An interpretable metaheuristic-optimized RVFL architecture for robust CO₂ emissions prediction in decarbonizing energy systems

Abstract

Abstract Sustainable energy transition requires accurate forecasting tools to guide decarbonization and environmental policy. This study develops an artificial intelligence-based framework to predict CO₂ emissions in Germany using advanced machine learning techniques. The proposed model integrates a Random Vector Functional Link network with the Whale Optimization Algorithm to enhance learning efficiency, optimize hyperparameters, and improve predictive performance. The analysis uses monthly data from 1990 to 2021 and incorporates key drivers of environmental development, including renewable energy consumption, economic growth, financial development, globalization, and geopolitical risk. The optimization process significantly reduces the need for manual parameter tuning and strengthens model robustness. The results show strong contemporaneous prediction (nowcasting) accuracy, with an out-of-sample R² of 0.9978, outperforming conventional machine learning benchmarks. However, under expanding-window validation, the proposed model, together with every benchmark, exhibits substantial performance deterioration across structural breaks, which is a limitation of the current study. To ensure transparency and application relevance, SHAP-based interpretability analysis evaluates the contribution of each variable. The findings reveal that renewable energy expansion plays the most significant role in reducing CO₂ emissions, confirming its central importance in Germany’s decarbonization strategy. Economic and financial factors also influence emissions dynamics, highlighting the interconnected nature of energy and economic systems. The study demonstrates how artificial intelligence can enhance data driven decision making in sustainable development. It provides a useful nowcasting predictive and interpretability tool for energy planning and environmental management under stationary conditions, while underscoring the need for regime-aware methods when the data-generating process shifts.