深度卷积神经网络对陕西地区气温预报误差订正效果研究

Correction Effect of Deep Convolutional Neural Network on Temperature Forecast Error in Shaanxi Province

当前气象预报工作中,数值模式预报处于核心地位,气象预报结果对其依赖性很强,但其预报结果存在一定误差。为了降低误差,本文提出了一种基于卷积神经网络的深度卷积神经网络(DCNN)订正方法,并基于欧洲中心的2 m温度预报资料和再分析资料在陕西地区进行了预报订正。通过对订正结果的分析表明,DCNN订正方法能有效提高预报准确率,预报误差越大订正效果越明显;其订正效果在一定范围内随着参数epoch的增大而提高;其前期的订正效果要明显优于后期。

In current meteorological forecasting work, numerical model forecasting is at the core, and the results of meteorological forecasting are highly dependent on it, but there are certain errors in its forecasting results. In order to decrease errors, the paper proposes a deep convolutional neural network (DCNN) correction method based on convolutional neural networks, and conducts prediction correction in Shaanxi region based on the 2 m temperature prediction data and reanalysis data from the ECMWF. The analysis of the correction outcome shows that the DCNN correction method can significantly improve the prediction accuracy, and the larger the prediction error, the more obvious the correction effect;the correction effect increases with the increase of parameter epoch within a certain range;the correction effect in the early stage is significantly better than that in the later stage.