Learning with deep Gaussian processes and homothety in weather simulation

Observations and numerical prediction models are the main methods for measuring and estimating the earth's energy balance parameters. However, the margin of error is very high when simulating meteorological parameters (e.g., latent heat and sensible heat) due to natural disturbances such as the coverage of sensors by clouds. To reduce the margin of error, a machine learning method representing model uncertainty can be useful to minimize fluctuations on measured parameters. In this paper, a learning method with Deep Gaussian Processes offering a probabilistic deep learning is proposed to analyze the numerical prediction model. Then, an optimizer based on a geometric transformation with homothety has been defined to perform the transport of the simulated data to the measured data. The proposed method is tested on the measurements of the global radiation parameter. A satisfactory result is obtained with a reduction of the error margin of about 98%. This means that the approach is applicable to other parameters to make the numerical forecast model more reliable.

Automated summary

This paper proposes a learning method using Deep Gaussian Processes for probabilistic deep learning to analyze numerical prediction models, and an optimizer based on geometric transformation to transform simulated data to measured data. The method is tested on measurements of global radiation parameter and achieves satisfactory results with a significant reduction in error margin.