The Use of Analog Ensembles to Improve Short-Term Wind Forecasting

Accurate short-term wind power forecasts become increasingly important as the penetration of wind generation on electric grids increases. Significant changes in wind speed that lead to large changes in power production are known as “ramps.” When wind power makes up more than a minor portion of power generation, accurate forecasting of ramps is needed several hours in advance to effectively and economically manage the electric grid. This is particularly important on small isolated electric grids, such as on the big island of Hawaii.

Hawaii provides an additional challenge in that significant wind power ramps are often caused by small-scale features that are not well observed, particularly over the data sparse ocean regions located upstream from the wind farms. In an effort to improve the forecasting of ramps in the 0 to 2-hour ahead time frame, the WindNET network of special observing stations was sited near several existing wind farms. The WindNET data has been assimilated into Frequently Updated Numerical weather prediction (FUN) models and used as input into statistical forecast techniques. While this has resulted in some forecast improvement, many ramps are not detected by this approach.

In an effort to increase the number of ramps that can be successfully predicted in the 0 to 2-hour ahead time frame, an analog ensemble forecast technique was developed. The analog ensemble method selects a historical sample of similar situations by picking cases that most closely resemble the current situation. There are several steps to this process. First, a set of observed or simulated “case-matching” variables is chosen. Second, the case-matching score components are computed. A case-matching score component is the difference between a case-matching variable for the forecast case and that same variable from a historical case. Finally, the case-matching score components are combined into a case-matching score that measures the “distance” between the forecast case and the historical case in case-matching variable space. The historical cases with the smallest case-matching scores are selected as ensemble members. The method can be combined with regime-based forecasting by allowing only ensemble members that are classified in the same regime as the current forecast case.

The analog ensemble method is particularly well adapted to generating probabilistic forecasts by inferring a forecast probability distribution from the distribution of ensemble members. The method can also generate deterministic forecasts by using the ensemble mean, the closest member to the mean or another method of generating a composite forecast. Well-chosen case-matching metrics can help increase the number of historical cases in the ensemble that had similar outcomes to the current forecasting situation. Such an ensemble has the potential to generate probabilistic and deterministic forecasts that have significantly greater skill in predicting ramp events than other short-term statistical techniques or the FUN approach.

The analog ensemble method was tested initially for the forecasting of wind speed at a wind farm on the southern tip of the island of Hawaii for a regime named “morning ramp up regime.” This regime is designed to detect situations where the onset of daytime heating is more likely to cause a return to strong east-northeasterly flow from weaker northeasterly flow that often prevails at night. Initial case-matching variables included only those that could be observed by an anemometer at the wind farm and 3 SODARs placed within 15 km upstream from the wind farm. Future plans call for the evaluation of an expanded set of case-matching variables including additional observed variables as well as output from several FUN simulations over Hawaii.

Analog ensemble forecast results were compared to a method that creates an ensemble by choosing all cases that fall within the regime and have an initial wind speed at the wind farm similar to the forecast case. This method, known as the regime and similar speed method, was chosen as a baseline since 1) its deterministic forecast outperformed a persistence-based forecast, and 2) its probabilistic ramp forecast outperformed a ramp forecast based on the diurnal ramp climatology. Preliminary results indicate that deterministic analog ensemble forecasts produced approximately a 3-5% improvement in the mean absolute wind speed forecast error in the first 60 minutes of the forecast period over the regime and similar speed method. The analog ensemble also showed a 5-10% improvement over the regime and similar speed method in the probabilistic ramp forecast of 30-minute ramp rate over the first 20 minutes. Refinement of the analog ensemble approach through improved case-matching parameters or more careful definition of the regime may provide additional benefit.

The conference presentation will review the project in greater detail and show the impact of the choice of case matching variables, size of the ensemble, regime definition and other factors on analog forecast performance.