Daily Solar Radiation Forecasting based on a Hybrid NARX-GRU Network in Dumaguete, Philippines

Abstract

In recent years, solar radiation forecasting has become highly important worldwide as solar energy increases its contribution to electricity grids. However, due to the intermittent nature of solar radiation caused by meteorological parameters, forecasting errors arise, and fluctuations in the power output of photovoltaic (PV) systems become a severe issue. This paper aims to introduce a forecasting hybrid model of daily global solar radiation time series. Meteorological data and solar radiation samples from Dumaguete, Philippines, are used to assess the forecasting accuracy of the proposed nonlinear autoregressive network with exogenous inputs (NARX) – gated recurrent unit (GRU) hybrid model. Four different models were trained using the meteorological and solar radiation data, which are the Optimizable Gaussian Process Regression (GPR), Nonlinear Autoregressive Network (NAR), NARX, and the proposed Hybrid NARX-GRU Network. Results show that the hybrid NARX-GRU model has a root mean square error (RMSE) of ~0.05 and a training time of 33 seconds. The proposed hybrid model has better forecasting performance compared to the three models which obtained RMSE values of 27.741, 39.82, and 28.92, for the GPR, NAR, and NARX, respectively. The simulation results demonstrate that the NARX-GRU model significantly outperforms the regression and single models in terms of statistical metrics and training efficiency. Furthermore, this study shows that the hybridized NARX-GRU model is able to provide an effective estimation for daily global solar radiation, which is important in the operation of PV plants in the country, specifically for unit commitment purposes

Keywords

• forecasting
• narx-gru
• neural network
• photovoltaic system
• solar radiation