Bayesian Regression-Based Power Loss Probabilistic Estimation of Dust-Accumulated PV Systems

• Published in: Chinese Control and Decision Conference pp. 3495 - 3501
• Main Authors: Li, Chunrong, Chen, Xinyi, Shen, Yu, Zhang, Kanjian, Fang, Shixiong, Wei, Haikun
• Format: Conference Proceeding
• Language: English
• Published: IEEE 16.05.2025
• Subjects: Accuracy; Adaptation models; Bayes methods; Bayesian regression neural network; Computational modeling; dust; Estimation; Feature extraction; feature pyramid network; Neural networks; photovoltaic power loss; power loss estimation; Predictive models; Systems operation; Uncertainty; uncertainty quantification
• ISSN: 1948-9447
• DOI: 10.1109/CCDC65474.2025.11090541

This paper proposes a Bayesian regression neural network (SolarBRNN) combined with a feature pyramid network (FPN) for photovoltaic power loss (PVPL) prediction. The model effectively captures multi-scale features from images by incorporating the feature pyramid module, enhancing its ability to detect the distribution of dust and its impact on power loss. Through Bayesian regression, the model can not only predict the mean power loss but also quantify the uncertainty of the predictions, providing more reliable information for system operation. Experimental results demonstrate that, compared to quantile regression methods, the SolarBRNN model significantly improves performance in terms of prediction standard deviation and the confidence interval coverage probability (CICP). Notably, the CICP indicator increases by 12%, indicating better prediction accuracy and stronger adaptability.