Multivariable neural network to postprocess short‐term, hub‐height wind forecasts

Abstract

The negative effects of environmental problems such as climate change, pollution, and energy-security have increased the pressure for cleaner and more efficient renewable energy sources. 1 Wind power is a fundamental part of the transition to renewable energy 2 ; in 2017, an estimated 17% of all renewable-generated electricity worldwide came from wind sources, 3 and wind energy corresponded to approximately 23% of all the renewable energy production capacity worldwide. 4 The potential for the exploitation of this resource is enormous; it is estimated that by 2050 it will be able to supply approximately 4.4 TW, which corresponds to roughly 37% of all of the end-use power supply in the world. 1

Improving Atmospheric Angular Momentum Forecasts by Machine Learning

“…However, ML methods are also rapidly advancing in Earth sciences and can solve a plethora of classification, data-augmentation, inversion, and modeling problems in this field (Irrgang et al, 2021;Lary et al, 2016;Salcedo-Sanz et al, 2020). Especially in numerical weather prediction, postprocessing forecast variables by ML has become an efficient tool to improve the prediction skills for specific application, e.g., postprocess short-term hub-height wind by multivariable neural network (Salazar et al, 2021), reducing the error in ECMWF lower stratosphere wind prediction by 2%-15% for wind speed and 15%-25% for direction (Candido et al, 2020). The advantage of such a postprocessing approach is the improvement of prediction skills for derived parameters such as AAM without the need of improving the whole numerical weather prediction system which is often beyond the research scope.…”

Deep generative model for probabilistic wind speed and wind power estimation at a wind farm

“…In a similar fashion, we present in Figure 4 the RMSE values for Raw, MOS and the averaged CVAE samples for every of the 24 forecast lead times. At all sites, there is a period of strong winds that extends from approximately the 9th forecasted hour (17:00 JST) to the 12th forecasted hour (20:00 JST) 17 . This translates into a higher RMSE in the raw forecast, as shown by the red line in Figure 4.…”

“…At all sites, there is a period of strong winds that extends from approximately the 9th forecasted hour (17:00 JST) to the 12th forecasted hour (20:00 JST). 17 This translates into a higher RMSE in the raw forecast, as shown by the red line in Figure 4. CVAE is able to generate forecasts with reduced error for all lead hours, outperforming MOS as well.…”

“…14,15 Both approaches rely on predictors provided by a dynamical model to reduce errors and biases in the atmospheric variable of interest. More recently, machine learning methods such as neural networks 16,17 have also been employed. Many of these approaches carry out parameter optimization based on a target or loss function.…”

“…We consider that the opportunities that data‐driven methods offer to generate ensembles and/or obtain uncertainty information from deterministic forecasts has not been sufficiently exploited. To the best of our knowledge, out of the few existing applications of generative models for NWP post‐processing, none uses a multivariate approach (such as the one in, 17 for deterministic forecasts), and none has been validated in a zone with complex geography an highly fluctuating winds such as the Japanese archipelago. Therefore, we introduce a method to generate probabilistic wind speed forecasts from single‐run deterministic NWP forecasts that satisfies all these requirements and that can be of practical utility for wind farm operators.…”