Do Machine Learning Approaches Offer Skill Improvement for Short-Term Forecasting of Wind Gust Occurrence and Magnitude?

Abstract

Wind gusts, and in particular intense gusts, are societally relevant but extremely challenging to forecast. This study systematically assesses the skill enhancement that can be achieved using artificial neural networks (ANNs) for forecasting of wind gust occurrence and magnitude. Geophysical predictors from the ERA5 reanalysis are used in conjunction with an auto-regressive term in regression and ANN models with different predictors, and varying model complexity. Models are derived and assessed for the warm (April–September) and cold (October–March) seasons for three high passenger-volume airports in the USA. Model uncertainty is assessed by deriving models for 1000 different randomly selected training (70%) and testing (30%) subsets. Gust prediction fidelity in independent test samples is critically dependent on inclusion of an autoregressive term. Gust occurrence probabilities derived using 5-layer ANNs exhibit consistently higher fidelity than those from regression models and shallower ANNs. Inclusion of the autoregressive term and increasing the number of hidden layers in ANNs from 1 to 5 also improve the model performance for gust magnitudes (lower RMSE, increased correlation, and model standard deviations that more closely approximate observed values). Deeper ANNs (e.g., 20 hidden layers) exhibit higher skill in forecasting strong (17 to 25.7 ms−1) and damaging (≥ 25.7 ms−1) wind gusts. However, such deep networks exhibit evidence of overfitting and still substantially underestimate (by 50%) the frequency of strong and damaging wind gusts at the three airports considered herein.

Short‐Term Forecasting of Wind Gusts at Airports Across CONUS Using Machine Learning

“…They require no assumptions about the form of relationships between the predictors and predictands (Abiodun et al., 2018), and have the potential to capture non‐linear systems characterized by complex predictor interactions (Gardner & Dorling, 1998; Reichstein et al., 2019). As such, ANNs offer the potential to improve prediction of sustained and gust wind speeds (Coburn & Pryor, 2022; Sallis et al., 2011; Sheridan, 2018; Wang et al., 2020) by allowing more complex predictor interactions, and may enhance skill for strong (17.5–25.7 m s −1 ) and damaging (≥25.7 m s −1 ) wind gust forecasting (http://www.weather.gov/mlb/wind_threat) (Pryor et al., 2014).…”

“…Other post‐processing tools have been developed that seek to describe wind gust occurrence and magnitude as linear or non‐linear functions of the larger‐scale meteorological context. The transfer functions employ generalized linear regression models (Yang & Tsai, 2019) or more complex machine learning algorithms (Coburn & Pryor, 2022) and are trained using in situ observations of wind gusts, as well as thermodynamic and dynamic predictors such as near‐surface air temperature, vertical temperature gradients, humidity, and wind shear within the atmospheric boundary layer (Coburn & Pryor, 2022; Kahl, 2020). For short‐term forecasting (i.e., forecast horizons <12 hr), statistically based forecasts have been shown to exhibit higher fidelity than direct output from NWP models (Wang et al., 2020).…”

“…For short‐term forecasting (i.e., forecast horizons <12 hr), statistically based forecasts have been shown to exhibit higher fidelity than direct output from NWP models (Wang et al., 2020). However, a key aspect of many previous analyses is an inability to correctly reproduce the most intense, and hence societally relevant, wind gust magnitudes (Coburn & Pryor, 2022; Gutiérrez et al., 2020).…”

“…While skill enhancement from artificial neural networks (ANNs) over regression is well established for some contexts (Pryor et al., 2017; Reichstein et al., 2019), relatively little previous work has considered applications of ANNs to wind gust forecasting. A recent analysis (Coburn & Pryor, 2022) (CP22 hereafter) systematically assessed skill for short‐term predictions of wind gust occurrence and magnitude using transfer functions based on predictors from the ERA5 reanalysis at three high passenger volume airports in the Northeastern United States of America: Chicago O’Hare (KORD), Boston‐Logan (KBOS) and Newark (KEWR) (Figure 1). That study found prediction fidelity is critically dependent on inclusion of an autoregressive (AR) term and is a function of the precise machine learning approach taken to generate the transfer functions.…”

“… Airports from which data are analyzed (KSEA (Seattle‐Tacoma), KLAX (Los Angeles), KPHX (Phoenix Sky Harbor), KDEN (Denver), KDFW (Dallas Fort Worth), KMSP (Minneapolis St. Paul), KATL (Hartford‐Jackson Atlanta), and KMCO (Orlando)). Stations from Coburn and Pryor (2022) are denoted by italics and are shown for reference. Also shown is the terrain elevation (color) and the mean wind speeds (m s −1 ) at 950 hPa from ERA5 (contours).…”