Improved Prediction of Near-Surface Winds for Solar Field Siting and Wind Load Design

Site-specific wind conditions impact the design and performance of the collector field in concentrating solar thermal (CST) central receiver plants and single-axis PV tracker arrays. Increasing fidelity and accuracy of near-surface wind model predictions can reduce uncertainty of heliostat design load calculations and yield more effective wind-stow strategies to reduce unnecessary plant downtime. Historical wind observation data at near-surface heights below 10m are sparsely distributed and not widely available for solar field siting considerations. Most numerical weather models provide wind data at 10m or higher, and often at coarse spatial and temporal resolution. Wind conditions at typical solar collector heights of 3m are significantly impacted by surface properties, such as terrain elevation and surface roughness. As a result, solar developers rely on wind data that is not tailored to the requirements of CST and tracking PV applications in terms of available heights and spatial resolution, leading to uncertainty in wind load design and operational performance predictions. This talk will present a machine learning-based approach that combines historical wind data from observation stations and atmospheric model outputs at 10m to generate bias-corrected near-surface wind data at 3m height. Long-term wind observation data at heights ranging from 1.5m to 8m with at least hourly sampling resolution were sourced as a validation set for correcting high-resolution atmospheric model data from the HRRR model (3km resolution) and ERA5 reanalysis (approximately 50km resolution). Two different ML-based techniques were investigated for reduced-dimensional representation of the atmospheric model data by integrating them with sparse observational data incorporating input features such as terrain and land surface properties and generating probabilistic outputs. Results show that ML-based approaches based on random forest quantile mapping of probability density functions and a benchmark neural network model are effective at generating historical near-surface wind time series datasets for selected sites with solar resource potential in the U.S. Southwest.