevent overfitting, early stopping is incorporated with a patience of 10 epochs and the restoration of the best weights. The training process spans over 50 epochs with a batch size of 1, allowing the model to iteratively learn and improve its understanding of the underlying temporal patterns. The LSTM model's capability to capture both short and long-range dependencies, combined with its intricate memory cell mechanisms, renders it a potent tool for forecasting CO₂ emissions trends. Here’s LSTM model summary (Table 3.).

Table 3

LSTM model summary
Layer (type)	Output Shape	Param #
LSTM	(None, 1, 264)	280,896
Dropout	(None, 1, 264)	0
LSTM_1	(None, 264)	558,624
Dense	(None, 32)	8,480
Dense_1	(None, 1)	33
Training Data	80%
Validation Data	20%
Loss Function	mean_squared_error
Optimizer	Adam
Epochs	50
Batch Size	1
Return Sequence	True
Stopping Criteria	Early stopping
Performance Metrics	R², RMSE, MAE

In this section, results for the evaluation of all the machine learning models are presented as well as with the future predicted values of CO₂ emission in Pakistan for the next 9 years from 2022 to 2030.

3.1 Artificial Neural Networks (ANN)

In the context of predicting CO₂ emissions, the implemented ANN model showcased promising performance across various evaluation metrics. To assess the accuracy of our model, predictions are made both on the test data and the training data. The mean absolute error (MAE) is calculated for both the training and test predictions. For the training data, the MAE was found to be 2.328, while for the test data, the MAE was 8.111. Furthermore, the root mean squared error (RMSE) is employed to measure the discrepancy between actual and predicted values. The RMSE for training predictions is 3.031, indicating a relatively low error in capturing training data patterns. Correspondingly, the RMSE for test predictions is 10.25, signifying its capability to generalize to unseen data. Additionally, the coefficient of determination (R²) is computed to gauge the goodness of fit between predicted and actual values. The R² value for training predictions is 0.993, illustrating the model's capacity to explain most of the variance in the training data. For the test predictions, the R² is 0.8614, indicating a reasonable degree of explanation for the variance in test data.

The training loss and validation loss are visualized over epochs, showcasing the convergence and generalization of the model during training (Fig. 5(A)). The loss curves exhibited a decreasing trend, indicating that the model was learning and adjusting effectively. For the testing dataset, the absolute difference between actual and test predicted values was visualized to understand the extent of deviation between them (Fig. 5 (B)). To visualize the model’s performance on training and testing data sets comprehensively, three dimensions plots are generated by using real and estimated values (Fig. 5(C&D)). Moreover, a future projection of CO₂ emissions is conducted using the trained ANN model. The predictions for the next 9 years are generated based on the model's learned patterns. The predicted values are then transformed to their original scale using the MinMaxScaler inverse transformation. The projected trend suggested an inclining pattern in CO₂ emissions from 2022 to 2030, with an estimated increase from 215.87 million tonnes in 2022 to 260.52 million tonnes in 2030. To visualize the model's performance comprehensively, the actual, predicted, and future CO₂ emissions are plotted against years. The plot (Fig. 5(E)) showcased the alignment between actual and predicted values for both training and test datasets, along with the projected values for the future. This demonstrated the ANN model's ability to capture underlying patterns and trends in CO₂ emissions data. The implemented ANN model exhibited noteworthy predictive capabilities, accurately estimating CO₂ emissions and showcasing its potential for future projections. The evaluation metrics and visualizations collectively emphasize the model's efficacy in capturing complex temporal patterns within the CO₂ emissions dataset.

3.2 Gated Recurrent Unit (GRU)

The GRU model, a specialized type of recurrent neural network, is employed to forecast CO₂ emissions with a focus on comprehending temporal dependencies. While evaluating GRU for training and test data, the mean absolute error (MAE) is calculated to measure the average magnitude of the errors between forecasted and actual values. The MAE for training set is 4.29, while the test MAE is 7.936. This signifies the model's ability to estimate CO₂ emissions with a reasonable degree of accuracy. To further check the predictive accuracy, the root mean squared error (RMSE) is computed for both training and test predictions. For training set, RMSE is 5.0895, while for the test, RMSE is 11.25. These values reflect the overall dispersion of prediction errors, and lower values suggest a better model fit. In addition to these metrics, the coefficient of determination (R²) is calculated to determine the proportion of the variance in the dependent variable that could be explained by the independent variables. The R² value for training predictions is 0.981, indicating a high degree of fit between forecasted and actual values in the training dataset. The R² value for test predictions is 0.8355, indicating a reasonable level of explanation for the variance in the test dataset.

