IEEE Power & Energy Magazine - November/December 2015 - 55

Current Experience
and Performance
of Solar Forecasting
A key question for operators and
other end users is "how well do solar
forecasts perform?" The answer
depends on a number of factors,
many of which are also relevant for
wind forecasting. For example, any
forecast performance metric should
be tied to the end-use application's
sensitivity to forecast error. However, many users (system operators, traders, utilities, etc.) may not
have a quantitative understanding
of their sensitivity to forecast error
and the related value of the forecast.
Therefore, a standard set of widely
used metrics is typically employed.
100%
Standardized metrics facilitate the
comparison of solar forecasts, but
they may not accurately inform a
specific user as to which forecasts
50%
provide the most value for their specific application.
Currently, the three most widely
used metrics for assessing deterministic point forecasts are mean
0%
(bias) error, mean absolute error
09:30
(MAE), and root mean square error
(RMSE). These metrics may be
10.07.2012
applied using different approaches.
This Map Shows the Relative Solar Power Production in Germany in Proportion
to the Installed Power.
The most widely used approach, as
followed here, is to express the metrics as a percentage of the installed figure 4. The real-time estimation of the solar power production on a zip code level
capacity over all daylight hours. in Germany based on online data from inverters. (Source: energy & meteo systems.)
Other approaches calculate the perand aggregates of facilities. The MAE rises rapidly
centage change in these metrics relative to a reference forecast
over the first few hours after the issue time. After about
(persistence or climatology) or as a percentage of average actual
six hours, the rate of error growth decreases substanproduction over the evaluation period. There is also a movement
tially. It can also be seen that forecasting performance
toward comparisons using "smart persistence," which assumes
improves with further forecast system development in
that relevant conditions such as cloud cover and temperature
the region, as 2014 shows a significantly lower error
remain the same but includes the underlying variation due to
rate than 2013 does.
changing solar angle in the baseline.
2) Variability in Solar Production. In general terms,
These metrics are simple to compute but one still needs to
absolute forecast performance is impacted strongly
interpret the resulting metric values. Forecast performance is
by the amount of atmospheric-based variability in
impacted by many factors, as described below.
the solar power production. For example, the abso1) Look-Ahead Time. One of the most intuitive factors
lute level of cloud impact on solar irradiance is lower
that impacts forecast performance is the forecast looknear sunrise or sunset, although the relative variation
ahead time (also known as the forecast time horizon or
(the percentage of average irradiance change) may be
lead time). Figure 5 shows an example development of
much larger. Thus, variability in production and the
the forecast error (error growth) over a forecast horizon
magnitude of forecast errors tend to follow the diurnal
of 85 h. The figure depicts the MAE over two years for
cycle. This is illustrated in Figure 6 for a nontracking
the aggregate solar generation in Spain, although the
facility in Texas. The relationship between the diurnal
pattern is somewhat typical for both individual facilities
november/december 2015

ieee power & energy magazine

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