Physically constrained neural network for solar spectral irradiance reconstruction

Accurate surface solar spectral irradiance data are indispensable for spectrally selective building envelopes and advanced photovoltaic systems, yet remain critically scarce due to the high cost of spectroradiometer networks and the stringent atmospheric parameter requirements of conventional radiative transfer models. This study proposes a Physically Constrained Neural Network (PCNN) that reconstructs full-band (280–2500 nm) solar spectral irradiance using readily‑available meteorological observations. The core innovation lies in embedding non‑negativity and energy closure as soft constraints within the loss function, guiding the optimization toward physically plausible solutions. Benchmarked against the standard NREL measurement dataset, the PCNN achieves substantially lower prediction errors than unconstrained models, with SHAP analysis confirming that the constraints induce physically grounded feature selection dominated by broadband irradiance. Field validation in Chengdu, China, further demonstrates that the PCNN achieves superior spectral reconstruction accuracy (R 2 = 0.9686 over 280–1000 nm) compared to SMARTS2 when the latter lacks its requisite atmospheric inputs. The proposed framework offers an accessible pathway for high‑resolution spectral irradiance assessment in regions lacking dedicated spectral monitoring infrastructure.