Hybrid Approaches in Time Series Forecasting: Integrating Statistical and Machine Learning Models

In the face of accelerating climate change, precise forecasting of carbon dioxide (CO₂) concentration thresholds has become critical for global policy and planning. This study explores a data-driven approach to modelling future CO₂ trajectories using advanced time-series prediction architectures, including LSTM, GRU, CNN-LSTM, and DNN models. Based on historical emission records, the modelling framework simulates the evolution of global CO₂ levels and estimates 2047 as a potential overshoot point where atmospheric concentrations may surpass 500 ppm—a threshold associated with irreversible environmental damage. Comparative performance across neural architectures highlights the DNN configuration as yielding the most consistent long-range forecasting outcomes. Analysis further indicates that a sustained 6.37% annual emission reduction would be necessary to avert this trajectory, while a 23.38% reversal rate could return concentrations to safer levels. The forecasting framework offers actionable insights for sustainability planning and early-warning environmental monitoring.