Projet de spécialité - Computationally-efficient high-resolution wind modeling using deep neural networks

De Ensiwiki
Aller à : navigation, rechercher

General information

This project aims at having a concrete impact in geoscience in particular by reducing the amount of time and electricity spent for computing high resolution wind predictions.

The overall idea is to replace a computationally highly costly high-resolution model of wind fields by a neural network (much lighter in term of memory and computational cost) that will predict a fine-scale wind field from a lower resolution model. This have implication for modelling wind and weather in complex terrain such as the Alps.

You will start the project using feedforward neural network and then try to improve your results with more complex convolutional neural networks.

Techniques: The language and libraries is left up to you, but Python and Keras/Pytorch seems a reasonable choice.

Detailed topic

Towards high-resolution wind fields through machine learning techniques

The knowledge of wind fields at high spatial resolution (~ 50 to 30 m) is critical to many applications including wind energy, weather forecasting, and risk prevention. In complex terrain, wind fields are especially hard to assess due to considerable spatial variability induced by processes occurring at different spatial and temporal scales. To overcome this difficulty, high-resolution atmospheric models can be run, that incorporate knowledge about the relevant processes and scales. However, for some applications like numerical weather prediction that currently runs at kilometric scales, modelling the atmosphere at higher resolution is computationally too expensive. The purpose of this study, is to draw on an existing database of simulations of wind-speed at high spatial resolution (30 m), to learn the link between high-resolution topography + coarser-scale wind speed and high-resolution wind field. When this link is learned and incorporated in a trained machine learning algorithm, it can be used to downscale coarse scale wind fields (that are usual products of weather forecast models) over known, high-resolution topography, and thus bypass the need of high-resolution, computationally expensive atmospheric model modelling.

Background: Helbig et al. (2017) have created such a database using the high-resolution ARPS atmospheric model, run on 9000 synthetic topographies. They have applied a non-linear regression method on simple topographic predictors, and proposed a downscaling method for wind speed based on these predictors.

The idea here, is to prolong this work by using more elaborated machine learning techniques. The students will use feedforward neural networks on preprocessed topographical features (slopes, aspects..), and as second step will generalize to predict directly from high-resolution topography and the coarse-scale wind speed, using convolutional neural networks. Additional work can be to predict, not only the wind speed, but also the wind direction (or both zonal and meridian components of the wind vector).

The performance of the new elaborated method will be assessed by cross-validation, and using the method by Helbig et al., 2017 for wind speed as a benchmark. If time allows, an evaluation of the new method as a downscaling tool can also be envisaged on real data from Meteo-France in the French Alps.


At Ensimag: Clovis Galiez (

Collaborators proposing the subject on the geoscience side: Isabelle Gouttevin ( and Nora Helbig (

Data provided

(need to be cited in the report in association with the contact persons if used for the project)

- database from Helbig et al., 2017 - contact person for questions or feedbacks :

- coarse scale wind speed from Meteo-France AROME model (1.3 km) over 2017-2018 - contact person :

- wind data from Meteo-France mountain stations over 2017-2018 - contact person:


Helbig, N., Mott, R., van Herwijnen, A., Winstral, A., & Jonas, T. (2017). Parameterizing surface wind speed over complex topography. Journal of Geophysical Research: Atmospheres, 122(2), 651-667.