IEEE Power & Energy Magazine - November/December 2015 - 56

Annual Solar Power Production Forecast MAE
Spain (Pinst = 4.4 GW)

MAE (% of Capacity)

5
4
3
2
1

2013
2014
1
3
5
7
9
11
13
15
17
19
21
23
25
27
29
31
33
35
37
39
41
43
45
47

0
Look-Ahead Period (h)

MAE or Actual
STDEV (% of Capacity)

figure 5. Forecast performance over the look-ahead period
of the forecast for aggregated 4.4 GW of PVs in Spain.

Forecast MAE and
Production Variability by Hour of the Day

25
20
15
10
5
0

6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21
Hour of the Day (Hour Ending in LST)
MAE: Dual Track
STDEV: Dual Track

MAE: No Track
STDEV: No Track

figure 6. Annual MAE versus variability (observed standard
deviation) by hour of the day for a nontracking facility (red)
and a nearby dual-axis tracking facility (black) in Texas.

cycle of variability and MAE is clearly evident in data
presented in this chart.
3) Specific Plant Attributes. The various plant attributes
of a solar generation facility are important as well, as
Figure 6 also illustrates. The solid lines show the variability and MAE for a dual-axis tracking facility that
is located close to (i.e., having very similar weather
regimes) the facility without tracking, as represented by
the dashed lines. The chart indicates that the variability
and MAE for both facilities is similar at midday and
near sunrise and sunset. However, the variability and
MAE of the dual-axis tracker is substantially higher
during the mid-morning and mid-afternoon hours. The
impact of this pattern on the overall forecast MAE for
the two sites is quite substantial.
4) Spatial Scale. A fourth factor that has a significant
impact on forecast performance is the spatial scale
56

ieee power & energy magazine

of the generation. Due to the well-known aggregation
effect, forecasts for geographically diverse aggregates
of solar generation facilities have smaller errors than
the forecasts for individual facilities in the aggregate.
Local effects, which are more random and more difficult to forecast, tend to average away when we look
at the aggregated forecast. A broad view of the effect
of aggregation on the performance of day-ahead solar
power production forecasts is depicted in Figure 7. This
chart shows the annual MAE over all daylight hours
for day-ahead forecasts for a broad spectrum, ranging
from individual centralized facilities to regional and
system-wide aggregates of centralized facilities and
distributed generation. The power of the aggregation
effect in reducing the absolute forecast errors is quite
evident. Note that many factors impact forecast performance for a given entity of a specific size, such as the
geographic diversity within the entity, the attributes of
the facilities (e.g., tracking versus nontracking, and so
forth), the amount of variability associated with local
weather regimes, and the causes of the variability.
5) Other Weather Phenomena. While forecasting
cloud cover is a predominant factor for solar forecasting ,there are some other phenomena that can have a
similar impact on the predictability of PV generation,
especially on day-ahead and longer time scales. These
include fog, snow, and dust. An example of the impact
of fog and snow on the performance of forecasts for
a large aggregate of generation facilities in Germany
is shown in Figure 8. The errors are much larger in
the spring and fall when fog is a significant factor in
power production variability.

Paths to Improved Solar Forecasting
Solar power forecasting is still a relatively new technology.
Individual methods have deficiencies, such as the lack of
attention that NWP models have traditionally paid to cloud
cover variables and the relative immaturity of sky imaging
techniques, while the blending of different forecasting techniques is still less than optimal. The fact that much PV generation is BTM means some of the data needed to set up models
may not be available, and methods are still being developed to
forecast this type of PV generation. Probabilistic forecasting
techniques that could significantly improve on the representation of uncertainty are still being explored. Finally, ineffective
communication of forecast information to the users and the
lack of end use tools for using forecast information hinder the
extraction of full value from forecasts.
A number of initiatives are currently underway in the
United States and Europe to improve some of these key areas
of solar power forecasting. These are covered in more detail
in forums such as the Utility Variable Generation Integration
Group's annual forecasting workshop and the energy track
of the American Meteorological Society, but the following
describes the major research areas, with specific examples:
november/december 2015

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