Global climate change driven by carbon dioxide (CO2) emissions necessitates accurate forecasting for effective mitigation strategies. Traditional forecasting models, such as ARIMA and statistical methods, often struggle to handle the non-linear and long-term temporal dependencies in CO2 emission data. This paper evaluates the performance of advanced Recurrent Neural Network (RNN) architectures, specifically Long Short-Term Memory (LSTM), Stacked LSTM, Bidirectional LSTM (Bi-LSTM), and Gated Recurrent Unit (GRU) with attention mechanisms, in forecasting long-term CO2 emissions. These models are compared against traditional machine learning algorithms like Random Forest and Gradient Boosted Trees (GBT). The Random Forest and Gradient Boosted Trees, implemented using PySpark framework for distributed computing, are leveraged for efficient big data handling, enabling scalable solutions for CO2 emission forecasting. Our experimental results, using metrics such as Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), and Mean Absolute Percentage Error (MAPE), show that the Stacked LSTM model achieved the lowest MAE of 25.795, RMSE of 80.678, and MAPE of 1.072, outperforming Random Forest (MAE: 72.4, RMSE: 310.5, MAPE: 1.32) and GBT (MAE: 75.6, RMSE: 356.8, MAPE: 1.57). The findings demonstrate that advanced RNN models, particularly the Stacked LSTM and Bi-LSTM with attention, provide superior performance over traditional models, effectively capturing complex temporal patterns in the data. This study offers valuable insights into CO2 emission forecasting approaches, contributing to the development of accurate tools for climate policy and strategy formulation.

Climate change and global warming demand immediate global action to protect future generations. Although deep learning is crucial for solving complex tasks, its high energy consumption and carbon emissions from extensive training processes necessitate improvements in efficiency to reduce its environmental impact. This research addresses the need for accurate carbon emission tracking to mitigate climate change and promote sustainable development. We evaluated deep learning models—specifically, the Temporal Fusion Transformer (TFT), Attention-based LSTM (Att-LSTM), and standard LSTM—in predicting CO2 emissions using datasets from World Bank Indicators (WBI) and Our World in Data (OWID). The results highlight TFT’s superiority, achieving the lowest mean squared error (MSE) in validation (0.0029) compared to Att-LSTM (0.0070) and LSTM (0.0072). Additionally, TFT demonstrates a lower root mean squared error (RMSE) in validation (0.0539) than both Att-LSTM (0.0712) and LSTM (0.0909), showcasing its enhanced predictive accuracy across evaluation tools like CarbonTracker, eco2AI, and CodeCarbon. Furthermore, TFT consistently outperforms LSTM in computational efficiency and carbon emissions, despite LSTM’s lower power usage but higher emissions in specific scenarios. These findings underscore TFT's pivotal role in formulating effective climate policies and promoting renewable energy sources, emphasizing its potential to significantly contribute to global warming mitigation and sustainability.

Climate change and global warming are urgent environmental issues demanding immediate action to safeguard future generations. The major contributor to the greenhouse effect, carbon dioxide $\left(\mathrm{CO}_{2}\right)$, primarily originates from industrial and transportation fossil fuel combustion. International agreements, like the Paris Agreement, call for a 30-35% reduction in CO2emissions compared to 2005 levels. This research aims to predict CO2emissions and raise awareness among SAARC nations and governments about the increasing trend. We introduce a novel predictive framework using Attention-based Long Short-Term Memory (A-LSTM) for CO2emissions analysis. The Attention mechanism assigns variable weights to input data, facilitating indirect connections between LSTM outputs and pertinent inputs. This enhances resource allocation in the A-LSTM model, overcoming computational constraints. We integrate input parameters encompassing CO2emissions from land-use changes, oil, natural gas, and coal combustion to forecast CO2emissions and correlate them with population and per capita GDP. Our comparative analysis conclusively demonstrates the superior performance of A-LSTM models over baseline LSTM models when applied to the CO2emission dataset sourced from Our World in Data (OWID) and World Bank Indicator database. Specifically, the LSTM model registers a MAPE of 24.968 and an RMSE of 0.34, whereas the Attention-based LSTM model showcases a marked improvement of 57% with a considerably lower MAPE of 10.5902 and an RMSE of 0.107.