Devising a forward propagation artificial neural network application technology for nowcasting weather elements

The object of this paper is the procedure of applying a forward-propagation artificial neural network of surface learning for the purpose of short-term forecasting one of the weather elements – the temperature of the near-surface air layer. Known methods to forecast weather elements, namely, hydrodynamic, physical-statistical, and synoptic, have been successfully supplemented in recent years by forecasting using artificial neural networks. It has become possible to build large networks for a large amount of training data for deep learning. However, the level of development of the theory of artificial neural networks does not make it possible to build the required network. Therefore, when solving applied tasks such as the one reported in this paper, the developer has to build a forecasting system blindly or based on some heuristic considerations, experimenting with neural networks. At the same time, the path of network complexity often does not lead to a qualitative improvement in forecasting results. Therefore, during the research, the main problem to be solved was to optimize the use of a simple neural network with a well-developed training algorithm for the purpose of nowcasting meteorological elements. The optimization criterion adopted was the surface air temperature short-term forecast accuracy at different time intervals. The parameters enabling the achievement of optimality are the parameters of the data that train the network and the parameters of the network itself. By selecting these parameters, a high accuracy rate for short-term forecasts of different timeliness has been achieved. The accuracy of a three-hour forecast reaches 100 percent. The same value is achieved for the forecast accuracy with a one-day lead time. Predicting temperature values for three days has an accuracy rate exceeding 90 percent