Horizon-adaptive time–frequency fusion for renewable power forecasting: empirical evidence of a consistent horizon–representation relationship across wind and solar systems

Accurate renewable power forecasting requires extracting information from both temporal dynamics and frequency-domain fluctuations, yet the optimal balance between these representations across forecasting horizons remains unclear. This study proposes a Horizon-Adaptive Time–Frequency Fusion (HATF) framework that learns horizon-dependent weighting between temporal and spectral features through a lightweight fusion module. Experiments on seven years of operational wind farm SCADA data demonstrate that HATF consistently outperforms recurrent, convolutional, and Transformer-based baselines across forecasting horizons from 15 min to 4 ℎ . Cross-dataset validation on an independent wind turbine dataset and a solar photovoltaic dataset further confirms the robustness of the approach; for the solar dataset, the auxiliary input channels of the model carry irradiance and ambient temperature rather than wind speed and wind direction, preserving the three-channel architecture across energy sources. Ablation studies show that adaptive fusion is essential for effectively integrating time-domain and frequency-domain information. A key finding is that the learned fusion coefficient exhibits a highly consistent approximately linear relationship with forecasting horizon across all datasets ( 𝑅 ⁢ ² > 0 . 9 8 ). Autocorrelation analysis indicates that this behavior is linked to the substantially different persistence timescales of low-frequency and high-frequency signal components, causing the model to gradually shift from frequency-dominant representations at short horizons to time-dominant representations at longer horizons. In addition, a Gaussian prediction head provides well-calibrated uncertainty estimates with near-nominal 95% prediction interval coverage. The results suggest that horizon-adaptive time–frequency fusion offers both an effective forecasting framework and an interpretable perspective on how predictive information evolves across temporal scales in renewable energy systems.