IEEE Power & Energy Magazine - November/December 2015 - 52

Ensemble Methods and Blending of Forecasts

Sky Imagers

Numerical Weather Prediction
with Model Output Statistics

Climatology

Satellite Imagery

Statistical Learning Methods

3
4
5
6
Time Horizon for Which Method Shows Value (Hours)

120

240

figure 1. An overview of solar forecasting methods, based on time horizon for which they show value.

motion. Clouds scatter some wavelengths of light more than
they do others, and these can be used to categorize clouds as
thick and thin as well as to differentiate them from aerosols
or dust. Consecutive images can estimate cloud velocity and
so can provide a very short-horizon forecast. If additional sky
imagers are available, triangulation can provide information on
cloud base and depth as well as differential advection speed at
different cloud levels. Sky imagers are of significantly decreasing value beyond approximately a 30-min time frame; moreover, they are expensive to use relative to most other forms of
very short-term prediction.

Satellite Imaging
Providing value for a longer look-ahead time frame, geostationary satellite data from networks such as the National
Oceanic and Atmospheric Administration (NOAA) Geostationary Operational Environmental Satellite network for
North and South America and the MeteoSat network for
Europe, Africa, and central Asia supply information about
cloud properties and movement. First, a physically driven
model predicts clear-sky conditions at a specific site, using
inputs like aerosol content, water vapor, elevation, and ozone.
This modeled clear-sky irradiance is then modulated by estimated irradiance derived from satellite images. Sequential
satellite images are combined to create cloud motion vector fields that can be used to predict future cloud locations.
This technique has been shown to be effective in forecasting
irradiance from one minute to as far as five hours ahead.
However, it performs less effectively during conditions when
clouds are rapidly forming or dissipating, such as convective
and marine layer cloud regimes.
52

ieee power & energy magazine

Numerical Weather Prediction
NWP systems have been the workhorse of forecasting
applications for many years. Traditionally, these have performed best for the time horizon from six hours to two
weeks; recently deployed rapid-update systems (such as
the NOAA High-Resolution Rapid Refresh system) can
provide value at shorter time scales. NWP methods have
been well described in previous articles in relation to
wind. These involve solving the Navier-Stokes equations to
model the weather, combined with modules that compute
other physical processes, such as radiative transfer, land
surface effects, and cloud microphysics parameterization.
Historically, these models have been optimized for predicting variables such as temperature, humidity, probability
of precipitation, and wind. Only recently has there been an
emphasis on improving prediction of surface solar irradiance.
Statistical learning methods are often used to correct for errors
in the NWP model output and to blend output from multiple
models in a process referred to as model output statistics. It is
typical for such methods to improve upon the raw forecast of
an NWP model by about 10-15%.

Ensemble Forecasting
Ensemble methods are important to deal with the uncertainty inherent to solar power forecasting, particularly during morning and evening ramps and with cloud cover and
fog. An ensemble consists of a collection of forecasts and
can represent the uncertainty in the calculation of the state of
the atmosphere inherent in forecasting methods. Two major
techniques are employed in creating an ensemble: the first
adds perturbations to the initial state of a forecasting model,
november/december 2015

Table of Contents for the Digital Edition of IEEE Power & Energy Magazine - November/December 2015