Short-term photovoltaic power forecasting with clear-sky normalization and physical decoding

Photovoltaic output exhibits pronounced nonstationary fluctuations across seasons and weather conditions, making accurate forecasting challenging. A physically structured forecasting framework integrating clear-sky references and intermediate-variable supervision is proposed to forecast mean power over the next 15 min interval. First, mutual information is used to select historical input features, and clear-sky irradiance and power are calculated. A shared time series mixer encoder extracts a historical representation, which is combined with target-time clear-sky and temporal features. Three prediction heads estimate the clear-sky plane-of-array irradiance index, ambient temperature, and effective performance factor. The predicted irradiance index is converted into absolute irradiance, module temperature is calculated using the nominal operating cell temperature relationship, and a physical decoder with fixed photovoltaic parameters produces the power forecast. A joint objective combines the clear-sky power index (KPV) loss with auxiliary losses for the three intermediate variables. The framework is evaluated using monitoring data from a 5.3 kW photovoltaic array at the Desert Knowledge Australia Solar Center in Australia. On the evaluated 2025 samples, the proposed method yields root mean squared error) and mean absolute error values 7.26% and 13.70% lower, respectively, than those of the KPV-based climatological persistence baseline.