Towards a Sustainable Future: Accurate Predictions of CO2Emissions Using Hybrid LSTM-GRU Models

Abstract

Growing carbon dioxide (CO₂) emissions are the primary driver of global climate change, creating an urgent need for robust and reliable predictive models to support effective mitigation strategies. Accurate long-term emission forecasting is essential for policymakers to design evidence-based interventions aligned with international climate commitments. This study proposes a hybrid layered deep learning framework for long-term CO₂ emission prediction up to 2030, integrating Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) architectures to capture complex temporal dependencies and nonlinear emission patterns. Historical emissions data were obtained from the Our World in Data database and applied to the seven largest CO₂-emitting countries, representing diverse economic structures and development trajectories. Multiple hybrid configurations were evaluated, including LSTM–LSTM, GRU–GRU, LSTM–GRU, and GRU–LSTM models, and their performance was assessed using Root Mean Square Error (RMSE). The results demonstrate that hybrid architecture consistently outperforms conventional single-layer models. The GRU– LSTM model achieved the lowest RMSE for China (520.81×10⁶ tons) and Japan (24.15×10⁶ tons), while the GRU–GRU model performed best for India (148.78×10⁶ tons). The LSTM–GRU configuration yielded superior results for Russia (73.03×10⁶ tons), and the LSTM–LSTM model showed the highest accuracy for the United States (263.14×10⁶ tons). These variations highlight the importance of country-specific modeling strategies when forecasting emissions. Overall, the results confirm the effectiveness of hybrid deep learning models in capturing long-term dynamics in CO₂ emissions. The proposed framework provides reliable, data-driven forecasts that support strategic climate planning and emission-reduction efforts aligned with the Paris Agreement.