Global forecasting of atmospheric CO <sub>2</sub> concentrations using a hybrid STL-Prophet-LSTM model
Zhenzhong Liu, Yulin Cai, Shili Meng, Zizheng Zhu, Xianglei Meng, Xinglu Wang, Lin Sun
Abstract
The increasing concentration of atmospheric carbon dioxide (CO2) poses a significant global challenge, underscoring the need for accurate predictions to better understand and mitigate climate change. While existing CO2 forecasting models are often restricted to specific scenarios or localized regions, a comprehensive global prediction framework remains lacking. To address this gap, we propose the SPL (STL-Prophet-LSTM) model, an integrated approach combining Seasonal-Trend decomposition using Loess (STL), the Prophet forecasting model, and Long Short-Term Memory (LSTM) networks. Leveraging monthly average CO2 concentration data from six globally distributed monitoring stations (NOAA), we forecasted CO2 trends over the next decade. Our results demonstrate the SPL model’s superior predictive performance, achieving an average RMSE of 0.67, MAE of 0.53, and R2 of 0.99—outperforming benchmark models (ARIMA, SARIMA, Prophet, and standalone LSTM). Projections reveal a concerning upward trajectory, with Northern Hemisphere CO2 levels expected to reach 450 ppm by 2032, compared to 438 ppm in the Southern Hemisphere, highlighting persistent hemispheric disparities. This study provides a robust methodological framework for global-scale CO2 concentration forecasting, offering critical insights for climate policy and mitigation strategies.