【作者】 刘杰；

【导师】 张彤；

【作者基本信息】 武汉大学 ， 地图制图学与地理信息工程， 2023， 博士

【摘要】 暴雨不仅危害农作物生长与收获,还有可能引发山洪暴发、河水泛滥、城市内涝等洪涝灾害以及山体滑坡、泥石流等地质灾害,给工农业生产、人民生命财产和城市发展带来巨大威胁。提高暴雨预测水平,尤其是提高暴雨预测的准确性,能够为防灾减灾提供风险信息与决策性建议,提高暴雨防灾减灾能力,对于保障社会经济发展与人民财产安全具有十分重要的意义。相较于传统数值方法,机器学习驱动的暴雨预测方法具有推理速度快、计算成本低、建模简单等优点。但由于大气混沌特性和时空异质性,机器学习驱动的暴雨预测面临着以下问题:(1)由于低发生率,暴雨历史样本数量稀少有限。基于有限的训练样本,很难训练鲁棒的预测模型来概括可靠的模式做出准确的预测。(2)在空间或时间上接近的暴雨之间存在复杂的时空相关性,在空间或时间上不接近的历史暴雨之间存在复杂的相似性。与其他时空事件或现象(如城市交通模式)相比,暴雨模式在空间和时间上更加难以捉摸。(3)暴雨的发生涉及许多大气因素,如温度、压力、湿度和风等,这些因素在不同地点和时间对暴雨发生及强度的影响是非线性变化的,概括这些因素和暴雨之间的动态因果关系是十分困难的。(4)现有暴雨预测数据驱动方法没有充分考虑暴雨相关的先验知识,可能无法产生准确且符合物理规律的预测结果。针对暴雨预测所面临的暴雨数据有限、事件时空相关性复杂以及大气因素关联复杂等问题,本文在先验门控卷积网络中显式集成了流体个别变化、时空模式等先验知识,提高深度学习驱动的暴雨预测模型的泛化能力和建模暴雨时空动态的能力。本文的主要研究内容如下:(1)流体个别变化指导的时空模式先验表征方法。为了充分利用时空模式先验缓解数据有限问题,提高模型的可靠性,本文设计了流体个别变化指导的门控卷积编解码网络,表征大气时空动态特征为时空事件表征,并将事件表征作为时空模式先验用于指导暴雨过程预测。考虑到暴雨多种影响因素的复杂时空非线性关系,本文根据描述大气属性时空动态的流体个别变化设计了隐藏状态物理更新门控,在潜在空间对隐藏状态添加物理约束,以符合大气运动物理规律的方式表征大气属性的时空动态特征。为了减少表征过程中的信息损失,本文提出了多层次重建损失,在编解码网络由浅到深的各个层次上保持大气的时空及属性特征。实验结果表明相较于3种先进深度学习驱动的表征方法,所提表征方法能够提供更有利于暴雨预测的表征结果。(2)时空模式先验感知的暴雨过程预测方法。考虑到时间空间不接近的暴雨之间的相似性,本文提出属性时空相似度来度量历史事件与当前事件在大气属性上的时空相似性。基于属性时空相似度矩阵,本文以时空图卷积的方式,将历史事件表征聚合成时空模式先验。然后,本文设计了时空先验激活层,通过先验激活门控将时空模式先验与当前的事件表征相结合,输出时空先验激活表征,为暴雨过程预测提供有用的信息。此外,为了使预测结果更加符合暴雨实际形成、发展和消散等规律,本文提出了以流体个别变化作为软约束的时空一致性损失。实验结果表明相较于4种先进深度学习驱动的时空预测方法,所提预测方法能够提供更好的暴雨预测性能。(3)基于卡尔曼门控优化网络的暴雨集合预测方法。为了尽可能降低大气运动的混沌特性、不完美的模型设计和优化等不确定性因素带来的预报误差,本文采用集合预测的方式改进单一的确定性预测。本文设计了时空模式先验分布扰动方案,从所表征的时空模式先验中学习噪音分布,为事件表征添加从所学分布中获取的扰动,并基于所提预测模型生成合理的集合成员。考虑到暴雨集合预测的状态空间方程的特点以及暴雨事件之间存在的复杂时空关联,本文提出了卡尔曼门控优化网络,根据由估计误差和预测误差计算的卡尔曼增益门控优化集合成员预测结果,实现准确的暴雨集合预测。实验结果表明所提集合预测方法能够提供比所比较的4种集合预测方法以及所提暴雨预测方法更好的暴雨预测性能。本文在深度学习建模非线性时空动态的强大能力的基础上,开展了先验门控卷积网络支持的暴雨时空过程预测研究,从“先验表征—暴雨预测—集合预测”三个方面集成流体个别变化、时空模式先验、卡尔曼滤波器等知识,提出了流体个别变化指导的事件时空模式先验表征、时空模式先验感知的暴雨过程预测和基于卡尔曼门控优化网络的暴雨集合预测方法。实验证明,本文所提方法在两个现实世界数据集上实现了准确的暴雨过程预测。本文所提方法能够提高暴雨预报的自动化水平,对工农业发展、交通运输与防灾减灾等实际应用具有重要意义。更多还原

