Predicting climate change: A comparative analysis of time series models for CO2 concentrations and temperature anomalies

Abstract

This study presents a novel, integrated modeling framework that combines machine learning (ML) techniques with physics-based approaches to forecast both CO₂ emissions and global temperature anomalies. Unlike prior research that typically addresses these components in isolation, this work concurrently applies and compares five advanced ML models—Long Short-Term Memory (LSTM), Extreme Gradient Boosting (XGBoost), Convolutional Neural Network (CNN), Facebook Prophet, and a hybrid CNN-LSTM—alongside two physics-based models: a zero-dimensional Energy Balance Model (EBM) and a simplified General Circulation Model (GCM) adapted from NASA's GISS framework.Using monthly global datasets from January 2000 to April 2024, obtained from the National Oceanic and Atmospheric Administration (NOAA) and the Scripps Institution of Oceanography, the models are evaluated based on predictive accuracy (RMSE, MSE, MAE, R²), scalability, and interpretability. Prophet demonstrated the highest accuracy for CO₂ emission forecasting (RMSE = 0.035), while LSTM achieved the best performance in temperature anomaly prediction (RMSE = 0.086). Physics-based models provided interpretable and computationally efficient long-term projections but lacked short-term flexibility.To facilitate reproducibility and practical application, we developed ClimateChange-ML, an open-source software package that implements all proposed models, includes trained weights, and provides full documentation and visualization tools.The novelty of this work lies in its dual-modeling strategy and comprehensive comparative evaluation, highlighting the complementary strengths of data-driven and physically grounded methods. This integrated approach offers a more holistic framework for climate forecasting across multiple temporal scales, providing valuable insights for both scientific understanding and climate policy planning.