AI Meets Energy: Forecasting The Future Of Country-Level Energy Consumption

Abstract

Accurate forecasting of national energy consumption is critical for enabling effective energy planning, sustainable development, and informed policy formulation. This study investigates and compares a diverse set of forecasting models—including deep learning architectures (LSTM, CNN-BiLSTM, Transformer), machine learning ensemble methods (XGBoost, LightGBM, CatBoost), and classical time series approaches (SARIMA)—to predict annual oil, gas, and renewable energy consumption across three major economies: the United States, China, and India. Leveraging a harmonized multi-decadal dataset, extensive preprocessing techniques were employed to ensure temporal consistency, normalize consumption metrics, and enhance feature representation. Each model was trained using consistent time-series cross-validation and evaluated using a standard suite of performance metrics. The research aims to assess the strengths and limitations of each modeling paradigm in the context of national-level energy forecasting, and to provide a foundation for data-driven model selection strategies in energy systems analytics.