A Nonlinear Autoregressive Exogenous (NARX) Neural Network Model for the Prediction of the Daily Direct Solar Radiation

Zina Boussaada, Octavian Curea, Ahmed Remaci, Haritza Camblong, Najiba Mrabet Bellaaj
2018 Energies  
The solar photovoltaic (PV) energy has an important place among the renewable energy sources. Therefore, several researchers have been interested by its modelling and its prediction, in order to improve the management of the electrical systems which include PV arrays. Among the existing techniques, artificial neural networks have proved their performance in the prediction of the solar radiation. However, the existing neural network models don't satisfy the requirements of certain specific
more » ... ions such as the one analyzed in this paper. The aim of this research work is to supply, with electricity, a race sailboat using exclusively renewable sources. The developed solution predicts the direct solar radiation on a horizontal surface. For that, a Nonlinear Autoregressive Exogenous (NARX) neural network is used. All the specific conditions of the sailboat operation are taken into account. The results show that the best prediction performance is obtained when the training phase of the neural network is performed periodically. Autoregressive Exogenous (NARX) model. In [6] , the authors use only the historical solar radiation as input of the forecaster, so the wavelet transformation is used to split the historical data into adapted series for the forecast. The authors apply the ARMA concept as a linear forecaster and the NARX model as a nonlinear form recognition instrument, in order to decrease the errors of the wavelet-ARMA forecast. In [2] also, only the historical data of the solar radiation is used and decomposed. This decomposition is performed via Local Mean Decomposition (LMD) and Empirical Mode Decomposition (EMD). The result is a set of nonlinear sub-components, so the authors of [2] combine two nonlinear prediction models. They use Least Squares Support Vector Machine (LSSVM) and Volterra models to predict the high-frequency subcomponents and low-frequency subcomponents, respectively. The final result is obtained by superimposing the sub-prediction results. In [7] four models of machine learning are compared: Random Forests (RF), neural network based on Multi-Layer Perceptron (MLP), Linear Regression (LR) and k-Nearest Neighbors (kNN). The objective is to demonstrate the efficiency of nonlinear models to forecast two times series: the hourly solar radiation on horizontal surface and the solar irradiance obtained on PV panels at different tilted surfaces. The forecast of the solar radiation is performed using three types of data inputs: ground weather data, satellite remote sensing data and calculations of sun position parameters (declination, hour, zenith, elevation and azimuth angles). References [3] [4] [5] 8] forecast the daily solar radiation using ANNs. In [3], its authors use the correlation criteria to determine the endogenous and the exogenous inputs to take into account. Four endogenous time lags are taken for the clear sky model of solar radiation. From several meteorological parameters only three are selected as exogenous inputs: relative humidity, sunshine duration and cloud cover. Based on the nature of the obtained data, the authors of [4] divide data differently: predicted data and statistical data, so they combine: 1/several statistical data as the calculated cloud ratio, the maximum hourly variation of the solar radiation, the solar radiation absolute daily variation between the day (t) and the day (t − 1), etc., and 2/numerical weather prediction data which contain the one day ahead forecast of the cumulative radiation each 3 h. As a result, a hybrid dataset is used as neural network input. In [5] , the authors don't classify the input data, but rather they train four neural networks with four combinations of input features, with the aim of considering the influence of several meteorological parameters on the prediction results. They conclude that the best performance is obtained when using the following inputs: the year day, the mean daily solar radiation at the top of atmosphere, the maximum sunshine hours, the mean daily air temperature, the mean daily relative humidity and the wind speed. In some other studies, the neural network technique is combined with other nonlinear models. For example, in [8] , the authors propose a diagonal recurrent wavelet neural network, in order to estimate the hourly solar radiation of the next day. It is a model founded on the combination of wavelet neural network with recurrent ANN and fuzzy technology, where the wavelet basis is implemented as activation function for the neurons and the input vector of the neural network contains defuzzificated data of the nebulosity. The results of these researches are advanced and give much better performances than the conventional prediction models. However, they cannot be generalized. They don't satisfy the requirements of certain specific situations. Among these situations, that of this research, whose main goal is the prediction of the direct solar radiation on a horizontal surface, in order to predict the solar radiation on a sloped surface, then the quantity of obtainable power from PV arrays in a race sailboat. Contrary to the aforementioned research studies, the proposed solution in this paper takes the change of localization into consideration and allows to take into account the other specific conditions (tilted surface, shadow . . . ) of the sailboat operation that will be considered in our future studies.
doi:10.3390/en11030620 fatcat:u2lhhywznfbmdn7kmgwciif6s4