【Abstract】 Rainstorm is not only harmful to the growth and harvest of crops,but also may trigger flooding disasters such as flash floods,river flooding,urban waterlogging,and geological disasters such as landslides and mudslides,posing a huge threat to industrial and agricultural production,people’s lives and properties,and urban development.Improving the rainstorm prediction skill,especially the accuracy of rainstorm prediction,can provide rainstorm risk information and decision-making suggestions for disaster prevention and mitigation,and improve the ability of rainstorm disaster prevention and mitigation,which is of great significance in safeguarding social&economic development and the safety of people’s property.Compared with traditional numerical methods,machine learning driven rainstorm prediction methods have the advantages of fast inference,low computational cost and simple modeling.However,due to the chaotic nature and spatiotemporal heterogeneity of the atmosphere,machine learning driven rainstorm prediction faces the following problems:(1)Due to the low occurrence rate,the number of historical rainstorm samples is sparse and limited.Based on limited training samples,it is difficult to train robust prediction models to generalize reliable patterns to make accurate predictions.(2)There are complex spatio-temporal correlations between rainstorms that are close in space or time,and complex similarities between historical rainstorms that are not close in space or time.Rainstorm patterns are more elusive and heterogeneous in space and time than other spatio-temporal events or phenomena(such as urban traffic patterns).(3)The occurrence of rainstorms involves many atmospheric factors,such as temperature,pressure,humidity,wind,etc.These factors have nonlinear effects on the occurrence and intensity of rainstorms at different locations and times,making it difficult to generalize the dynamic causal relationships between these factors and rainstorms.(4)Existing data driven rainstorm prediction methods do not fully consider the prior knowledge of rainstorms,and may not make accurate and physically consistent predictions.To address the problems of limited rainstorm data,complex spatiotemporal correlations of events and correlations of atmospheric factors faced by rainstorm prediction,we explicitly integrate prior knowledge,such as substantial derivative and spatiotemporal patterns of events,in the gated convolutional networks to improve the generalization ability and the rainstorm spatiotemporal dynamics modeling ability of the deep learning driven rainstorm prediction.The main research contents are summarized as follows:(1)Substantial derivative guided representation of spatio-temporal pattern prior.In order to make full use of the spatio-temporal pattern prior to alleviate the problem of limited data and improve the reliability of the model,a substantial derivative guided gated convolutional encoder-decoder network is proposed to represent the spatio-temporal dynamics of the atmosphere as spatio-temporal event representations,which can be used as spatio-temporal pattern prior to guide the rainstorm process prediction.Considering the complex spatio-temporal non-linear relationships of multiple influencing factors of rainstorms,a physical update gate of the hidden state is designed according to the substantial derivative describing the spatio-temporal dynamics of atmospheric attributes.Physical constraints are added to the hidden states in the potential space to represent the spatio-temporal dynamics of atmospheric attributes in a manner consistent with the physical laws of atmospheric motion.In order to reduce the information loss in the representation process,a multi-level reconstruction loss is proposed to maintain the spatio-temporal and attribute features of the atmosphere at each level of the encoder-decoder network from shallow to deep.Experimental results show that the proposed representation method is able to provide representation results that are more beneficial to rainstorm prediction compared with three advanced deep learning driven representation methods.(2)Spatio-temporal pattern prior informed rainstorm process prediction.Considering the similarity between rainstorms that are not close in time or space,the attributed spatio-temporal affinity is proposed to measure the spatio-temporal affinity of the atmospheric attributes of historical events and current event.Based on the attributed spatio-temporal affinity matrix,the historical event representations are aggregated into spatio-temporal pattern prior by means of spatio-temporal graph convolution.Then,a spatio-temporal prior activation layer is developed to combine the spatio-temporal pattern prior with the current event representation through prior activation gates to output the spatio-temporal prior activation representation,which can provide useful information for rainstorm process prediction.In addition,in order to make the prediction results more consistent with the actual formation,development and dissipation patterns of rainstorms,a spatio-temporal coherence loss is proposed with the substantial derivative as a soft constraint.Experimental results show that the proposed rainstorm prediction method can provide better rainstorm prediction performance compared with four advanced deep learning driven spatio-temporal prediction methods.(3)Kalman gated optimization network based rainstorm ensemble prediction.In order to minimize prediction errors due to uncertainties in the chaotic nature of atmospheric motion and imperfect model design&optimization,an ensemble prediction method is adopted to improve deterministic rainstorm predictions.A spatio-temporal pattern prior distribution based representation perturbation scheme is proposed to learn the noise distribution from the represented spatio-temporal pattern priors,and add perturbations obtained from the learned prior distribution to the event representation.Then reasonable ensemble members can be generated based on the proposed prior perturbation scheme and rainstorm prediction model.Considering the characteristics of the state-space equations for rainstorm ensemble prediction and the complex spatio-temporal correlations between rainstorm events,a Kalman gated optimization network is developed to optimize ensemble member prediction results to achieve accurate rainstorm ensemble prediction based on the Kalman gain calculated from the estimation and prediction errors.Experimental results show that the proposed ensemble prediction method can provide better rainstorm prediction performance than compared four ensemble prediction methods and the proposed rainstorm prediction method.Based on the powerful ability of deep learning for modeling non-linear spatio-temporal dynamics,prior gated convolutional networks are developed.Prior knowledge of substantial derivative,the spatio-temporal patterns and Kalman filter is integrated into processes of‘prior representation-rainstorm prediction-ensemble prediction’,developing methods of substantial derivative guided representation of spatio-temporal pattern prior,spatio-temporal pattern prior informed rainstorm process prediction,and Kalman gated optimization network based rainstorm ensemble prediction.Experimental results show that the proposed method achieved accurate rainstorm prediction on two real-world data.The proposed methods can improve the automation level of rainstorm prediction and are of great significance for practical applications such as industrial&agricultural development,transportation,and disaster prevention&mitigation.更多还原