Abstract: Base on deep gated recurrent unit （GRU） neural network, an hourly prediction model of air temperature in Shimen county are proposed by using hourly surface meteorological data （air temperature, air pressure, dew point temperature, relative humidity and water vapor pressure）, besides the accuracy of the model is analyzed. As the result indicated, the accuracy of the deep GRU model gradually declines with the prediction horizons, whose coefficients of determination, mean absolute error, mean values of root square error and the prediction accuracy are 0.996~0.906, 0.359~1.974, 0.510~2.562 ℃ and 62.529% respectively, which has higher performance than the prediction model based on autoregressive integrated moving average （ARIMA）, and reflects turning changes of air temperature as well. The hourly air temperature prediction accuracies in Shimen from April to July in 2019 are contrasted of the deep GRU model and numerical weather predictions （NWPs） from European Center in addition to Japan. The result shows that the deep GRU model would be a beneficial supplement to the air temperature short-impending prediction within 12 hours due to its prediction accuracies are all better than NWPs in 3, 6 and 9 h horizons.

Key words: deep learning, GRU, hourly air temperature prediction, Shimen area