This study presents a comprehensive machine learning approach to model final energy consumption patterns across approximately 150 countries using Natural Gradient Boosting (NGBoost). Unlike traditional econometric methods that rely on point estimates, this approach provides probabilistic forecasts with uncertainty quantification based on historical data from 1960 to 2022. The model incorporates macroeconomic indicators (GDP and population), energy prices, and climate variables (average temperature) to predict annual final energy consumption. Results demonstrate strong predictive accuracy across diverse geographical and temporal contexts, with actual consumption consistently falling within 99% confidence intervals. Beyond quantifying statistical uncertainty, the probabilistic framework enables the systematic exploration of deep uncertainty through structured “what-if” scenario analyses, allowing policymakers to test alternative assumptions about energy transitions and policy effectiveness. By employing an uncertainty-aware prediction methodology, the framework provides not only accurate forecasts but also robust confidence intervals, enabling scenario analysis and stress-testing for energy transition planning. This methodological advance provides policymakers and analysts with a tool that integrates data-driven insights with explicit uncertainty quantification, thereby enhancing support for long-term energy strategies.
