Optimization of photovoltaic power output predictions: a comparative analysis of artificial neural network algorithms with varied hidden layers

Abstract

This paper explores the impact of hidden layers in the prediction of the photovoltaic power output of a polycrystalline Photovoltaic module using an artificial neural network with four different algorithms: Levenberg-Marquardt optimization, Bayesian regression, scaled conjugate gradient and scaled resilient propagation. Error function estimations were designed in MATLAB software and trained between 1–20 hidden layers’ configurations. Solar irradiance, ambient temperature, module temperature, wind speed and relative humidity are the five inputs to the artificial neural network model. The predictability of the power output using the four algorithms at a varied number of hidden layers was investigated using a time series seasonal average data set of 631 points obtained under outdoor conditions. 70% of the data set was used for training while validation and testing used 15% each. Results showed that all algorithms exhibited commendable prediction performance across all algorithms, with average mean square errors of 0.03042, 0.02679, 0.078 and 0.0709 for Levenberg-Marquardt optimization, Bayesian regression, scaled conjugate gradient and scaled resilient propagation, respectively. High coefficients of determination further confirmed accuracy, with values of 0.98669 for Levenberg-Marquardt optimization, 0.98996 for Bayesian regression, 0.96046 for scaled conjugate gradient and 0.96541 for scaled resilient propagation. The Bayesian regression algorithm outperformed other algorithms. These findings offer valuable insights for enhancing PV system performance and promoting sustainable energy solutions dyes, D3 showed the best properties compared to D1 and D2 for dye-sensitized solar cell applications.