Lead‐Time‐Conditioned Convex Stacked Ensemble for Medium‐Term Electricity Load Forecasting

ABSTRACT

Increasing renewable penetration raises the value of accurate electricity‐load forecasting. Medium‐term load forecasting (MTLF) is complicated by lead‐time‐dependent error propagation, evolving seasonal effects and cross‐market heterogeneity. This paper evaluates nine base learners and proposes a lead‐time‐conditioned convex (LTCC) stacked ensemble to address their lead‐time‐dependent performance variability. LTCC combines a static, out‐of‐fold‐fitted convex prior per horizon with an online exponential‐weights update that re‐normalises the blend from realised past losses alone, keeping weights non‐negative and unit‐sum with no learning rate to tune. The framework is assessed with full‐year rolling weekly forecasts on Spain, France, Greece and Portugal, against the base learners and an eight‐method ensemble ablation ladder. Across 12 country–metric comparisons, LTCC is competitive with, but does not uniformly dominate, the strongest specialised comparator, and none of these differences is statistically significant. Against the base‐learner mean, LTCC improvement is positive in every case. By rank, LTCC attains the lowest mean rank of 18 methods and the smallest shortfall to the per‐cell best. An ablation shows the OOF‐fitted prior and online update are not interchangeable. These results support cross‐market consistency, rather than per‐cell dominance, as the main advantage, and a parameter‐free online adaptation rule as a defensible alternative to a fixed blending rate.