Do Classical Methods Still Win? Revisiting Forecasting Strategies for Curtailment Mitigation in Brazil

Background: The increase in curtailment in Brazil highlights structural limitations in the electrical system and reinforces the need for more accurate demand forecasts to support dispatch, storage, and expansion decisions. Although deep learning and hybrid models have advanced rapidly, classical statistical methods remain widely used. This study compares statistical, machine-learning, and hybrid forecasting models using monthly Brazilian electricity consumption data (2004–2023) to assess which techniques offer the best accuracy and efficiency for curtailment-oriented planning. Results: Six models were evaluated: SARIMAX, Holt–Winters, LSTM, XGBoost, ES-RNN, and N-BEATS. Performance was assessed using sMAPE and computational metrics. Holt–Winters achieved the lowest error (2.03 ± 0.06%) with the fastest inference and smallest storage. ES-RNN provided the second-best accuracy (2.18 ± 0.06%) but required significantly more computation. SARIMAX also performed well (2.3 ± 0.1%) with fast inference but high storage demand. LSTM and N-BEATS showed moderate accuracy, with LSTM having the highest computational cost. XGBoost was the least accurate (6.4 ± 0.2%) despite low latency. Conclusions: Classical methods, particularly Holt–Winters, outperformed more complex AI-based models for this dataset, highlighting that methodological sophistication does not always guarantee superior results. Accurate, low-cost forecasting can aid curtailment mitigation and support more reliable renewable integration.