Wind‐Topo: Downscaling near‐surface wind fields to high‐resolution topography in highly complex terrain with deep learning

Abstract

Predicting wind flow in highly complex terrain like the Alps is a challenge for all models. When physical processes need to be resolved in a spatially explicit manner, grids with high horizontal resolution of a few hundred meters are often required and drastically limit, in many cases, the extent and duration of the simulations. Many surface process models, like the simulation of heterogeneous snow cover across a season, however, need long time series on large domains as inputs. Statistical downscaling can provide the required data, but no model can reach the desired resolutions effectively and provide temporally resolved wind speed and direction on highly complex topography. The assessment of the potential for wind energy in the Alps, a promising player in the energy transition, is an example where the current shortcomings cause strong limitations. We present "Wind-Topo", a novel approach based on deep learning that discovers some of the interactions between high-resolution topography and coarser-resolution states of the atmosphere to generate near-surface wind fields with a 50-m resolution. In our test case, we use a large number of measurement stations in Switzerland to train the model and an operational weather prediction model (COSMO-1) as predictor. Wind-Topo employs a custom architecture that analyses the state of the atmosphere on various scales and associates it with high-resolution topography. A dedicated loss function leads to good scoring metrics as well as accurate wind-speed distributions at 60 independent stations used for a thorough validation. 50-m resolution wind fields are generated efficiently and exhibit several expected orographic effects like ridge acceleration, sheltering, and deflection. Furthermore, the bias and mean absolute error from COSMO-1 at the alpine validation stations, which are 0.72 and 1.77 m⋅s −1 respectively, are reduced to −0.07 and 1.21 m⋅s −1 .

Terrain-Guided Flatten Memory Network for Deep Spatial Wind Downscaling

“…[62], USRNET [41] and the newly proposed TIGAM are superior to the remaining methods. WindTopo [31] outperforms most of the single image super resolution methods and the performance of Restormer [45] is quite competitive to TIGAM. From this table, it should be noted that the proposed TIGAM outperforms other methods consistently on all of the evaluated metrics (i.e., MAE, SSIM, and PSNR) in terms of both U and V components, demonstrating the effectiveness of the proposed method.…”

“…Analysis. According to early works [30], [29], [31] and after a statical analysis, we say there is high correlation between terrain and wind, especially between terrain std and wind std. Unfortunately, how this relevance is mathematically formulated is unknown and has not been revealed.…”

“…To verify the superiority of the proposed TIGAM, 20 kinds of methods including both of traditional interpolation methods such as bilinear, bicubic and nearest interpolation, and advanced deep learning methods such as SRCNN (Super Resolution Convolutional Neural Network) [32], SubPixel [34], VDSR (Very Deep Super Resolution network) [39], FSRCNN (Fast Super Resolution Convolutional Neural Network) [33], EDSR (Enhanced Deep residual Super Resolution network) [35], SRDenseNet (Super Resolution DenseNet) [56], LAPSRN (LAPlacian Super Resolution Network) [57], RDN (Residual Dense Network) [58], RCAN (Residual Channel Attention Network) [36], DBPN (Deep Back-Projection Network) [59], RNAN (Residual Non-local Attention Network) [60], SRFBN (Super Resolution FeedBack Network) [61], SAN (Second-order Attention Network) [62], USRNET (Unfold Super Resolution Network) [41], RFANet (Residual Feature Aggregation Network) [40], CSNLA (Cross-Scale Non-Local Attention network) [63], NLSN (Non-Local Sparse Attention network) [64] SwinIR [44], Restormer [45] and WindTopo [31] etc are discussed in detail. Specifically, two categories of experiments are conducted, i.e.,…”

“…SwinIR [44] is a recently proposed dedicated image super resolution structure based upon Swin Transformer [65]. WindTopo [31] is a recently proposed deep learning based wind downscaling method using high-resolution topograph. Apart from that, Restormer [45] is a recently proposed transformer architecture especially designed for highresolution image restoration, such as denoising, deblurring, and deraining.…”

“…Given the wind-terrain interaction, the downscaling method effectively improves the error metrics and spatial distributions correlation between the downscaled wind field and observation. Based on convolutional neural network, Dujardin and Lehning [31] propose a near-surface wind fields downscaling approach which considers the state of the atmosphere on various scales and its interaction with highresolution topography. The downscaling method perform well in generating 50-m resolution wind fields, especially under the effects of complex topography like ridge acceleration, sheltering, and deflection.…”