An Application of an Analog Ensemble for Short-Term Solar Power Forecasting

- Indicates paper has been withdrawn from meeting
- Indicates an Award Winner
Tuesday, 6 January 2015: 11:30 AM
224B (Phoenix Convention Center - West and North Buildings)
Stefano Alessandrini, NCAR, Boulder, Colorado; and L. delle Monache, T. Brummet, S. E. Haupt, and G. Wiener

The efficient integration of solar power in the energy market is limited by its natural variability and predictability. A cost-effective utilization of solar energy over a grid strongly depends on the accuracy and reliability of the solar power forecasts available to the Transmission System Operators (TSOs). In several countries the legislation requires solar power producers to pay penalties proportional to the errors of day-ahead energy forecasts, which makes the accuracy of such predictions a determining factor for producers to increase their revenues. To this end, probabilistic predictions can provide accurate deterministic forecasts along with a quantification of their uncertainty, as well as a reliable estimate of the probability to overcome a certain production threshold.

The analog ensemble (AnEn) technique has been developed by NCAR (Delle Monache et.al 2011, 2013) and it has been extensively tested for the probabilistic prediction of both meteorological variables and wind power. We will present a novel application of AnEn to solar power forecasting (from photovoltaic panels). The AnEn is based on an historical set of deterministic predictions and observations of the quantity to be predicted. For each forecast lead time and location, the ensemble prediction of a given variable is constituted by a set of measurements of the past (i.e., 1-hour averages of solar power). These measurements are those concurrent to past deterministic predictions for the same lead time and location, chosen based on their similarity to the current forecast. The meteorological variables used to identify the past forecast similar to the current one are called analog predictors. The variable to be predicted, the predictand, is the 1-hour average of the produced solar power. One of the advantages of applying AnEn to solar power predictions is that a radiation to power conversion curve specific for each production unit is not necessary as such conversion is built-in the AnEn approach.

Data from eight solar power production units are used to test the AnEn for solar power forecasting. These units are located in the Sacramento metropolitan area, in California. One-hour average power data are available for an 8-month period. Also, in the proximity of the production units hourly averaged measurements of 2-m air temperature (T2M), global horizontal irradiation (GHI) and direct normal irradiation (DNI) are available.

The deterministic predictions used to generate the AnEn are those from NCAR's DICast system, which is available for the whole period covered by solar power measurements. Every DICast run starts at 1200 UTC and it includes 72 hourly lead times. The forecast time series of GHI, cloud cover (CC), DNI, and T2M have been computed at the solar farm locations and used as analog predictors for the AnEn metric computation. The hourly average azimuth angle (AZ) and solar elevation angle (EL) have been computed separately and then added to the set of analog predictors. The latter allow to define the sun position and to take into account possible obstacle's shadows (e.g., buildings or mountains).

The “analogs” are selected from a “training period” defined by the first 5-months of the entire data set. The remaining part (3 months) of the dataset is used to estimate the AnEn performance. To mimic real-time operations, for each forecast the training goes from the start of the period up to one day before the date the forecast was issued.

An in-depth analysis of important attributes of probabilistic predictions generated with the AnEn will be presented. These attributes include statistical consistency, reliability, resolution, sharpness, and the spread-skill relationship. The AnEn provides reliable, sharp, and statistical consistent probabilistic solar power predictions, at a fraction of the real-time computational cost of traditional ensemble methods.