Deep Learning for Improving Numerical Weather Prediction of Daily Maximum Air Temperature

Objective forecast of maximum temperature is an important part in numerical weather prediction (NWP). Due to the influence of complex factors such as atmospheric dynamic processes, physical processes and local topography and geomorphology, the prediction of near-surface meteorological elements in the numerical weather model often has deviation. In recent years, on the one hand, meteorological observations expand rapidly, making traditional error correct method difficult to deal with the massive data. On the other hand, artificial intelligence has an increasingly obvious advantage in processing big data.   

In view of this, here we correct the European Centre for Medium-Range Weather Forecasts (ECMWF) Integrated Forecast System (IFS) both general statistics and local events, by post-processing its maximum temperature output with a deep neural network. Based on the fully connected neural network, four sensitivity experiments are designed in order to investigate the importance of auxiliary variable, time-lagged variable and the effectiveness of embedding layer in the neural network. The observations of basic meteorological elements of totally 2238 basic weather stations and the output of NWP during 15 January 2015 to 31 December 2020 are employed. The training period is from 15 January 2015 to 31 December 2019, and the rest part is test period. The results show that the forecast error of daily maximum air temperature from the IFS in test period is reduced greatly by the sensitivity experiments, which add auxiliary variables, daily maximum air temperature with 1-2 lag days and embedding layer structures and their combination. The root mean square error is reduced by 29.72%-47.82% and the accuracy of temperature forecast are increased by 16.67%-38.89%, and the effects for Qinghai-Tibet Plateau is especially remarkable where the forecast error of IFS model is very high. It is preliminarily proved that the fully connected neural network with embedding layer has better overall performance than the raw fully connected neural network, and the features also affect the forecast errors and forecast skills of the model. Besides, the prediction error of neural network model with embedding layer is more stable when auxiliary variables and lag time variables are added. Positive forecasting techniques are available for almost all stations in the study, and it is possible to reduce the mean absolute error to less than 1℃ at many stations. Tests of 1-year daily maximum temperature forecasts at four in situ stations show that FCNN forecasts with embedded layers are closest to observations, both for the whole year and for extreme points.