Furthermore, the training loss and validation loss are plotted over epochs to observe the model's learning behavior. The loss curves (Fig. 6(a)) depict a declining trend, suggesting that the model is effectively learning from the data while maintaining good generalization. The plot (Fig. 6(b)) showcases the deviation between predicted and actual values over time for test dataset, providing insights into specific periods of model performance. To visualize the model’s performance on training and testing data sets comprehensively, three dimensions plots are generated by using real and estimated values (Fig. 6(c) and (d)). For future predictions, a projection is made for the next 9 years based on the GRU model's learning. The projected CO₂ emission values are transformed to their original scale using the MinMaxScaler inverse transformation. The projected trend indicates an exponential pattern in CO₂ emissions over the forecasted years. The overall performance of the GRU model was visualized through a comprehensive plot depicting actual, predicted, and future CO₂ emissions across different years. This plot (Fig. 6(e)) highlighted the model's effectiveness in capturing CO₂ emissions trends. The GRU model demonstrated its capability to predict CO₂ emissions with favorable results across various evaluation metrics. The analysis underscored the model's ability to understand temporal dependencies and offered insights into future CO₂ emission trends. The visualizations and metrics collectively emphasize the model's suitability for predicting CO₂ emissions with reasonable accuracy.

3.3 Long Short-Term Memory (LSTM)

The Long Short-Term Memory (LSTM) model, designed to capture long-range dependencies and temporal patterns, is employed to forecast CO₂ emissions. A rigorous evaluation of the LSTM model's performance is conducted, encompassing various metrics to assess its predictive capabilities. The mean absolute error (MAE) is calculated for both training and test data to quantify the average magnitude of errors between forecasted and actual values. The model exhibited a training MAE of 1.953 and a test MAE of 7.936, indicating its ability to closely approximate CO₂ emission values. To gauge the model's predictive accuracy comprehensively, the root mean squared error (RMSE) is calculated for both training and test predictions. The training RMSE is 2.606, and the test RMSE is 10.45, reflecting the overall accuracy of the model's predictions and its ability to capture variations in CO₂ emissions. For training predictions, the R² value is 0.995, signifying a strong alignment between predicted and actual values in the training dataset. The R² value for test predictions is 0.8586, indicating a substantial level of variance explanation in the test dataset.

Future predictions for next 9 years from 2022 to 2030 are forecasted from trained LSTM model. The projected CO₂ emission values showcased an increasing trend over the forecasted years. These predictions provide insights into the expected trajectory of CO₂ emissions in the future. Training and validation loss over increasing epochs are plotted (Fig. 7(a)). These decreasing curves show the model’s ability to learn efficiently over increasing epochs. Absolute difference plot (Fig. 7(b)) shows the error between actual test data and predicted test data through LSTM trained model. To visualize the model’s performance on training and testing data sets comprehensively, three dimensions plots are generated by using real and estimated values (Fig. 7(c) and (d)). The overall performance of the LSTM model is visualized through a plot (Fig. 7(e)) that displayed actual, predicted, and future CO₂ emissions across different years. This plot underscored the model's aptitude in forecasting CO₂ emissions and capturing temporal trends. Overall, the LSTM model showcases exceptional performance in predicting CO₂ emissions. Its ability to capture long-range dependencies within time series data is evident through consistently low MAE and RMSE values, along with high R² values. The future projections further emphasize the model's proficiency in estimating emissions trends.

