Short Term Solar Radiation Forecasts Using Weather Regime Dependent Artificial Intelligence Techniques

- Indicates paper has been withdrawn from meeting
- Indicates an Award Winner
Wednesday, 5 February 2014: 9:30 AM
Room C204 (The Georgia World Congress Center )
Tyler C. McCandless, NCAR, Boulder, CO; and S. E. Haupt and G. S. Young
Manuscript (1.6 MB)

Accurate short term forecasts of solar radiation can provide utility companies and independent system operators with vital information for regulating the fast response of generators to the variability of solar energy. Forecasting in this time range requires first building a clear-sky model, then predicting the variability around that model based on observed data.

The goal of our effort is to provide short term (up to three hour) probabilistic forecasts that are calibrated to specifically address the short term variability in solar radiation. The first step in forecasting solar radiation is to build a clear-sky model for each location. The clear-sky model uses a ten-day sliding window average for days that are defined as clear-sky days. Days are defined as clear-sky based on the variability of the each day's irradiance time series compared to the climatological average clear-sky days for each location.

The next step of our method is a regime identification method of forecasting solar irradiance in order to estimate the type and timing of variability due to cloudiness. A k-nearest neighbor (KNN) technique is used to cluster similar irradiance time series while Self-Organizing Maps (SOMs) are used on Numerical Weather Prediction (NWP) data for a domain encompassing the forecast region. An Artificial Neural Network (ANN) is then trained on each regime in order to more accurately capture the physics, and therefore predictive ability, of the weather pattern identified by the SOMs. The preliminary results of our forecasting method quantify the predictive ability for solar irradiance forecasts from fifteen minutes to three hours lead time.