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

Abstract

Accurate clear-sky detection (CSD) is essential for reliable data analysis and performance assessments in photovoltaic (PV) systems. However, many advanced machine learning (ML) models function as “black boxes”, limiting their interpretability and trustworthiness. This study presents an interpretable Artificial Intelligence (AI) framework that combines high predictive performance with deep insight into model decision-making. Using a hand-labelled dataset from a fixed-tilt PV system in Golden, Colorado, USA, with 1min plane-of-array (POA) measurements of global horizontal irradiance (GHI), direct normal irradiance (DNI), and diffuse horizontal irradiance (DHI), a Categorical Boosting (CatBoost) classifier is developed for CSD. The model is iteratively refined through a closed-loop diagnostic process guided by SHapley Additive exPlanations (SHAP). Misclassified instances are analysed using dimensionality reduction via Uniform Manifold Approximation and Projection (UMAP) and clustering, revealing distinct, physically-grounded failure modes such as “cloud enhancement”, where reflected or scattered sunlight temporarily increases irradiance, and “hazy but stable conditions”, where thin atmospheric haze slightly attenuates sunlight without introducing variability. Insights from this analysis inform targeted feature engineering, yielding a refined model with high classification performance quantified by an F1-score of 97.3%, along with substantially reduced false positive (1.99%) and false negative (7.0%) rates, reflecting both overall accuracy and balanced sensitivity to clear-sky and non-clear-sky periods. This interpretable framework improves the reliability of clear-sky filtering for downstream PV applications, including fault detection and diagnosis (FDD) and long-term performance loss rate (PLR) estimation, and provides a transferable methodology for developing trustworthy AI models in energy systems.