Abstract. Near-surface winds over complex terrain generally feature a large variability at the local scale. Forecasting these winds requires high-resolution numerical weather prediction (NWP) models, which drastically increase the duration of simulations and hinder them in running on a routine basis. Nevertheless, downscaling methods can help in forecasting such wind flows at limited numerical cost. In this study, we present a statistical downscaling of WRF (Weather Research and Forecasting) wind forecasts over southeastern France (including the southwestern part of the Alps) from its original 9 km resolution onto a 1 km resolution grid (1 km NWP model outputs are used to fit our statistical models). Downscaling is performed using convolutional neural networks (CNNs), which are the most powerful machine learning tool for processing images or any kind of gridded data, as demonstrated by recent studies dealing with wind forecast downscaling. The previous studies mostly focused on testing new model architectures. In this study, we aimed to extend these works by exploring different output variables and their associated loss function. We found that there is no one approach that outperforms the others in terms of both the direction and the speed at the same time. Finally, the best overall performance is obtained by combining two CNNs, one dedicated to the direction forecast based on the calculation of the normalized wind components using a customized mean squared error (MSE) loss function and the other dedicated to the speed forecast based on the calculation of the wind components and using another customized MSE loss function. Local-scale, topography-related wind features, which were poorly forecast at 9 km, are now well reproduced, both for speed (e.g., acceleration on the ridge, leeward deceleration, sheltering in valleys) and direction (deflection, valley channeling). There is a general improvement in the forecast, especially during the nighttime stable stratification period, which is the most difficult period to forecast. The result is that, after downscaling, the wind speed bias is reduced from −0.55 to −0.01 m s−1, the wind speed MAE is reduced from 1.02 to 0.69 m s−1 (32 % reduction) and the wind direction MAE is reduced from 25.9 to 15.5∘ (40 % reduction) in comparison with the 9 km resolution forecast.

中文翻译：

---

使用卷积神经网络缩小地面风预报

摘要。复杂地形上的近地表风通常在局部范围内具有较大的变化性。预测这些风需要高分辨率数值天气预报（NWP）模型，这大大增加了模拟的持续时间并阻碍了模拟的正常运行。然而，降尺度方法可以帮助以有限的数值成本预测此类风流。在这项研究中，我们将法国东南部（包括阿尔卑斯山西南部）的 WRF（天气研究和预报）风力预报从原来的 9 公里分辨率降尺度到 1 公里分辨率网格（1 公里 NWP 模型输出为用于拟合我们的统计模型）。降尺度是使用卷积神经网络 (CNN) 执行的，卷积神经网络是处理图像或任何类型网格数据的最强大的机器学习工具，最近处理风预报降尺度的研究证明了这一点。之前的研究主要集中在测试新的模型架构。在本研究中，我们的目标是通过探索不同的输出变量及其相关的损失函数来扩展这些工作。我们发现没有一种方法能够同时在方向和速度方面优于其他方法。最后，通过组合两个 CNN 获得最佳整体性能，一个专用于基于使用定制均方误差 (MSE) 损失函数计算归一化风分量的方向预测，另一个专用于基于计算风分量并使用另一个定制的 MSE 损失函数。当地规模、与地形相关的风特征在9公里处的预报效果不佳，现在可以很好地再现速度（例如，山脊加速、背风减速、山谷遮挡）和方向（偏转、山谷窜流）。预报普遍有所改善，特别是在夜间稳定层结期，这是预报最困难的时期。结果是，降尺度后，风速偏差从-0.55减小到-0.01 m s-1，风速MAE从1.02减小到0.69 m s-1（减小32%），风向MAE从与 9 公里分辨率预测相比，25.9 至 15.5∘（减少 40%）。