Growth in computational performance, the amount of accumulated data about the environment and experience of handling such amounts of data leads to an increase in the number of applications of data analysis. Weather forecasting is one of these areas. Weather forecasting uses a variety of data and meteorological models describing the physical processes in the atmosphere. Machine learning algorithms can correct some errors of these models and improve weather forecasts. To improve the temperature forecast, we added to the training data for different models readings from the nearest meteorological stations. This technique proved to be useful for the short-term temperature prediction. We evaluated the results of experiments by the changing of the root-mean-square error. Yandex.Weather service successfully uses the improved model for forecasting. The described technique has already been added to the Yandex.Weather and used on an ongoing basis in production. In addition, we suggest possible directions of this task development.