IEEE Power & Energy Magazine - November/December 2015 - 53

while the second uses different numerical models or physics schemes. Both provide insights into the likelihood of
extreme events, such as ramps of solar output due to changes
in cloud cover, fog, or other atmospheric features. Ensembles can be applied to any forecasting approach, although
they are most well developed for NWP.
Ensemble forecasts for solar generation are used in power
markets where the uncertainty of forecasting leads to price
volatility. They may also be employed to help in determining
reserves or scheduling generation. In such real-world applications, percentiles can be used to make the uncertainty visible
and so more easily interpreted. Figure 2 shows how percentile "bands" change in size around a mean value (white line)
and the true generation (black line). Percentiles represent a
value below which the given percentage of the outcomes is
expected to fall. For example, 10% of the outcomes should
fall below the tenth percentile (P10) value, while 90% of the
outcomes should fall below the 90th percentile (P90) value.

Distributed and Behind-the-Meter
Solar PV Forecasting
A number of factors can impact how solar forecasts are
developed and used, particularly when the solar generation is
connected to the distribution network or is BTM:
✔ Whether the generation of a commercial or residential
solar power system is recorded on its own or aggregated with customer load.
✔ Whether telemetered meter data is available in real time.
✔ Whether detailed static data (i.e., metadata) is available (the plant location, the PV geometry, nearby
obstructions, hardware information, and the like).

At a large solar plant, meter data is typically available
in near-real time and site metadata is generally known. For
smaller distribution connected plants, dedicated meter data
is usually recorded but is often not telemetered in real time.
Metadata may be difficult to obtain as well. For BTM solar,
real-time generator data is rarely known, meter data is often
net of load, and metadata is difficult to obtain.
Therefore, distribution-connected PVs, in particular BTM
PVs, can be more difficult to forecast. If detailed metadata
is available, the data can be combined with irradiance and
weather data, typically from satellite and/or NWP sources as
described earlier. For example, information about all PV systems in the state of California is recorded and then combined
with high-resolution irradiance values and weather predictions
to forecast output for the entire state. Such a "bottom-up"
approach is employed by the California Independent System
Operator to predict the total contribution of BTM solar on their
grid, as shown in Figure 3.
This approach can be quite cost-effective, particularly in
areas having a large number of distributed PVs (California,
for example, has more than 200,000 unique PV systems).
It allows for wide-area aggregation to support independent
system operators in utility-wide prediction and also for
regional analysis, where distribution system operators and
planners need to understand localized impacts. One potential benefit of this approach is that the particulars of each
system are explicitly captured. For example, a PV system
that is oriented toward the east will produce a different
power curve than a south- or west-facing system, so the
fleet composition will influence the resulting power profile
for the overall system.

100
90
80
70

(%)

60
50
40
30
20
10

M
ay
02
:0
0
04
:0
0
06
:0
0
08
:0
0
10
:0
0
12
:0
0
14
:0
0
16
:0
18 0
:0
0
20
:0
0
22
:0
19 0
M
ay
02
:0
0
04
:0
0
06
:0
0
08
:0
0
10
:0
0
12
:0
0
14
:0
0
16
:0
18 0
:0
0
20
:0
0
22
:0
20 0
M
ay

figure 2. An example of a two-day solar power forecast with nine percentiles (P10 through P90) and the actual measurements (black dotted line.)
november/december 2015

ieee power & energy magazine

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