BENCHMARKING STATISTICAL, MACHINE LEARNING, DEEP LEARNING, AND HYBRID FORECASTING MODELS FOR GLOBAL RENEWABLE ENERGY CONSUMPTION: A WALK-FORWARD CROSS-VALIDATION STUDY WITH STRUCTURAL BREAK ANALYSIS

Energy independence and resilience have become critical policy priorities as geopolitical tensions, supply disruptions, and price volatility expose the vulnerability of fossil-fuel-dependent energy systems. Accurate forecasting of renewable energy consumption is therefore essential for effective energy transition planning, infrastructure investment, and monitoring progress toward international climate targets such as Sustainable Development Goal 7 (SDG-7). In macro-energy policy practice, quantitative forecasts underpin scenario design, capacity planning, and assessment of alignment with net-zero pathways, yet the annual frequency and short length of globally comparable time series severely constrain the effective application of data-intensive forecasting methods. This study benchmarks 13 forecasting model families-spanning baselines (Naive, Random Walk with Drift, Linear Trend), classical statistical methods (ETS, Damped ETS, Theta, ARIMA), machine learning (XGBoost), deep learning (GRU, LSTM, N-BEATS), an additive model (Prophet), and a novel ETS-GRU hybrid - against the World Bank EG.FEC.RNEW.ZS indicator (1990-2020). All models are evaluated under a unified 5-window expanding walk-forward cross-validation protocol with a 3-year forecast horizon, nested hyperparameter tuning, multi-seed deep learning robustness checks, Diebold–Mariano tests, Model Confidence Set analysis, and bootstrap inference.