Wind Speed Prediction Using Machine Learning Methods Based on Atmospheric Stability and Terrain Type

As wind energy production increases and wind turbine hub heights rise, accurately extrapolating wind speed to higher altitudes is critical for reliably forecasting wind energy production. The power law model has conventionally been used to extrapolate mean wind speed measured at lower levels to turbine-relavant heights. This project explores machine learning (ML) techniques to capture the dynamics of wind behavior and to test ML models for wind speed extrapolation by validation with the established power law model. One year of wind data at multiple levels was collected from a 106-meter meteorological tower in Cedar Rapids, Iowa. Data pre-processing includes outlier removal, identification of failed sensor readings, and data imputation using singular value decomposition (SVD). Terrain was classified as complex or open based on surface features and dominant wind directions. Atmospheric thermal stability was characterized by the bulk Richardson number (Rib). We observed wind shear exponents ranging from 0.12 to 0.39 depending on stability and terrain. Wind speeds at 80m, where a large amount of wind data were not recorded, were predicted using SVD and a Shallow Decoder Network (SDN) and evaluated against recorded values. We calculated the wind shear exponents using predictions from both SVD and SDN models and found it changed by 0.003 on average for the SVD model and 0.010 for the SDN model. Future efforts will evaluate the turbulence intensities of the predicted wind speeds with the SVD and SDN methods.