Clustering analysis for classification and forecasting of solar irradiance in Durban, South Africa.
Loading...
Date
2017
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Classification and forecasting of solar irradiance patterns has become increasingly important for
operating and managing grid-connected solar power plants. A powerful approach for classification
of irradiance patterns is by clustering of daily profiles, where a profile is defined as irradiance
as a function of time. Classification is useful for forecasting because if the class of a day can be
successfully forecast, then the irradiance profile of that day will share the general pattern of the class.
In Durban, South Africa (29.871 °S; 30.977 °E), beam and diffuse irradiance profiles were recorded
over a one-year period and normalized to a clear sky model to reduce the effect of seasonality, from
which several variables were derived, namely minute-resolution beam, hourly-resolution beam and
diffuse, and hourly-resolution beam variability. To these variables, individually and in combination,
k-means clustering was applied, and beam irradiance was found to be the one that best distinguishes
between sky conditions. In particular, clustering of hourly-resolution beam irradiance produced four
classes with diurnal patterns characterized as sunny all day, cloudy all day, sunny morning-cloudy
afternoon, and cloudy morning-sunny afternoon. These classes were then used to forecast beam
and diffuse irradiance for the day ahead, in association with cloud cover forecasts from Numerical
Weather Prediction (NWP) output. Two forecasting methods were investigated. The first used k-means
clustering on predicted daily cloud cover percentage profiles from the NWP, which was a
novel aspect of this research. The second used a rule whereby predicted cloud cover profiles were
classified according to whether their averages in the morning and afternoon were above or below
50%. From both methods, four classes were obtained that had diurnal patterns associated with the
irradiance classes, and these were used to forecast the irradiance class for the day ahead. The two
methods had a comparable success rate of about 65%. In addition, hour-ahead forecasts of beam
and diffuse irradiance were performed by using the mean profile of the forecast irradiance class
to extrapolate from the current measured value to the next hour. The method showed an average
improvement of about 22% for beam and diffuse irradiance over persistence forecasts. These results
suggest that classification of predicted cloud cover and irradiance profiles are potentially useful for
development of class-specific, multi-hour irradiance forecast models.
Description
Doctor of Philosophy. University of KwaZulu-Natal, Durban. 2017.