Characterizing Wind Fields and Evaluating Model Performance in Forested Complex Terrain in southwest Washington

Accurate forecasting of wind flows a few hundred meters above the ground is crucial for optimizing wind power generation, but is particularly difficult to do in complex terrains influenced by orography and forest canopies. The challenge lies in the variability of surface properties and exchange processes, which can affect near-surface wind profiles differently under diverse weather forcing conditions. By leveraging observation datasets collected from a tall (74 m) meteorological tower, profiling lidar, and scanning lidar at the Wind River Experimental Forest (WREF) in southwestern Washington State, we characterize the observed wind flows and evaluate the simulations from the High-Resolution-Rapid-Response (HRRR) model under various weather conditions. We develop a machine-learning framework, called TCLViT, for the classification of weather systems. TCLViT integrates a Vision Transformer (ViT) to capture spatial relationships and Temporal Contrastive Learning (TCL) to account for temporal context information. This combination enables the extraction of spatial-temporal features in the meteorological variables, which are then processed through a hierarchical clustering algorithm. This approach identifies five typical weather patterns over the first phase of the lidar deployment from late summer to fall 2023.

Weather patterns induce significant variations in rotor-disk (~100-300 m above ground) wind conditions, both temporally and vertically, at the WREF site. Notable model errors are observed in vertical wind profiles due to the lack of canopy representation. This underrepresentation of canopy processes also impacts rotor-disk winds. Under strong large-scale forcing, wind speeds at 200 meters are low with minimal diurnal variation, and the model substantially overestimates wind speeds. Conversely, under weak large-scale forcing, thermal contrasts and mountain-valley flows drive diurnal wind speed variations, with higher winds in the early afternoon. The modeled winds peak in the late afternoon, with a pronounced overestimation during nighttime. During the second phase of lidar deployment, covering winter 2023 to summer 2024, we will apply the same analysis to classify weather patterns, characterize wind flows, and evaluate model performance in cold and transitional seasons.