【作者】 李慧；

【导师】 任鹏；

【作者基本信息】 中国石油大学(华东) ， 信息与通信工程， 2020， 硕士

【副题名】以风场预测为例

【摘要】 当前,气象要素的预测多采用数值模式的方法。此类方法预测结果的准确性依赖于构建的模型的可靠性。但是,气象要素复杂多变,影响因素众多,难以全面、有效地对其进行建模。因此,数值模式预测方法的结果存在一定的误差。另外,数值模式的结果形式单一,仅能定量表示预测结果,缺乏直观的表达。生成对抗网络近年来被广泛应用在预测任务并表现突出。生成对抗网络可以学习数据在时间维度上的发展模式,刻画相邻时刻的数据之间的内在联系。本文提出利用生成对抗网络进行气象要素场预测,对当前的数值模式的预测方法进行补充和提升。本文以风场预测为例,进行了两项任务,第一项为基于卫星云图的台风云系预测,第二项为基于数值模式预测结果的全球风场预测。针对台风云系预测,本文构建了基于多层次生成对抗网络的台风云系预测模型。模型以历史时段台风卫星云图数据为输入,输出结果为未来时段台风云图。模型包括全局生成器和局部判别器两个网络。全局生成器根据历史时段的实测台风云图序列,学习台风云系的运动规律,并预测未来时段的台风云图。局部判别器网络对输入其中的生成未来云图和真实未来云图判别真假。两个网络在对抗训练中不断促进彼此的表征能力。最终,全局生成器生成的未来台风云图无法被局部判别器判别出真假。业务实现时,训练好的全局生成器可用作预测未来时段台风云图。基于真实的卫星云图数据的实验证明:本模型有效地表示台风云系的“时空”演变过程,预测未来时段台风云图,为当前的台风预测提供了全新参考和直观的结果展示,对台风的防灾减灾工作具有重要的参考价值。针对全球风场预测,本文构建了基于生成对抗网络的全球风场预测模型。数值模式的风场预测结果存在一定误差,根据部分地点的实测数据构建的再分析数据的准确度高于预测数据。本文构建的预测模型学习预测数据与其对应的再分析数据之间的内在联系,对预测数据进行校正,获得生成的“再分析”数据。模型的基础结构为生成对抗网络。生成器以数值模式的预测结果为输入,输出接近于再分析数据的生成“再分析”数据。判别器的输入为实际的再分析数据和生成的“再分析”数据,判别器对二者的来源进行区分。业务实现时,运用训练好的生成器对数值模式的预测结果进行校正,获得更高准确度的风场预测结果。本文通过对来自欧洲中期天气预报中心的全球风场预报数据和再分析数据进行实验,模型生成的“再分析”数据比预测数据更接近于真实的再分析数据,准确度更高。综上,针对气象要素预测任务中的台风云系预测,本文提出了基于多层次生成对抗网络的台风云系预测方法,是对当前预测方法结果的可视化补充。针对全球风场预测任务,本文提出了基于生成对抗网络的全球风场预测模型,提升数值模式预测结果准确度。更多还原

【Abstract】 At present,the prediction of meteorological elements mostly adopts the method of numerical models.The accuracy of the results of such methods depends on the reliability of the model.However,the meteorological elements are complex and changeable,with many influencing factors,making it difficult to model them comprehensively and effectively.Therefore,there is a certain error in the results of the numerical model prediction method.In addition,the result format of the numerical model is single and lacks intuitive expression,which can only quantitatively represent the prediction result.Generative adversarial networks have been widely used in prediction tasks in recent years with outstanding performance.Generative adversarial networks can learn the development pattern of data in the time dimension and characterize the inherent relationship between data at adjacent moments.We propose to use the generative adversarial networks to predict the meteorological element field,supplementing and improving the current prediction methods of numerical models.We use wind field prediction as an example to perform two tasks,the first one is the prediction of typhoon clouds based on satellite cloud image,and the second one is the global wind field prediction based on the prediction results of numerical model.Aiming at typhoon clouds prediction,we build a typhoon clouds prediction model based on the hierarchical generative adversarial networks.The model takes typhoon satellite cloud image in the historical period as input,and the output result is the typhoon cloud image in the future period.The model includes two networks,a global generator and a local discriminator.The global generator learns the movement rules of the typhoon clouds based on the measured typhoon cloud image sequence in the historical period,and predicts the typhoon cloud images in the future period.The local discriminator network discriminates whether the generated future cloud image and the real future cloud image are true or false.The two networks continuously promote each other’s representation capabilities in adversarial training.Finally,the generated future typhoon cloud images could not be discriminated by the local discriminator.During commercial operation,the trained global generator can be used to predict the typhoon cloud images in the future.The typhoon clouds prediction model learns the evolution of the typhoon clouds from the typhoon cloud image sequence,and generates the typhoon cloud images at the future moment,which is the visual prediction result of the typhoon clouds.Experiments based on real typhoon satellite cloud image data prove that this model effectively represents the“temporal and spatial” evolution of the typhoon clouds and predicts typhoon cloud images in the future and provide a new reference and intuitive results display for the current typhoon prediction,which has important reference value for typhoon disaster prevention and reduction.Aiming at the global wind field prediction,we build a global wind field prediction model based on the generative adversarial network.There is a certain error in the wind field prediction result of the numerical model.There is a certain error in the wind field prediction result of the numerical model.The accuracy of the reanalysis data constructed based on the measured data of some locations is higher than the predicted data.The global wind field prediction model we constructed learns the inherent relationship between the prediction data and its corresponding reanalysis data,corrects the prediction data,and obtains the generated "reanalysis" data.The basic structure of the model is a generative adversarial network.The generator takes the prediction result of the numerical model as an input,learns the relationship between the prediction result and the reanalysis data and outputs the generated "reanalysis" data as close as possible to the reanalysis data.The input of the discriminator is the true reanalysis data and the generated "reanalysis" data.The discriminator distinguishes between the two sources.In commercial operation,the trained generator can be used to correct the prediction results of the numerical model to obtain more accurate wind field prediction results.Through experiments on global wind field prediction data and reanalysis data from the European Center of Medium Weather,the "reanalysis" data generated by the model is closer to the real reanalysis data than the predicted data and has higher accuracy.In summary,for the prediction of typhoon clouds in the prediction task of meteorological elements,we propose a typhoon clouds prediction method based on hierarchical generation adversarial network,which is a visual supplement to the results of the current prediction methods.For the global wind field prediction task,we propose a prediction method based on the generative adversarial network model,which improves the accuracy of the numerical model prediction results.更多还原