Calibration of Gridded Wind Speed Forecasts Based on Deep Learning

Abstract

The challenges of applying deep learning (DL) to correct deterministic numerical weather prediction (NWP) biases with non-Gaussian distributions are discussed in this paper. It is known that the DL UNet model is incapable of correcting the bias of strong winds with the traditional loss functions such as the MSE (mean square error), MAE (mean absolute error), and WMAE (weighted mean absolute error). To solve this, a new loss function embedded with a physical constraint called MAE_MR (miss ratio) is proposed. The performance of the UNet model with MAE_MR is compared to UNet traditional loss functions, and statistical post-processing methods like Kalman filter (KF) and the machine learning methods like random forest (RF) in correcting wind speed biases in gridded forecasts from the ECMWF high-resolution model (HRES) in East China for lead times of 1–7 days. In addition to MAE for full wind speed, wind force scales based on the Beaufort scale are derived and evaluated. Compared to raw HRES winds, the MAE of winds corrected by UNet (MAE_MR) improves by 22.8% on average at 24–168 h, while UNet (MAE), UNet (WMAE), UNet (MSE), RF, and KF improve by 18.9%, 18.9%, 17.9%, 13.8%, and 4.3%, respectively. UNet with MSE, MAE, and WMAE shows good correction for wind forces 1–3 and 4, but negative correction for 6 or higher. UNet (MAE_MR) overcomes this, improving accuracy for forces 1–3, 4, 5, and 6 or higher by 11.7%, 16.9%, 11.6%, and 6.4% over HRES. A case study of a strong wind event further shows UNet (MAE_MR) outperforms traditional post-processing in correcting strong wind biases.