Forecasting solar irradiance at short horizons: Frequency and time domain models

Abstract

A key issue in integrating solar power into the grid is short-term forecasting. Up to now, most solar forecasting has been run in the time domain. But since the data are dominated by the diurnal cycle, frequency domain methods naturally recommend themselves. Several time series models are tested over horizons of 15, 30 and 45 min, and 1–3 h. The methods include persistence, regressions in levels, ARIMAs, and a frequency domain model, based on Fourier transformation of the ARIMA. The model coefficients are time-varying, and are estimated using moving windows. Models are estimated on the actual irradiance data, the clear sky and clearness indexes. Forecasting tests are run using data from six sites in the continental United States. The persistence and regression forecasts achieve similar degrees of accuracy at the 15 min horizon, with the frequency domain model showing marginally higher errors. At the 45 min horizon, the performance of the persistence, regression and frequency domain forecasts is so close as to be nearly indistinguishable. At the 1 h horizon, the ARIMA as applied to the clear sky index achieves the most accurate forecast. At the 2 h horizon, the frequency domain model predicts more accurately. At 3 h, the ARIMA and frequency domain models are close. None of the models consistently dominate the others. Model performance differs as a function of the resolution of the data and the forecast horizon. The accuracy of the forecasts also depends on the degree of time variation in the coefficients. The optimal window width for estimating the coefficients was found to be 480 h.