Anomaly detection in all-sky images: An approach using robust ensemble modeling of cloud cover fraction and prediction bounds

作     者：da Rocha, Vinicius Roggerio Fisch, Gilberto Costa, Rodrigo Santos Ruano, Antonio 

作者机构：Natl Inst Space Res Ave Astronautas 1758 BR-12227010 Sao Jose Dos Campos SP Brazil Univ Algarve Fac Sci & Technol P-8005139 Faro Portugal Univ Lisbon Inst Super Tecn IDMEC P-1049001 Lisbon Portugal 

出 版 物：《ENGINEERING APPLICATIONS OF ARTIFICIAL INTELLIGENCE》 (Eng Appl Artif Intell)

年 卷 期：2025年第143卷

核心收录：

学科分类：0808[工学-电气工程] 08[工学] 0811[工学-控制科学与工程] 0812[工学-计算机科学与技术（可授工学、理学学位）] 

基　　金：Coordenacao de Aperfeicoamento de Pessoal de Nivel Superior - Brasil (CAPES) [001, 88887.601400/2021-00] CNPq [409711/2021-7] INCT-Mudancas Climaticas Fase 2 (FAPESP) [2014/50848-9] INCT-Mudancas Climaticas Fase 2 (CNPq ) [465501/2014-1] INCT-Mudancas Climaticas Fase 2 (CAPES/FAPS) [16/2014] Fundacao para a Ciencia e a Tecnologia (FCT) 88887.885659/2023-00 

主　　题：Multi-Objective Genetic Algorithms Radial Basis Function Neural Networks Convex hull All-sky images Cloud cover fraction Prediction bounds 

摘      要：All-sky images (ASI) are widely used for sky monitoring, particularly in solar energy generation applications. Alignment issues and interferences demand a detection process of problematic images. With high sampling frequencies (1-2 images per minute), automating this process is crucial for managing large datasets and enabling integration into automatic systems, which is the objective of this work. For this purpose, a robust ensemble model, using the ApproxHull and Radial Basis Function (RBF) neural networks combined with Multi Objective Genetic Algorithms (MOGA) tools, was developed to compute the cloud cover fraction of each image. By computing the deviation between this result and the one obtained by the equipment, and by assessing if it lies within prediction bounds obtained in the design phase, an automatic method for detecting anomalies in All-sky images was obtained. ASI data collected during the Green Ocean (GoAmazon) Experiment 2014/5 was employed. The proposed approach obtained a Probability of Interval Coverage (PICP) similar to the user- specified level of confidence for several sets within what was classified as a gooddataset, while being able to detect anomalies found within a baddataset.