Nonparametric Conditional Heteroscedastic Hourly Probabilistic Forecasting of Solar Radiation

Abstract

We develop a new probabilistic forecasting method for global horizontal irradiation (GHI) by extending our previous bootstrap method to a case of an exponentially decaying heteroscedastic model for tracking dynamics in solar radiance. Our previous method catered for the global systematic variation in variance of solar radiation, whereas our new method also caters for the local variation in variance. We test the performance of our new probabilistic forecasting method against our old probabilistic forecasting method at three locations: Adelaide, Darwin, and Mildura. These locations are chosen to represent three distinct climates. The prediction interval coverage probability, prediction interval normalized averaged width and Winkler score results from our new probabilistic forecasting method are encouraging. Our new method performs better than our previous method at Adelaide and Mildura; regions with a higher proportion of clear-sky days, whereas our previous method performs better than our new method at Darwin; a region with a lower proportion of clear-sky days. These results suggest that the ideal probabilistic forecasting method might be climate specific.

Probabilistic Forecasting of Wind and Solar Farm Output

“…The usual procedure then is to square the z t and construct a forecast for the variance using either an ARCH or GARCH model, whichever is appropriate. However, as found in [6], an exponential smoothing forecast model is an efficient alternative. For the estimate of the variance at time t, S t , we have…”

“…Significantly, the simpler version of Fourier series plus ARMA performed at least as well over both continental and island locations. There have also been various articles on probabilistic forecasting of solar radiation, using a variety of methods (see [4][5][6][7] as examples).…”

“…For example, the distribution has a higher variance and skewness in the middle of the day in summer than at any other time. In [6], the authors exploit both this systematic change in variance, as well as the conditional change in variance to construct the prediction intervals. The conditional change in variance is typical of a series displaying the ARCH effect, the autoregressive conditional heteroscedastic or variance clustering effect.…”

“…In [10], ARMA and generalised autoregressive conditional heteroscedastic (GARCH) models are used to construct prediction intervals for solar radiation. They do not take into account the systematic change in variance that is evident in [5,6]. There are several other articles we could refer to here but we would like to simply point to [11], where the authors use a deep convolutional neural net for their forecasting.…”

A short-term solar radiation forecasting system for the Iberian Peninsula. Part 1: Models description and performance assessment

“…Arbizu-Barrena et al (2017) presented a similar hybrid approach (CIADCast) which also advects and diffuses the MSG CI images within the WRF model. Other authors have proposed probabilistic approaches to improve short-term solar radiation forecasting (Boland and Grantham, 2018).…”