A hybrid approach to forecasting solar irradiance using ARIMA-Neural Networks Models

Abstract

If solar power is to be integrated into the national grid in large volumes, it should be backed up by accurate short-term solar irradiance forecasting information. This is because system de signers in the solar power markets utilise short-term forecasting information in the early stages of setting up solar power systems to properly design and size the solar power harvesting system such as the photovoltaic (PV) system. However, the unsteady and varying nature (mainly due to atmospheric mechanisms and diurnal cycles) of solar energy resource is often a hindrance to receiving high-intensity levels of solar radiation at the ground level. As such, there has been a growing demand for accurate solar irradiance predictions that adequately captures the mixture of linear and nonlinear behaviour in which solar radiation presents itself on the earth’s surface. Among the time series based forecasting techniques, autoregressive integrated moving average (ARIMA) models have been widely applied in forecasting because of their ability to handle lin ear component embedded in the time series data. On the other hand, artificial neural networks (ANNs) are capable of handling the nonlinear feature in the time series data that cannot be prop erly captured by traditional linear models (e.g. ARIMA models). In this study, we build hybrid models by blending seasonal autoregressive integrated moving average (SARIMA) models (to cap ture linear component) and neural network autoregression (NNAR) models (to capture nonlinear component) to form SARIMA-NNAR models which we utilise to model global horizontal solar irradiance (GHI) data collected from RVD-GIZ solar radiometric station located in the Alexandra Bay, Northern Cape, South Africa. Overall, comparative results with four GHI data series show that the SARIMA-NNAR model is superior over the NNAR model and SARIMA model in terms of forecasting performance. A brief exploration of the harmonically coupled neural network au toregression (HCNNAR) models revealed that these models are capable of effectively modelling the inherent periodic sinusoidal component in the solar irradiance data observed on the earth’s surface with some level of accuracy. Hence, the study proposes the use of these models for future studies on modelling and forecasting solar irradiance data.