3.4 Ensemble Approach

The Ensemble method serves as a robust approach for forecasting CO₂ emissions. In presented study, we integrated the predictions obtained from the Artificial Neural Network (ANN), Gated Recurrent Unit (GRU), and Long Short-Term Memory (LSTM) models to create an ensemble model for final output. This approach aims to use the strengths of individual models and mitigate potential weaknesses, resulting in a more accurate and reliable prediction. To begin, the predictions from each individual model are combined by averaging the predicted values. The aggregated predictions are then inversely transformed to their original scale using the previously applied scaling technique. Subsequently, the Mean Absolute Error (MAE) and Root Mean Squared Error (RMSE) are computed for the ensemble approach. On training dataset, the MAE for ensemble is calculated to be 2.3169, with an associated RMSE of 3.1172. While, on the test dataset, the MAE is of 7.876, and the RMSE is found to be 10.5043. A high R² value of 0.996 in the training predictions by ensemble approach denotes a strong agreement between the actual and predicted values in the training dataset. The test dataset exhibits a considerable degree of variance explanation, as indicated by the R² value of 0.869 for test predictions.

The predictive performance of the ensemble model is also depicted through the year-wise predictions for future CO₂ emissions. As part of this research’s investigation, we forecasted CO₂ emissions for the years 2022 to 2030 by using ensemble approach. Notable predictions include an anticipated CO₂ emission of approximately 216.75 million tonnes for the year 2022, gradually increasing to around 302.50 million tonnes by the year 2030. These forecasts provide insights into the expected trajectory of CO₂ emissions and contribute to a comprehensive understanding of future environmental trends. The graph showing the difference between forecasted and actual values of CO₂ emission is presented below (Fig. 9).

Table 4 provides a detailed summary of performance metrics, including Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), and R², obtained from each model as well as the ensemble method. The results indicate that the proposed ensemble approach enhances the effectiveness of individual models and address potential weaknesses, leading to more accurate and reliable predictions. The ensemble method achieves this by initially combining predictions from each individual model through the averaging of predicted values. Similarly, the errors are amalgamated by averaging the performance metric values of each individual model. This collaborative approach helps leverage the strengths of each model while minimizing the impact of their respective weaknesses, resulting in an improved predictive performance overall.

Table 4

Summary of performance metrics (MAE, RMSE, and R², obtained from each model as well as the ensemble method
Performance	Training data				Test Data
Metric	ANN	GRU	LSTM	Ensemble Approach	ANN	GRU	LSTM	Ensemble Approach
MAE	2.328	4.29	1.953	2.3169	8.111	7.936	7.941	7.876
RMSE	3.031	5.0895	2.606	3.1172	10.25	11.25	10.45	10.5043
R²	0.993	0.981	0.995	0.996	0.8614	0.8355	0.8586	0.867

The predictions from each model as well as from the ensemble method are compiled and presented in a comprehensive table below (Table. 5). The final results concluded in this research are presented in graphical form (Fig. 9), showing the actual and predicted CO₂ emissions, which can help stakeholders get insights visually.

Table 5

Future predicted values by all models
Year	ANN Predicted (million-tonnes)	GRU Predicted (million-tonnes)	LSTM Predicted (million-tonnes)	Ensemble Predicted (million tonnes)
2022	215.87	217.47	216.92	216.75
2023	221.38	225.90	224.03	223.72
2024	226.91	236.10	231.80	231.39
2025	232.47	248.63	240.35	239.89
2026	238.04	264.36	249.82	249.38
2027	243.63	284.61	260.39	260.08
2028	249.23	311.50	272.29	272.24
2029	254.86	348.63	285.81	286.23
2030	260.52	402.55	301.32	302.50

The main comparisons between the earlier researches and this suggested study are outlined in the table below (Table 6). From the table, it is clear that this research’s proposed work is unique and distinctive. DL models showed excellent results on testing dataset giving very low RMSE value highlighting credibility of the models.

Table 6

Comparison between previous studies and proposed study.
Reference	Context	Method	Dataset	Prediction Years	RMSE Value
Alam et al. [24]	Gulf Countries	ARIMA, ANN	55 (1960–2014)	2015–2025	-
Habadi et al. [32]	Middle East	ARIMA	20 (1996–2015)	2014–2015	-
Eko Cahyono et al. [33]	Indonesia	GAINS	10 (2009–2019)	1990–2050	-
Nassef et al. [26]	Saudi Arabia	FFNN, ANFIS, LSTM	67 (1936–2020)	2020–2030	15.42
Wang C et al. [34]	China	SVR, RF, ANN	144 (1985–2020)	-	5.98
Proposed Study	Pakistan	ANN, GRU, LSTM	76 (1946–2021)	2021–2030	10.50

