Generation of Horizontal Hourly Global Solar Radiation From Exogenous Variables Using an Artificial Neural Network in Fes (Morocco)

Abstract

This paper presents a comparison between multilayer perceptron (MLP) and neural autoregressive with exogenous inputs(NARX) for generating hourly global solar radiation in the city of Fes(Morocco). The results from this analysis are essential for the analytic modelling of the long term performance of solar energy systems. MLP and NARX were created, trained and tested using MATLAB. Four hourly measured variables over five years and one calculated parameter were used as input to the models and horizontal hourly global solar radiation as target. Models with different numbers of neurons in the hidden layer as well as different combinations of inputs were experimented with. The regression coefficient (R 2 )and two error statistics, namely, normalized root mean square error (nRMSE) and normalized mean bias error (nMBE) are used to evaluate different models. According to theses statistics, the best model is NARX with 5 inputs. The generalization of this model over unseen data and its ability to produce accurate forecasts show a good accuracy ( nRMSE = 15%, nMBE=0.036% and R 2 =0.95). For the wide use of the proposed model with available data, different sizes for different periods of data were used for the learning process. Depending on the accuracy required for the generated values, our model gives quite good results with relatively small sets of training data. As such, the proposed model shows a good ability to generate hourly solar radiation values from more available and cheaper data namely temperature and relative humidity.