7.3 A Multi-Stage Regime-Dependent Machine Learning Approach to Short-Term Wind Power Forecasting

Tuesday, 14 January 2020: 11:30 AM
256 (Boston Convention and Exhibition Center)
Tyler C. McCandless, NCAR, Boulder, CO; and S. Naegele and S. E. Haupt

The Kuwait Institute for Scientific Research (KISR) is in the midst of installing the Shagaya Renewable Energy Park in western Kuwait that includes wind power and both photovoltaic and concentrated solar power. As the park is being installed, renewable energy forecasting methods are being developed on the current installed capacity, including short-term wind power forecasts utilizing machine learning techniques. In a hub height wind climatology analysis, it was determined that there are dominant seasonal and diurnal patterns that affect wind power production. On the seasonal scale, shamal winds produce high capacity factors for June through August. On the diurnal scale, nocturnal low-level jets produce stronger wind speeds and higher wind power in the overnight hours. Therefore, the machine learning models should capture both dominant patterns for more accurate short-term wind power production. We test machine learning approaches that are trained on all of the data, such as random forests and artificial neural networks, and compare this to multi-stage models that first classify the seasonal regime (i.e. shamal vs non-shamal), then classify the diurnal pattern with the bulk Richardson number in order to quantify the lower boundary layer stability, and then apply machine learning models in each of these regimes separately. The models are trained on approximately one year of data for five 2 MW wind turbines and the test results are presented on the following year of data for predictions from 15-minutes out to 6-hours.
- Indicates paper has been withdrawn from meeting
- Indicates an Award Winner