Рекурентні нейронні мережі для задачі уточнення чисельних метеорологічних прогнозів

Abstract

This paper briefly describes examples of how deep learning can be applied to geoscientific problems, as well as the main difficulties that arise when scientists apply this technique to the problems of meteorological forecasting. This paper aims at comparing the two most popular types of recurrent neural network architectures, namely the long short-term memory network and the gated recurrent unit when they are used to improve 2m temperature forecast results obtained using numerical hydrodynamic methods of meteorological forecasting. An efficiency comparison of architectures of recurrent neural networks was performed using the root-mean-square error. It is shown that all models with gated recurrent units are more efficient than models with long short-term memory. Thus the best architecture of recurrent neural networks for solving the problem of improving numerical meteorological forecasts has been revealed.