An interpretable AI framework for clear-sky detection in photovoltaics monitoring

Special Issue on ‘EU PVSEC 2025: State of the Art and Developments in Photovoltaics', edited by Robert Kenny and Carlos del Cañizo

Open Access

Table 5

Mapping of SHAP-identified failure patterns to engineered feature solutions. New abbreviations are defined as follows: CV_ghi,30 is the coefficient of variation for GHI over a 30 min window; Δk_d,z is the deviation of the diffuse fraction from its zenith-specific median; and $M_{R_{d}}$ is the match score between the measured and theoretical direct-to-diffuse ratio.

Identified Failure Pattern	Key SHAP Indicators	Engineered Feature Solution
Cloud Enhancement (False Positive)	High and unstable values for k_t and R_d.	CV_ghi,30: Quantifies long-term stability to distinguish transient irradiance spikes from sustained clear conditions.
Hazy but Stable Conditions (False Negative)	Moderately low but stable values for k_t and k_b.	Δk_d,z: Contextualizes the diffuse component by zenith angle, helping the model learn the signature of haze versus clouds.
Dawn/Dusk Ambiguity (False Positive)	High θ_z with low and inconsistent R_d values.	$M_{R_{d}}$ : Compares the measured component balance to a theoretical model to better interpret attenuated, low-light conditions.

Current usage metrics show cumulative count of Article Views (full-text article views including HTML views, PDF and ePub downloads, according to the available data) and Abstracts Views on Vision4Press platform.

Data correspond to usage on the plateform after 2015. The current usage metrics is available 48-96 hours after online publication and is updated daily on week days.

Initial download of the metrics may take a while.