Improving Near-Surface Winds with ML-Based Model Corrections for CSP and Tracking PV Collector Design and Performance

Site-specific wind conditions impact the collector field of concentrating solar power (CSP) plants and tracking photovoltaic (PV) utility plants in multiple ways. Structural design of the collectors must be optimized to avoid wind damage and costly overdesign. Misaligned reflectors due to wind and gust-induced vibrations reduce the optical efficiency of the collector field, leading to power production losses. Moreover, effective wind-aware stow strategies can reduce unnecessary downtime. Despite its importance, historical wind data close to the surface are not widely available. Most numerical weather models provide wind data at 10 m or higher, and often at coarse spatial and temporal resolution. Because of its proximity to the surface, wind conditions at CSP collector heights at 2-3m are greatly impacted by surface properties, such as terrain and land use, and simple extrapolation might not be accurate. As a result, solar developers rely on wind data that is not tailored to the requirements of CSP and tracking PV applications in terms of available heights and spatial resolution, leading to uncertainty in wind-related design and operational planning. To close the data gap, we propose a machine learning-based approach that combines historical wind data from observation stations and atmospheric model output to create more accurate historical near-surface wind data at 2-3m height relevant to CSP and PV. We identified over 2800 long-term wind observation data records at 1.5 m - 8 m height with an average 20-year operation time and at least hourly sampling resolution, covering the entire continental U.S. with a focus on key CSP deployment regions in the U.S. Southwest. These observations serve as a validation set for correcting high-resolution atmospheric model data from the HRRR model (3 km resolution, available since 2014) and ERA5 reanalysis (approximately 50 km resolution, since 1940). We test two different ML-based techniques to correct and enhance site-specific historical wind data: First, a novel framework, called Sig-PCA, for correcting biases in large-scale physics-based model outputs, such as atmospheric simulations, by integrating them with sparse observational data. The core challenge addressed is the high dimensionality of model outputs and the sparsity of observations. Sig-PCA leverages path signatures for unsupervised dimensionality reduction of numerical model data, extracting important spatiotemporal patterns. These reduced representations are then fed into a two-step neural network process: a reconstruction network and a corrective network. The major advantage of using summary statistics is that the amount of data necessary to train a neural network is greatly reduced compared to using full gridded, long-term numerical model data. The second proposed method is quantile mapping of the probability density functions of model and observation wind data for each station with mapping parameters learned from local site and climatological characteristics. Besides wind data from numerical models and observations, both approaches incorporate further input features such as terrain and land surface properties. The goal is to generate historical near-surface wind time series datasets for selected potential CSP sites in the U.S. Southwest. At the AMS meeting, we will present the wind products used and the performance of the two proposed ML-based approaches to correct numerical weather model output, along with an overview of the expected wind datasets. Our analysis highlights the potential for integrating improved near-surface wind datasets into CSP and PV planning tools for better-informed site selection, improved stow strategies, and structural collector design optimization. Future work will focus on expanding the methodology to generate gridded near-surface wind datasets for broader application in solar resource assessment.