Monitoring Shading-Related Power Deficits in Photovoltaic Arrays Using an Adaptive Neural Reference

Shading reduces photovoltaic power output, while changes in irradiance and temperature can produce substantial variations during normal operation. Distinguishing shading-related output deficits from these normal variations requires a reliable estimate of expected array power. This study develops an adaptive normal-power reference using measurements from a neighboring photovoltaic array and environmental sensors. A control-anchored neural model learns the normal inter-array power relation and applies an environmental correction; departures of measured target power from this reference generate shading alarms calibrated on separate normal dates. Retrospective evaluation uses measured alternating-current power and recorded shading interventions at the La Réunion rooftop plant. Across three runs, environmental adaptation reduces normal-reference mean absolute error by 6.68–7.41 W in the low-control-power stratum relative to the control-only branch. On 193,864 commonly eligible observations, false-positive rates are 7.430–7.723% and false-negative rates are 6.471–6.524%, improving both rates over direct neural networks with similar trainable parameter counts. Detection performance is close to strong boosted-tree references, with 99.763% observation coverage. Daily power traces illustrate how the reference distinguishes shading-associated departures from shared output fluctuations and reveal the remaining alarm errors. The results support normal-power reference estimation as a practical basis for photovoltaic shading monitoring, with particular reference-accuracy benefits at low control power.