Optimization of hyperparameters of neural networks using high-performance computing for prediction of precipitation

The paper describes the procedure for tuning hyperparameters of the neural network for analyzing spatial meteorological data using the tools of the hybrid high-performance computing system. The comparison of precipitation forecasting accuracy has been carried out on the basis of such methods as grid and random searches. It has been demonstrated that even with a relatively small number of random choices of combinations of hyperparameters, it is possible to obtain an accuracy comparable to brute force, with moderate time costs. These results show the ability to automatically build a neural network architecture based on the general model for solving applied problems.