Physics-Guided Deep Residual Learning for High-Accuracy Reconstruction of PV Plant AC Power from DC String Electrical Measurements

Engineering notes

Engineering notes will be added by the Power for Solar editorial team.

Chinese explanation / 中文解读

中文解读待补充：本站会优先为光伏效率、钙钛矿太阳能电池、储能技术、太阳能热利用、BIPV、并网技术等高价值论文补充中文说明。

Original abstract

Reconstructing photovoltaic plant AC active power accurately from DC-side electrical measurements is useful for digital twin calibration, sensor failure tolerance, and SCADA validation. By combining a physically motivated DC to AC estimate based on aggregated array DC power with a lightweight deep residual model that learns systematic deviations attributable to inverter non-idealities, operating regimes, and environmental conditions, this work proposes a physics-guided deep residual learning framework that reconstructs inverter AC power. The model makes use of DC string measurements, internal temperature, and meteorological variables from an hourly dataset covering 2024/05/01 to 2025/07/31 with 6778 samples. With much better daytime performance as 8.10% of sMAPE, the suggested reconstruction achieves MAE, RMSE, sMAPE, and R2 as 0.1907 kW, 0.2704 kW, 34.73%, and 0.9999, respectively. Residual diagnostics versus inverter temperature demonstrate low bias across operating ranges and reveal unpredictable error growth at high temperatures, consistent with known thermal derating effects in inverters. The results show that physics-guided residual learning can provide high accuracy AC power reconstruction from DC string measurements suitable for operational monitoring and digital twin applications.

Links and sources

• PDF from original source

Comments