卷积神经网络在ENSO预报中的应用

Application of Convolutional Neural Network in ENSO Prediction

为提高对ENSO的预报能力,同时针对用机器学习方法做气候预报时观测资料不足的问题,基于深度卷积神经网络(convolutional neural networks,CNN)架构,以CMIP6模式资料和GODAS观测资料为数据集,训练出一个应用于ENSO预报的神经网络。结果表明,在训练神经网络时引入CMIP6模式资料能提高数据量,解决了机器学习中观测资料不足的问题。在时效为1~9个月的后报实验中,神经网络的表现优于传统的动力模式和统计模式。对照实验显示模式数据的加入以及采用集合预报的方法有利于改善预报效果,热含量数据的加入则表现出负面效果。对后报实验的结果分析显示,神经网络的预报准确度存在年内和年际变化,其中年内变化与ENSO预报中普遍存在的春季预报障碍有关。实验结果显示卷积神经网络在ENSO预报中的有效性。

In order to improve the forecasting of ENSO and to address the problem of insufficient observational data when using machine learning methods for climate forecasting,a neural network based on a deep convolutional neural network(CNN) architecture was trained for ENSO forecasting using CMIP6 model data and GODAS observations as the dataset.The results show that the introduction of CMIP6 model data in training the neural network can improve the amount of data and solve the problem of insufficient observation information in machine learning.In the hindcast experiment with a lead time of lto 9 months,the performance of the network is better than the traditional dynamic models and statistical models.Control experiments show that the introduction of model data and the use of ensemble forecast are conducive to improving the prediction effect,while the addition of heat content data shows a negative effect.The analysis of the results of the hindcast experiment shows that there are annual and interannual variation in the accuracy,and the interannual variation is related to the spring predictability barrier which is generally present in ENSO forecast.The results of the experiments show the effectiveness of convolutional neural networks in ENSO forecasting.