Long-term hourly load forecasting (LTLF) is essential for strategic power system planning, yet improvements are often pursued through increasing model complexity rather than enhancing structural representation. This study demonstrates that carefully designed feature engineering—explicitly incorporating calendar decomposition, special-day identification, and climatic-year substitution—substantially improves forecasting accuracy across five European countries. By restructuring the input representation of annual demand into normalized hourly profiles driven by calendar and climatic factors, the proposed framework achieves an average MAPE reduction of approximately 25% relative to baseline formulations, consistently across all case studies. Multiple machine learning models are evaluated (MLR, GRNN, ANN, GBT, LSTM, CNN, SVR, DNN), with GRNN providing the best overall trade-off between accuracy and robustness (average MAPE of 2.77% for the test year). A climatic substitution analysis further shows that inter-annual weather variability induces an intrinsic dispersion that effectively defines a practical performance ceiling for deterministic LTLF models. The results indicate that structured feature representation exerts a stronger influence on performance than incremental increases in algorithmic complexity. The proposed framework offers an interpretable and computationally efficient approach for generating long-term hourly load scenarios under climatic uncertainty.
Paulos et al. (Mon,) studied this question.