Short‑term photovoltaic power forecasting must model both the regular solar cycle and weather‑driven fluctuations, whose relative importance varies with the forecast horizon. This paper proposes a hierarchical ensemble that fuses temporal neural networks, historical analogs, climatology baselines, and gradient‑boosted trees. Solar geometry and numerical weather predictions describe the expected generation conditions, while horizon‑specific convex weights combine the complementary outputs. A subsequent calibration step uses available historical forecast errors to correct recent bias. The framework is evaluated on the public PVDAQ dataset (15–240 minute horizons) and on three GEFCom2014 solar zones at hourly horizons up to four hours. On PVDAQ the ensemble attains a daylight capacity‑normalized mean absolute error of 4.315%, a 4.11% reduction versus a full‑feature LightGBM and a 6.03% reduction versus a fine‑tuned Chronos‑2 under identical calibration. Expert‑removal experiments reveal redundancy within the ensemble. Across three training seeds on GEFCom2014, learned fusion outperforms equal weighting but matches LightGBM performance. The results support horizon‑specific fusion as an effective strategy, while its advantage depends on the dataset and evaluation period.
$$\text{MAE}=\frac{1}{N}\sum_{i=1}^{N}|y_i-\hat{y}_i|$$
Review