Abstract
Reliable 6–24 h regional wind-power forecasts support reserve scheduling, balancing, dispatch, and market operation, but evaluation is difficult when regional telemetry is unavailable and reanalysis is mistaken for an operational weather forecast. This study develops a reproducible proxy-forecasting framework for Dobrich and Stara Zagora, Bulgaria, combining physics-derived features, user-specified 70–80% effective-output scaling scenarios, a Temporal Transformer, and evolutionary Fixed-Share Hedge. The data comprise 400 ERA5 NetCDF files for 2015–2024, 58,560 regional hourly rows, 196 capacity-register records, and 960 sparse national-generation observations. The chronological protocol uses 2015–2021 for training, 2022 for validation, 2023 for ensemble calibration, and 2024 for testing. In the strict historical-information track, Hedge attains a mean capacity-normalized mean absolute error (cNMAE) of 7.575% and a worst district–horizon value of 12.614% against the model-derived regional proxy. Against the same proxy, a non-operational oracle experiment using target-time ERA5 reanalysis yields 0.272–0.427%. Under the oracle information set, the standalone Transformer has a worst cNMAE of 3.033%, while four-expert Hedge has a worst cNMAE of 0.430% by assigning virtually all weight to histogram gradient boosting. These oracle values are proxy-recovery diagnostics, not regional operational-generation errors. The results provide a leakage-controlled benchmark that explicitly separates strict and oracle information regimes.

This publication has 16 references indexed in Scilit: