A Si-photodiode radiometer based on artificial neural networks for daily instantaneous global horizontal solar radiation measurements

Abstract

A low cost Si-photodiode based radiometer that uses a transimpedance amplifier was designed, developed and used to measure daily instantaneous global solar radiation on a horizontal platform. The design together with a reference CMP11 pyranometer were connected to a data logger and daily measurements recorded at 10 s intervals for a period spanning six months. The measurement period extended from June to November 2017. The photodiode suffered from nonlinearities and spectral mismatches but it was corrected using an artificial neural networks (ANNs). Daily voltage data measured over a period of six months recorded at Mafikeng were used to create, train and test ANN that uses the Levenberg-Marquardt (LM) algorithm, and the sigmoid as the activation function, with photodiode voltages as inputs and global solar radiation measured by the reference CMP11 as targets. Two separate ANNs for summer and winter were developed, validated and tested using the data recorded on clear days in both winter and summer. ANN estimated irradiances were further corrected via a method that involved the design of an average difference correction function for that period. Daily deviation functions based on the simple difference between ANN estimated and CMP11 measured irradiances were obtained. The average deviation function for each season was estimated using insolation obtained on clear days in that season and it was used to adjust radiation of that whole season. An average difference correction function (dcf) was then calculated for that season by simply negating the average deviation function. This average difference correction function for each season was saved on the computer and applied to correct all season ANN estimated irradiances. For winter, this function was found to resemble the negated second derivative of the sigmoid function. Radiometer performance analysis shows average oot-mean square errors (rmse) of the measured irradiance dropping from ∼ 16 W/m² to ∼ 9 W/m² in summer and from ∼ 27 W/m² to ∼ 7 W/m² in winter. This shows that a reliable, low cost radiometer was successfully designed that could be used for renewable energy projects currently running on the Mafkeng campus. Overally ANN and signal processing techniques have been successfully applied to develop a reliable radiometer that could be used for long term measurements in place of the ex-pensive standard reference. The novelity of the radiometer and unique correction method designed shows that one can use any photodiode even if its spectral response does not match the reference. In addition, there may even be no need to worry about the exact horizontal setting of the photodiode. In most cases, empirical methods and models have been developed and used to calculate and predict global solar radiation for a particular location. This involves using data previously collected from meteorological stations. In such cases, only results concerning mismatches have been reported. This may be the first time where ANN estimated irradiances are themselves corrected to improve the accuracy of the ANN estimations.