In this study, an in-depth analysis was conducted to forecast CO₂ emissions in Pakistan using three distinct predictive models i.e. ANNs, GRU, and LSTM.
The primary objective was to assess the accuracy and reliability of these models in projecting CO₂ emissions over the next decade. Each model was developed and trained using historical data spanning several decades, enabling model to capture the intricate trends and patterns of CO₂ emissions in Pakistan. Each model’s effectiveness was evaluated through various performance metrics, including MAE, RMSE, and R². Additionally, comprehensive and clear visualizations were generated to gain deeper insights into each model’s predictive capabilities.
The ANN model achieved an MAE of 2.328, RMSE of 3.031 and R² of 0.993 on training set, while an MAE of 8.111, RMSE of 10.25 and R² of 0.8614 on test set. The GRU model exhibited an MAE of 4.29, RMSE of 5.0895, and R² of 0.981 on training set, while an MAE of 7.936, RMSE of 11.25 and R² of 0.8355 on test set. The LSTM model showcased impressive results with an MAE of 1.953, RMSE of 2.606, and R² of 0.995 on training set, while an MAE of 7.941, RMSE of 10.45 and R² of 0.8586 on testing set.
To further enhance final predictions, an ensemble approach was employed that combines the strengths of the ANN, GRU, and LSTM models. The ensemble model achieved an MAE of 2.3169 and RMSE of 3.1172, and R² of 0.996 on training set, while an MAE of 7.876 and RMSE of 10.504, and R2 of 0.869 on test se, confirming its efficacy in delivering reliable forecasts.
This approach offers a consolidated view of the anticipated CO₂ emissions, aiding policymakers, researchers, and stakeholders in making informed decisions to address environmental challenges effectively. Moreover in future, investigating the impact of external factors such as economic indicators and policy changes could offer a more comprehensive understanding of CO₂ emission dynamics.

The following abbreviations are used in this manuscript:

CO₂	Carbon Dioxide
AI	Artificial Intelligence
ANN	Artificial Neural Network
GRU	Gated Recurrent Unit
LSTM	Long Short-Term Memory
MAE	Mean Absolute Errors
RMSE	Root Mean Squared Errors
ARIMA	Autoregressive Integrated Moving Average
CRI	Climate Risk Index
MT	Million Tonnes
ML	Machine Learning
DL	Deep Learning
R²	Coefficient of determination
RF	Random Forest
SVR	Support Vector Regressor

Competing interests

The authors have no conflicts of interest to declare that are relevant to the content of this article.

Author Contribution

NHT and NS conceived the idea. TN and HFN developed the theory and the methodology. TN and NS performed the computational work, and RNS, NS and MHB verified the proposed methodology. All authors discussed the findings and contributed to the final manuscript.

Ethical and informed consent for data used

Not applicable.

Data Availability and Access

The historical carbon dioxide (CO₂) emissions data for Pakistan is sourced from the repository "Our World in Data" available at https://ourworldindata.org/co2/country/pakistan. The data that support the findings of this study are available from the corresponding author upon reasonable request.

