Numerical Weather Prediction Postprocessing with Convolutional Neural Network

Numerical weather prediction (NWP) is one of the main ways to provide weather forecasting, while its precision still needs to be improved due to the incomplete theoretical understanding and computational complexity as well. Good news is that, the error of NWP usually shows certain spatial and temporal pattern, which gives us a chance to capture the error distribution rule using historical forecast and its corresponding observations. In this paper, we use the Convolution Neural Network (CNN) method, which has been proved to be very effective in the data mining of spatiotemporal-type data, to derive the error distribution rule of NWP and correct its forecasting. The training data are from ECMWF high resolution10-day forecast and the High-Resolution China Meteorological Administration Land Data Assimilation System (HRCLDAS) respectively in October, 2017. The elements used for CNN training are air temperature, pressure, humidity, and wind. Corrected forecasting in November, 2017 is used for validation. Results show that, compared with the original prediction, the accuracy of corrected prediction increases in general for each element, the mean absolute errors within 48-hour forecasting decrease 15.6%, 32.1%, 15.9%, and 27.4% respectively for air temperature, pressure, humidity, and 10-meter wind speed. It indicates that the CNN method can be used for improving NWP performance, thus provide us an effective way to improve the accuracy of weather forecasting.