Guided learning for photovoltaic power regression in the absence of key information

Accurate PV power regression is essential for grid stability and efficient energy use. Although deep learning has advanced forecasting performance, many models depend on solar-irradiance measurements that are often unavailable in practice. We address this limitation with a guided-learning framework that jointly estimates irradiance and regresses PV power, enabling accurate operation without irradiance at inference time and broadening applicability across locations. Using manually collected PV power from three sites and co-located meteorological data, we perform correlation analysis to select input variables and evaluate multiple network architectures under three training strategies and differing installed capacities. Results demonstrate that deep regression models generally outperform classical baselines but are sensitive to irradiance quality; in contrast, the proposed method remains robust when irradiance is missing or unreliable. Importantly, the framework is architecture-independent and can be integrated into diverse deep networks with minimal changes. Comprehensive statistical tests confirm significant accuracy gains over baselines that does not use irradiance. At the test site, the proposed method reduces hourly RMSE by 0.6 kW and daily RMSE by 1.07 kW relative to the baseline. Overall, the framework delivers accurate and deployable PV power predictions without requiring irradiance at run time and scales across locations and capacities. • Analyze PV power links to meteorological variables to build an optimal input set. • Propose guided deep learning for PV power regression without irradiance sensors. • Validate gains via t-tests and bootstrap; reduce hourly and daily RMSE vs baselines.