Impact of Targeted Measurements and Advanced Machine Learning Techniques on 0-3 Hr Ahead Rapid Update Wind Generation Forecasts in the Tehachapi Wind Resource Area.

A multi-institution project lead by the University of California, Davis is underway to improve short-term (0-15 hours ahead) forecasts of significant wind ramps in the Tehachapi Wind Resource Area (TWRA) of California by gathering data from a targeted network of meteorological instruments and the customization and application of state-of-the-art physics-based and machine learning prediction methods This paper focuses on the component of the project that employed a combination of data from a targeted network of 6 remote sensing devices and the use of a time series type statistical prediction model based on a machine learning method called Extreme Gradient Boosting (XGBOOST) were used to produce 15-minute updates of 0-3 hour ahead forecasts of the 15-minute TWRA average power production. Experiments were conducted to determine the relative impact of data from different sources. In all experiments the prediction model was trained on a rolling 24-month data sample and forecasts were made for a sample of 12 months. The addition of the predictors from the targeted sensor network resulted in an average MAE reduction relative to the forecasts from a baseline method of 5.9% over the entire 0-3 hour forecast period. The MAE reduction associated with the targeted sensor data ranged from a low of 0.1% for a 15-minute forecast to a high of 8.1% for a 180-minute forecast. The majority of the benefit from the targeted sensor network was for look-ahead periods longer than 60 minutes. The impact was dominated by data from one sensor, which was a radar wind profiler located the furthest distance away in the prevailing upstream direction from the area of the TWRA generation facilities. This sensor accounted for more than 50% of the overall benefit of the data from the targeted sensor network in this application. The sensor with the second largest impact was a mini-sodar that was located a very short distance upstream of the TWRA generation facilities. This sensor contributed about 15% to the overall MAE reduction associated with the data from the targeted sensor network. [ABSTRACT FROM AUTHOR]