A physics-informed TA-GMM-HMM framework for probabilistic weather regime clustering and photovoltaic power prediction

As solar photovoltaics (PV) penetration increases in modern power grids, accurate short-term power forecasting becomes critical for effective planning and real-time operations. Current forecasting methods face two limitations, reliance on hard clustering of weather regimes, which discards transitional states such as partly cloudy conditions and neglects the temporal dynamics essential for PV generation modeling, and omission of solar geometry parameters (azimuth, zenith, clear-sky irradiance), which are widely used in irradiance forecasting but overlooked in PV power prediction despite their direct physical relationship to panel output. To address these gaps, this study proposes a physics-informed forecasting framework that integrates solar geometry features with a novel Time-Aware Gaussian-Mixture Hidden-Markov Model (TA-GMM-HMM). TA-GMM-HMM produces soft regime assignments that preserve transitional states, with weighted time encodings and sticky transitions that enforce temporal coherence and minimize flicker. These probabilistic regime memberships, combined with solar-physics features and adaptive signal processing techniques, drive a hybrid Automatic Relevance Determination–Extra Trees forecasting model. Empirical validation on three rooftop PV sites in Lisbon demonstrates error reductions of 27–78% (RMSE) compared to prior work, confirming the effectiveness of this probabilistic, physics-informed framework for operational PV forecasting in grid integration contexts.