An End-to-End Physics-Informed Dynamic Harmonic Mixer Framework for Efficient Day-Ahead Photovoltaic Power Forecasting

A critical deployment gap hinders the industrial adoption of physics-informed deep learning for day-ahead photovoltaic forecasting. Despite theoretical robustness, physics-integrated models face persistent deployment barriers from unstable joint training and high sensitivity to Numerical Weather Prediction (NWP) uncertainty. To address this, we propose PV-DynHarm-Mixer, an end-to-end framework that places the physical prior in the representation rather than at the input. Its three components include: (1) a VAE-based Adaptive Dynamic Harmonic Encoder (ADHE) that learns diurnal and sub-diurnal components from measured history and removes the need for external forecasts; (2) a lightweight multi-scale temporal MLP-Mixer; and (3) two adaptive training mechanisms, validation-guided regularization and a progressive cautious optimizer, that stabilize the joint learning. On a full test year at the Desert Gardens site of the DKASC Yulara Solar System, the model attains the lowest nRMSE (5.54%), MAE (31.06 kW), and R2 (0.875) of ten compared models, and is the only learning model with positive forecast skill over persistence at the native five-minute resolution (+3.7%). The harmonic representation is shown to carry this advantage: removing the encoder is the single ablation that pushes the model below persistence, the learned diurnal frequency remains locked at one cycle per day as the look-back window extends from one to three days while nRMSE improves monotonically to 5.18%, and a model trained at one site stays above persistence at a second installation after only one month of local data. One complete 24 h forecast costs 5.86 × 109 FLOPs and 44.59 MB of runtime memory, using only on-site measurements together with deterministic calendar and astronomical features. The advantage is concentrated in high-insolation regimes; on predominantly calm winter days, clear-sky persistence remains the stronger reference.