On the matter of choosing a neural network for the task of short-term forecasting based on the results of microwave radiometric sounding of the atmosphere

Abstract

The article studies the matters of choosing a neural network and its operation efficiency evaluation for the short-term forecasting task of weather parameters in the ground-level air further to the results of microwave radiometric sounding of the atmosphere. The following tasks and sequence for neural network forecasting of weather parameters based on the results of multi-frequency microwave radiometric sounding of the atmosphere: analysis and measurement data preprocessing by a multi-range microwave radiometric system for the purpose of retrieving informative features about the atmospheric status; forecasting the dynamics pattern in key weather parameters: air temperature and humidity, atmospheric pressure, clouds, rain situation; predicted result verification based on data from extra observation sources for the adaptive timing purpose of neural-network model parameters; calculation of quality and reliability indicators of weather parameters prediction to assess the proposed approach efficiency. There are examined three options of neural-network development programs that are out in the open. There is performed training and testing of neural networks for forecasting weather parameters of the ground- layer air via output signals of eleven channels in the quad-band microwave radiometric system that provides measurements of thermal radio radiation in the ground- layer air in four bands with center frequencies of 4 GHz, 10 GHz, 22 GHz and 37 GHz and data from the weather station located in Murom town. There was proposed and accomplished teaching sample creation   for a neural network with data perturbation to exclude the influence of temporal trend variations in weather parameters. There are obtained numerical estimates for the neural-network operation efficiency for forecasting weather parameters - correlation factor and mean deviation. The conclusions on the comparative analysis of neural-network operation are made. The attained results manifested different problem solving efficiency of forecasting weather parameters based on microwave radiometric measurements, which enabled to conclude that it is necessary to solve the structural optimization task for the utilized neural network.