Domain-informed CNN architectures for downscaling regional wind forecasts

High-resolution wind speed forecasts are of great importance to the wind energy industry, from short-term energy forecasting and trading to longer-term resource assessment and planning. Generating high-resolution regional wind forecasts currently requires compute-intensive numerical models to downscale from a global forecast. Black-box AI models, once trained, can produce results in a fraction of the time and cost; however, they tend to produce smoothed outputs, are not interpretable and generalise poorly. The domain-informed AI architecture presented in this work seeks to address these problems by incorporating prior static fields directly into the model architecture. Specifically, the proposed approach combines two sequential U-Nets – the first upsamples the input wind fields and expands the number of feature maps, a fusion layer then injects prior static data such as topography, and a second U-Net generates the final output wind field. This approach improves all performance metrics versus a baseline U-Net model and generalises better to out-of-sample scenarios. In addition, this study compares the performance of several loss functions, including standard pixel-wise measures such as mean-squared error, structural similarity and frequency-focused functions, and a function based on Wiener filter theory. All loss functions, with the exception of the Wiener loss, perform comparably and tend to attenuate higher-frequency detail. Although the Wiener loss encourages higher frequencies, it over-estimates amplitudes. A composite Wiener-L1 loss function balances generating high-frequency detail and correctly predicting amplitudes.