Hybrid Machine Learning and Generative AI Approach for Carbon Footprint Forecasting and Sustainable Policy Recommendations

Abstract

Mitigation of global climate change depends on accurate carbon emission forecasting and implementation of actionable, data-supported policy frameworks. This thesis proposes a unique hybrid architecture that blends state-of-the-art machine-learning techniques with Generative AI to pursue this purpose. Via an ambitious global sustainable energy dataset, we implement a multi- step approach to analyze country-level CO2 emissions' estimations and predictions. The authors first develop time-series forecasting models such as ARIMA, XGBoost, and Long Short-Term Memory (LSTM) networks to capture emissions trends for the period 2026-2035. The second step clusters countries into different sustainability archetypes using K-Means, making judgments based upon selected economic, energy, and emission indices. Lastly, the study harnesses the predictive insights and the profile information obtained from the clustering via the Large Language Model (LLM) to create context-driven and professionally articulated mitigation strategies. Results show that the non-linear models, particularly XGBoost and LSTM, exhibit much better performance than the classical statistical baselines in relating complex emission dynamics. The XGBoost model has shown high predictive accuracy of the R2 scored 0.998 and Mean Absolute Percentage Error of (MAPE) 1.38 while LSTM model performed fairly well with an R2 of 0.987 and MAPE of 3.50. Moreover, through the effective Generative AI application, quantitative forecasts were turned into qualitative policy propositions, which immensely suit high-impact sectors. Thus, this dissertation showcases how the hybrid AI systems may actually change the game by linking raw environmental data with strategic decision-making to provide a sturdy instrument for policymakers in achieving the global sustainable development goals.