• Tidak ada hasil yang ditemukan

3.2 Forecast horizon

3.2.3 Day-ahead forecasts

For industry-related purposes, day-ahead forecasts are required for operational planning, program- ming backup, maintenance and transmission scheduling, as well as for planning of reserve usage and peak load matching. From 6 hours up to several days ahead solar irradiance forecasts rely mainly on NWP models.

For the present work, day-ahead forecasts were the focus and particularly how the combination of clustering of irradiance from ground-based instruments and cloud cover forecasts from the NWP could be used to produce a day-ahead forecast of irradiance. The classes produced by clustering de- scribe the diurnal irradiance patterns that can be correlated with diurnal cloud cover patterns. Cloud cover forecasts from the NWP is easily accessible through AccuWeather, thus meeting the need for high spatial and temporal resolution satellite images that are expensive and difficult to obtain.

Furthermore, AccuWeather provides a day-ahead forecast that is appropriate for combination with clustering, since clustering of irradiance profiles finds patterns that is on the diurnal-scale. Consid- ering this, the present study focused on day-ahead forecasts of irradiance for Durban.

An evaluation of three NWP forecasts i.e the North American Model (NAM), Global Forecast System (GFS) and European Centre for Medium-Range Weather Forecasts (ECMWF) for predict- ing global irradiance was conducted by Mathiesen and Kleissl (2011). For all models, MBE and RMSE exceeded 30 and 110 W/m2, respectively. A general under-prediction in cloud cover was also observed.

Using the National Digital Forecast Database (NDFD), Perez et al. (2007) derived surface global irradiance from the sky cover product to produce a forecasting model. Results showed that for 8-26 and 26-76 hour forecast horizons, the relative RMSE for hourly-averaged global irradiance was 38 and 40%, respectively.

Remund et al. (2008) evaluated three NWP-based GHI forecasts (ECMWF, NDFD and GFS/WRF) in the USA, reporting relative RMSE values ranging from 20% to 40% for the day ahead forecast horizon. Similar results were reported by Perez et al. (2010), where NWP-based irradiance forecasts in several places in the USA were evaluated.

Currently, NWP models are unable to predict the position and extent of cloud fields precisely.

This is primarily due to their relatively coarse spatial resolution (1-20 km), rendering them inef- ficient at resolving micro-scale physics associated with cloud formation. However, NWP models have the advantage of producing forecasts over long time horizons (15-240 hours) and have been shown to be more accurate than satellite-based models for time horizons exceeding 4 hours (Inman et al., 2013). Mesoscale NWP models (such as MM5 and WRF) have higher spatial and temporal resolution and may provide better accuracy in resolving cloud. However, even mesoscale models may not be able to capture cloud movement on the short time scales required at solar power plants, since their output is generally hourly.

Prediction of solar irradiance for more than one day ahead is not restricted to NWP-based mod- els, but can also include statistical techniques. An example of such a study that was undertaken by Martin et al. (2010). The authors presented a comparison of linear and non-linear statistical models applied to half daily values of global solar irradiance with a temporal horizon of 3 days. It was found that the neural network model yielded the best results. Other studies using similar techniques include Paoli et al. (2010, 2014). Techniques using ANNs and other artificial intelligence methods for modelling and forecasting of solar irradiance are presented by Mellit (2008).

It is evident from this discussion that certain forecasting techniques are successful within certain time horizons. Therefore, the choice of forecasting model is strongly dependent on two factors: (i) the forecast horizon and (ii) the available data at a particular site. Ground-based imagers have a maximum forecast horizon of approximately 30 minutes. Statistical methods have been success- fully applied to forecast solar irradiance for time horizons ranging from several minutes to a few hours ahead. Satellite imaging is most accurate in producing forecasts up to 6 hours ahead. Fore- casts beyond the 6 hour time horizon and up to several days ahead are most accurate if derived from

NWP models.

The focus of this research was the use of clustering for classification and forecasting of irra- diance for the day-ahead. Clustering was first used to understand and classify the solar irradiance patterns in Durban. These classes have mean profiles that describe the diurnal irradiance patterns.

The classes produced by the clustering were then combined with cloud cover forecasts from the NWP to forecast an irradiance class for the day ahead.

Cluster analysis

Cluster analysis is a technique used for exploring and identifying interesting patterns and distri- butions, and discovering natural groupings within data. This chapter reviews previous studies on clustering, and thereafter focuses on the details of two clustering techniques i.e. hierarchical clus- tering andk-means clustering and explains the purpose of each. Pre-processing of the data that is applied prior to the clustering techniques is also discussed. The minute-resolution horizontal beam irradiance fraction is used an example to illustrate all of the above-mentioned techniques.