California has a goal of generating 33% of its power from renewable energy
sources by 2020. The vagaries of solar radiation will require more accurate
forecasting technologies if such a utility grid is to be successfully built and
operated.
Some attempts to forecast solar power have relied on super computers running NWP
(numerical weather prediction) models. However current NWP models are
unable to forecast cloud density, formation and movement accurately. Other
attempts have focused on predicting solar radiation based on satellite images of
cloud movements, but these models do not forecast the dissipation or
formation of clouds, nor the cloud’s opacity to solar radiation.
Past modeling attempts have lacked an accurate, hourly dataset of solar
radiation that extends back several years. The Solar Data Warehouse now provides
ground-based measurements of solar radiation and other weather parameters in
near-real time for 5000 US locations. It has been demonstrated that this dataset
is more accurate than satellite-based observations and the National Solar
Radiation Database.
There is an old adage: if you want to forecast the weather, predict no rain for
the next 24 hours and you will be right most of the time. In a similar way, many
solar models are based on minimizing average error, but this may not be the best
measure of a solar forecasting system for balancing the electrical grid. To
illustrate: We predict the solar radiation in Fontana, CA during May will be 90%
of the clear sky maximum. The graph above (click for larger version) shows the forecasts from this simple model
overlaid onto the actual observations. For three weeks during May 2009, we see
the average error is only 50 W/m2. These would be great results except that most
of the period was cloud-free. Clearly, there is more to a good model than simply
low average errors. We feel it is important to judge the success of a model on
how well it is able to forecast solar radiation on
difficult days rather than the average error across all days. We
believe we have modeled these difficult days successfully for the L.A. basin in
our Phase I project. Please look through our
results and references.
Out-of-sample testing means applying the model to a significant amount of data
it hasn't seen before (e.g. live forecasting). Without out-of-sample
tests, you are simply judging a model on how well it can do curve fitting. The Phase I
accuracy was better than any published out-of-sample forecasting
results, but clearly some improvements could be made.
For our Phase II project, we created forecasts over a 50 x 50 km grid for the
Greater L.A. region with improved dampening in the model and some new
proprietary techniques. This study was presented at the 2011 ASES conference in Raleigh.
Currently we believe we have the most accurate solar forecasting system
demonstrated by true out-of-sample testing. The next natural step will be
to develop a utility-grade solar forecasting system using live data as well as
cost and risk control measures.
ASES 2011 Links:
Phase II Paper
PowerPoint
Presentation