Hybrid multi-task deep learning framework for joint forecasting of solar and wind power

Abstract

This study presents the adaptive three-expert ensemble (A3E), a reproducible framework for joint one-hour-ahead solar and wind power forecasting. A3E combines temporal, physically informed, and high-generation Extra Trees experts through a causal local-error gate and is evaluated under a strictly chronological, leakage-free protocol. After timestamp reconciliation and causal feature construction, the primary source yielded 8,700 model-ready observations: 5,775 for training, 1,463 for validation, 719 for internal testing, and 743 for a same-source December holdout. The primary scale-neutral comparison used Macro train-SD normalized root mean square error (NRMSE) on 672 common forecast origins. A3E achieved 0.1978, compared with 0.2314 for the official photovoltaic (PV)-Client 2026 comparator and 0.2340 for TimeMixer. However, validation-optimized static weights achieved 0.2017 on the complete 719-origin internal test, compared with 0.2025 for adaptive A3E; the difference was not statistically significant under dependence-aware analysis. On the independent CISO and ERCO systems, TimeMixer outperformed A3E, demonstrating source-dependent model rankings. Event evaluation used 98 CISO and 320 ERCO high-generation events. The results support the value of heterogeneous expert integration, but not a universal accuracy advantage of adaptive weighting.