B. Lin, M.Y. Raza, Analysis of energy related CO2 emissions in Pakistan, J Clean Prod. 219 (2019) 981–993. https://doi.org/10.1016/J.JCLEPRO.2019.02.112.
United Nations. Climate actions notes. 2021., (n.d.).
S. Aftab, A. Ahmed, A. Chandio, B. Korankye, A. Ali, W. Fang, Modeling the nexus between carbon emissions, energy consumption, and economic progress in Pakistan: Evidence from cointegration and causality analysis, Energy Reports. 7 (2021) 4642–4658. https://doi.org/10.1016/j.egyr.2021.07.020.
D. Eckstein, M.-L. Hutfils, M. Winges, Global climate risk index 2019, Who Suffers Most from Extreme Weather Events. 36 (2018).
Y. Zhu, F. Zhang, S. Jia, Embodied energy and carbon emissions analysis of geosynthetic reinforced soil structures, J Clean Prod. 370 (2022) 133510. https://doi.org/10.1016/J.JCLEPRO.2022.133510.
Y. Wu, Z. Yang, B. Lin, H. Liu, R. Wang, B. Zhou, J. Hao, Energy consumption and CO2 emission impacts of vehicle electrification in three developed regions of China, Energy Policy. 48 (2012) 537–550. https://doi.org/10.1016/J.ENPOL.2012.05.060.
Y. Zhang, J. Wang, F. Hu, Y. Wang, Comparison of evaluation standards for green building in China, Britain, United States, Renewable and Sustainable Energy Reviews. 68 (2017) 262–271. https://doi.org/10.1016/J.RSER.2016.09.139.
R. Wu, J. Zhang, Y. Bao, Q. Lai, S. Tong, Y. Song, Decomposing the Influencing Factors of Industrial Sector Carbon Dioxide Emissions in Inner Mongolia Based on the LMDI Method, Sustainability. 8 (2016). https://doi.org/10.3390/su8070661.
F. Ma, A. Sha, R. Lin, Y. Huang, C. Wang, Greenhouse Gas Emissions from Asphalt Pavement Construction: A Case Study in China, Int J Environ Res Public Health. 13 (2016). https://doi.org/10.3390/ijerph13030351.
A.V.; A.R.U.; d′Arge, R.C. Kneese, Economics and the Environment: A Materials Balance Approach; Routledge: London, UK, 2015, n.d.
C. Wang, F. Wang, X. Zhang, Y. Yang, Y. Su, Y. Ye, H. Zhang, Examining the driving factors of energy related carbon emissions using the extended STIRPAT model based on IPAT identity in Xinjiang, Renewable and Sustainable Energy Reviews. 67 (2017) 51–61. https://doi.org/10.1016/J.RSER.2016.09.006.
G. Liu, H. Yang, Y. Fu, C. Mao, P. Xu, J. Hong, R. Li, Cyber-physical system-based real-time monitoring and visualization of greenhouse gas emissions of prefabricated construction, J Clean Prod. 246 (2020) 119059. https://doi.org/10.1016/J.JCLEPRO.2019.119059.
Q. Liu, J. Hao, Regional Differences and Influencing Factors of Carbon Emission Efficiency in the Yangtze River Economic Belt, Sustainability. 14 (2022). https://doi.org/10.3390/su14084814.
M. Gong, S. Zhen, H. Liu, Research on the nonlinear dynamic relationship between FDI and CO2 emissions in the “One Belt, One Road” countries, Environmental Science and Pollution Research. 28 (2021) 27942–27953. https://doi.org/10.1007/s11356-021-12532-6.
R.P.; M.M.C.; M.E.F. Masini, Machine learning advances for time series forecasting. J. Econ. Surv. 2021, (n.d.).
J. Crespo Cuaresma, J. Hlouskova, S. Kossmeier, M. Obersteiner, Forecasting electricity spot-prices using linear univariate time-series models, Appl Energy. 77 (2004) 87–106. https://doi.org/10.1016/S0306-2619(03)00096-5.
S. Elsworth, S. Güttel, Time series forecasting using LSTM networks: A symbolic approach, ArXiv Preprint ArXiv:2003.05672. (2020).
L. Amarpuri, N. Yadav, G. Kumar, S. Agrawal, Prediction of CO2 emissions using deep learning hybrid approach: A Case Study in Indian Context, in: 2019 Twelfth International Conference on Contemporary Computing (IC3), 2019: pp. 1–6. https://doi.org/10.1109/IC3.2019.8844902.
S. Kumari, S.K. Singh, Machine learning-based time series models for effective CO2 emission prediction in India, Environmental Science and Pollution Research. (2022). https://doi.org/10.1007/s11356-022-21723-8.
Y. Qi, W. Peng, R. Yan, G. Rao, Use of BP Neural Networks to Determine China’s Regional CO₂ Emission Quota, Complexity. 2021 (2021) 6659302. https://doi.org/10.1155/2021/6659302.
A. Abdullahi, R.S.A. Usmani, T. Pillai, M. Marjani, I. Abaker, I. Hashem, An Optimized Artificial Neural Network Model using Genetic Algorithm for Prediction of Traffic Emission Concentrations, International Journal of Advanced Computer Science and Applications. Volume 12 (2021) 2021. https://doi.org/10.14569/IJACSA.2021.0120693.
M. Khan, M. Khan, Application of ANFIS, ANN and fuzzy time series models to CO 2 emission from the energy sector and global temperature increase, Int J Clim Chang Strateg Manag. 11 (2019). https://doi.org/10.1108/IJCCSM-01-2019-0001.
M. Mutascu, CO2 emissions in the USA: new insights based on ANN approach, Environmental Science and Pollution Research. 29 (2022) 68332–68356. https://doi.org/10.1007/s11356-022-20615-1.
T. Alam, A. AlArjani, A Comparative Study of CO2 Emission Forecasting in the Gulf Countries Using Autoregressive Integrated Moving Average, Artificial Neural Network, and Holt-Winters Exponential Smoothing Models, Advances in Meteorology. 2021 (2021) 1–9. https://doi.org/10.1155/2021/8322590.
Z. Zuo, H. Guo, J. Cheng, An LSTM-STRIPAT model analysis of China’s 2030 CO2 emissions peak, Carbon Manag. 11 (2020) 577–592. https://doi.org/10.1080/17583004.2020.1840869.
A.M. Nassef, A.G. Olabi, H. Rezk, M.A. Abdelkareem, Application of Artificial Intelligence to Predict CO2 Emissions: Critical Step towards Sustainable Environment, Sustainability. 15 (2023). https://doi.org/10.3390/su15097648.
Our World in Data, Pakistan: CO2 Country Profile, (n.d.). https://ourworldindata.org/co2/country/pakistan (accessed August 17, 2023).
M. Khashei, M. Bijari, An artificial neural network (p, d, q) model for timeseries forecasting, Expert Syst Appl. 37 (2010) 479–489. https://doi.org/10.1016/J.ESWA.2009.05.044.
F. and Z.J. and Y.G. Zhang Xu and Shen, Time Series Forecasting Using GRU Neural Network with Multi-lag After Decomposition, in: S. and L.Y. and Z.D. and E.-A.E.-S.M. Liu Derong and Xie (Ed.), Neural Information Processing, Springer International Publishing, Cham, 2017: pp. 523–532.
G. Shen, Q. Tan, H. Zhang, P. Zeng, J. Xu, Deep Learning with Gated Recurrent Unit Networks for Financial Sequence Predictions, Procedia Comput Sci. 131 (2018) 895–903. https://doi.org/10.1016/J.PROCS.2018.04.298.
B. Lindemann, T. Müller, H. Vietz, N. Jazdi, M. Weyrich, A survey on long short-term memory networks for time series prediction, Procedia CIRP. 99 (2021) 650–655. https://doi.org/10.1016/J.PROCIR.2021.03.088.
M. Habadi, C. Tsokos, Statistical Forecasting Models of Atmospheric Carbon Dioxide and Temperature in the Middle East, Journal of Geoscience and Environment Protection. 05 (2017) 11–21. https://doi.org/10.4236/gep.2017.510002.
W. Eko Cahyono, Parikesit, B. Joy, W. Setyawati, R. Mahdi, Projection of CO2 emissions in Indonesia, Mater Today Proc. 63 (2022) S438–S444. https://doi.org/10.1016/J.MATPR.2022.04.091.
C. Wang, M. Li, J. Yan, Forecasting carbon dioxide emissions: application of a novel two-stage procedure based on machine learning models, Journal of Water and Climate Change. 14 (2023) 477–493. https://doi.org/10.2166/wcc.2023.331.

No competing interests reported.

floatimage11.jpeg

Download PDF

Version 1

posted

You are reading this latest preprint version

Predictive analytics for environmental impact: Deep learning and AI ensemble strategies in forecasting CO2 emissions

Status:

Version 1

Abstract

Figures

1. Introduction

2. Methodology

2.1. Data Pre-Processing

2.2. Machine Learning Models

2.2.1. Artificial Neural Networks (ANN)

2.2.2. Gated Recurrent Unit (GRU)

2.2.3. Long Short-Term Memory (LSTM)

3. Results and Discussion

3.1 Artificial Neural Networks (ANN)

3.2 Gated Recurrent Unit (GRU)

3.3 Long Short-Term Memory (LSTM)

3.4 Ensemble Approach

Summary and conclusions

Abbreviations

Declarations

Competing interests

Author Contribution

Ethical and informed consent for data used

Data Availability and Access

References

Additional Declarations

Supplementary Files

Status:

Version 1