Intelligent techniques, harmonically coupled and SARIMA models in forecasting solar radiation data: A hybridization approach

Abstract

The unsteady and intermittent feature (mainly due to atmospheric mechanisms and diurnal cycles) of solar energy resource is often a stumbling block, due to its unpredictable nature, to receiving high-intensity levels of solar radiation at ground level. Hence, there has been a growing demand for accurate solar irradiance forecasts that properly explain the mixture of deterministic and stochastic characteristic (which may be linear or nonlinear) in which solar radiation presents itself on the earth’s surface. The seasonal autoregressive integrated moving average (SARIMA) models are popular for accurately modelling linearity, whilst the neural networks effectively capture the aspect of nonlinearity embedded in solar radiation data at ground level. This comparative study couples sinusoidal predictors at specified harmonic frequencies with SARIMA models, neural network autoregression (NNAR) models and the hybrid (SARIMA-NNAR) models to form the respective harmonically coupled models, namely, HCSARIMA models, HCNNAR models and HCSARIMA-NNAR models, with the sinusoidal predictor function, SARIMA, and NNAR parts capturing the deterministic, linear and nonlinear components, respectively. These models are used to forecast 10-minutely and 60-minutely averaged global horizontal irradiance data series obtained from the RVD Richtersveld solar radiometric station in the Northern Cape, South Africa. The forecasting accuracy of the three above-mentioned models is undertaken based on the relative mean square error, mean absolute error and mean absolute percentage error. The HCNNAR model and HCSARIMA-NNAR model gave more accurate forecasting results for 60-minutely and 10-minutely data, respectively. Highlights HCSARIMA models were outperformed by both HCNNAR models and HCSARIMA-NNAR models in the forecasting arena. HCNNAR models were most appropriate for forecasting larger time scales (i.e. 60-minutely). HCSARIMA-NNAR models were most appropriate for forecasting smaller time scales (i.e. 10-minutely). Models fitted on the January data series performed better than those fitted on the June data series.