J4.6 Evaluating Bias Correction Methods for Gridded WRF Output Using Floating LiDAR Buoy Measurements

Monday, 29 January 2024: 5:45 PM
347/348 (The Baltimore Convention Center)
Polina Khapikova, TGS, Houston, TX; California Institute of Technology, Pasadena, CA; and K. Brennan, S. J. Eichelberger, PhD, B. Lasscock, and A. Sansal

Accurate assessment of the wind resource is crucial for risk mitigation during the development of offshore wind projects. Weather Research and Forecast (WRF) simulations are commonly used to provide spatially complete information about atmospheric conditions in areas of offshore wind development, but are known to have biases that introduce uncertainty and risk into wind resource assessments. Bias correcting the WRF wind output would reduce such risk, but several challenges emerge. First, few observations exist over the ocean, particularly measurements at wind turbine hub height, which ranges from 120 to 160 meters. To-date, the gold standard for measurements is floating LiDAR buoys, which provide vertically resolved measurements of wind speed and direction at a single location. Because of the high cost of floating LiDAR buoy deployments, measurement campaigns are often limited in length to 1-2 years at a time. Thus, wind resource measurements from LiDAR buoys have sparse coverage in both space and time, but these data are the best available solution for determining and correcting model biases. Further, for applications in the energy sector, it is desired to reduce uncertainty in wind energy and not just wind speed. The energy produced by wind turbines depends nonlinearly on wind speed, due to both the relationship between wind energy and wind speed and the turbine cut-in and cut-out values. We therefore desire a bias correction method that incorporates both wind energy and wind speed biases into its methods and evaluation.

Here, we evaluate different methods for bias correcting gridded WRF output, using a suite of 12 floating LiDAR buoy observations along the US East Coast. Observations are densest in the New York Bight region, but spatial coverage extends from offshore Massachusetts down to the Virginia-North Carolina border. Our bias correction process occurs in two steps: a point model is first trained at the location of each buoy and then corrections are applied spatially to the entire WRF domain. We investigate the performance of multiple statistical and ML-based point model implementations centered around the GradientBoost algorithm. We analyze the impact of varying the implementation of GradientBoost and systematically tuning model hyperparameters using Optuna. Further we investigate approaches to ensure biases in wind energy are reduced, including incorporating a quantile mapping and defining a custom loss function. We use a variety of metrics to evaluate model performance, which elucidate wind speed and energy biases before and after the correction is applied. Bias correction skill is assessed both spatially and temporally in and out of sample to rank the correction schemes and determine the preferred solution. The results of this validation analysis help quantify WRF model biases and estimate uncertainty of bias-corrected WRF model data in offshore conditions along the US East Coast. Such information is critical for accurately assessing the wind resource conditions and for optimal design of future offshore wind measurement campaigns.

- Indicates paper has been withdrawn from meeting
- Indicates an Award Winner