Power prediction of soiled PV module with neural networks using hybrid data clustering and division techniques. (August 2016)

• Record Type: Journal Article
• Authors: Pulipaka, Subrahmanyam; Kumar, Rajneesh
• Is Part Of: Solar energy. Volume 133(2016)
• Page Start: 485
• Page End: 500
• Publication Date: 2016-08
• Subjects: Clustering -- Data division -- Neural network -- Power prediction; Solar energy -- Periodicals; Solar engines -- Periodicals; 621.47
• DOI: 10.1016/j.solener.2016.04.004
• Languages: English
• ISSNs: 0038-092X

Abstract

Graphical abstract: Highlights: A hybrid data clustering using K -means and FCM is proposed. A new data division technique for neural network training is developed. Artificial soiling experiments are conducted to obtain the data for developing the model. The soil on the panel is characterized based on particle size composition, XRD and FTIR. The neural network model developed accurately predicts the output of a soiled panel. Abstract: The performance of a neural network model depends on the data used to develop the model and nature of data preprocessing used in its training algorithm. This paper proposes a hybrid clustering algorithm to preprocess the data that is used for neural network training as well as data division technique to predict the power output of a soiled solar panel. Deterministic characteristics of soil on the panel namely particle size composition, X-ray diffraction (XRD) analysis and Fourier transform infrared spectroscopy (FTIR) analysis are used to model the soil. Artificial soiling experiment is conducted with each soil sample on the panel at irradiance level in the range of 200–1200 W/m 2 for a set of 18 tilt angles to collect data of short circuit current ( Isc ) and open circuit voltage ( Voc ) leading to power output. NNR (Neural Network random) with random data division, NNF (Neural network fuzzy) with Fuzzy C Means clustering before training, NNK (Neural network k -means) with K -means data clustering and NNH (Neural network hybrid) with hybrid data clustering and division techniques are developed using these data. The performance of these networks for a known and unknown soil samples is compared. For a known soil sample, NNH performed better (Maximum percentage error of 2%) as compared to NNR (−6.3%), NNF (−6.7%) and NNK (13%). In case of unknown soil sample NNH outperforms other models with a maximum error margin as low as (−10%) as compared to NNR (50%), NNF (−81%) and NNK (124%).