October 24, 2020


Connecting People

Cloud Cover Nowcasting with Deep Learning

Cloud protect nowcasting is an endeavor to forecast the position of clouds on a limited time scale. Contrary to physical-modeling-based mostly procedures used in regular meteorology, it extrapolates pixels from wind vectors. A modern research applies deep convolutional networks to forecast the position of clouds on a 1h30’ time scale.

Picture courtesy of NASA

Styles based mostly on different neural network architectures have been in a position to master to predict the movement of cloud handles. The outcomes exhibit that the instructed strategy outperforms the two regular impression extrapolation and physical modeling procedures and involves much less computational time. The main limitation of the neural network architectures is that it can’t predict the emergence of new clouds. The cloud protect nowcasting can be employed to forecast the electrical power generation of solar panels or to enhance satellite impression photographs.

Nowcasting is a industry of meteorology which aims at forecasting temperature on a limited expression of up to a few several hours. In the meteorology landscape, this industry is fairly unique as it involves individual tactics, this sort of as info extrapolation, wherever traditional meteorology is commonly based mostly on physical modeling. In this paper, we concentration on cloud protect nowcasting, which has many application spots this sort of as satellite photographs optimisation and photovoltaic vitality generation forecast.
Pursuing modern deep learning successes on numerous imagery tasks, we applied deep convolutionnal neural networks on Meteosat satellite photos for cloud protect nowcasting. We present the outcomes of various architectures specialized in impression segmentation and time sequence prediction. We picked the finest types according to device learning metrics as well as meteorological metrics. All picked architectures showed significant enhancements over persistence and the well-regarded U-Net surpasses AROME physical design.

Link: https://arxiv.org/abdominal muscles/2009.11577