The system is learning.

Both PJM and ERCO grids show remarkably similar model performance hierarchies: trend models achieve 5.27% and 6.71% MAPE respectively, while weather and momentum models struggle in the 31-34% range. This 5-6x performance gap is driving the ensemble adaptation in predictable directions, though with notable regional variations.

The adaptive weighting systems are responding differently to the same performance signals. PJM has concentrated 84% weight on the trend model with minimal weather/momentum allocation (8.2%/7.9%), while ERCO maintains more balanced weights (40%/35%/25%). This suggests ERCO's system may be accounting for factors beyond pure historical accuracy, or the trend model's dominance is less stable there.

The worst errors cluster around similar timestamps and magnitudes (-48% to -51%) across both grids, with all major failures occurring during overnight hours (00:xx to 01:xx). The timing pattern suggests systematic blind spots in handling overnight load transitions, particularly during March-April periods when seasonal patterns may be shifting.

Open question: Why does ERCO maintain higher weather model weighting (35%) despite its poor individual performance (34.11% MAPE) compared to PJM's near-abandonment (8.2%) of a similarly performing weather model (31.79% MAPE)?

The trend model demonstrates remarkably superior individual performance across both grids (5.27% MAPE for PJM, 6.71% for ERCO) compared to weather and momentum models, which both hover around 31-34% MAPE. This performance gap drives the adaptive weighting algorithm to allocate 84% weight to trend in PJM and 40% in ERCO, suggesting the trend model captures fundamental load patterns that weather and momentum models miss.

Both grids exhibit strikingly similar error patterns: all worst-case errors occur during late-night/early-morning hours (predominantly 00:xx and 01:xx timestamps) with consistent underestimation around -49% to -51%. The temporal clustering suggests systematic challenges in capturing overnight load dynamics, possibly related to industrial scheduling changes, demand response programs, or baseload transitions that occur during low-demand periods.

Weight allocation differs significantly between grids despite similar individual model performance rankings. ERCO maintains more balanced weights (40/35/25) while PJM heavily concentrates on trend (84/8/8). This divergence may reflect different grid characteristics - PJM's larger scale and complexity potentially making trend patterns more reliable, while ERCO's smaller system benefits from incorporating weather and momentum signals.

The ensemble achieves nearly identical MAPE (19.88% vs 19.75%) despite different weighting strategies, raising questions about whether the adaptive algorithm has found different local optima or whether grid-specific factors necessitate distinct ensemble compositions for equivalent performance.