Predicting solar radiation at high resolutions: A comparison of time series forecasts

The increasing use of solar power as a source of electricity has led to increased interest in forecasting radiation over short time horizons. The relevant horizons for generation and transmission can range from as little as 5 minutes to as long as several hours. Forecasting experiments are run using six data sets, at resolutions of 5, 15, 30, and 60 min, using the global horizontal component. The data exhibits nonlinear variability, due to variations in weather and cloud cover. Nevertheless, the dominance of the 24-h cycle makes it straightforward to build predictive models. Forecasting tests are run using regressions in logs, Autoregressive Integrated Moving Average (ARIMA), and Unobserved Components models. Transfer functions, neural networks, and hybrid models are also evaluated. All the tests use true out-of-sample forecasts: The models are estimated over history prior to the start of the forecast horizon, the data is forecasted, and the predicted values are compared with the actuals. In nearly all the tests, the best results are obtained using the ARIMA in logs, with time-varying coefficients. There are some exceptions. At high resolutions, a transfer function using cloud cover is found to improve over the ARIMA. In a few cases, the neural net or hybrid models can improve at very high resolutions, on the order of 5 min. The success of the ARIMA is attributable mainly to its ability to capture the diurnal cycle more effectively than other methods. 2008 Elsevier Ltd. All rights reserved.