[2405.12614] Environment friendly modeling of sub-kilometer floor wind with Gaussian processes and neural networks

[Submitted on 21 May 2024]

View a PDF of the paper titled Environment friendly modeling of sub-kilometer floor wind with Gaussian processes and neural networks, by Francesco Zanetta and a pair of different authors

View PDF

Summary:Precisely representing floor climate on the sub-kilometer scale is essential for optimum decision-making in a variety of functions. This motivates the usage of statistical strategies to supply correct and calibrated probabilistic predictions at a decrease value in comparison with numerical simulations. Wind represents a very difficult variable to mannequin as a result of its excessive spatial and temporal variability. This paper presents a novel method that integrates Gaussian processes (GPs) and neural networks to mannequin floor wind gusts, leveraging a number of knowledge sources, together with numerical climate prediction (NWP) fashions, digital elevation fashions (DEM), and in-situ measurements. Outcomes exhibit the added worth of modeling the multivariate covariance construction of the variable of curiosity, versus solely making use of a univariate probabilistic regression method. Modeling the covariance allows the optimum integration of noticed measurements from floor stations, which is proven to scale back the continual ranked likelihood rating in comparison with the baseline. Furthermore, it permits the direct era of practical fields which can be additionally marginally calibrated, aided by scalable strategies akin to Random Fourier Options (RFF) and pathwise conditioning. We talk about the impact of various modeling decisions, in addition to totally different levels of approximation, and current our outcomes for a case research.