毕业生知识库

由于大气运动的时空复杂性和相关性，气象预测是一项极具挑战性的任务。
过去几十年，传统的基于物理模型的数值预报方法不断完善，预报质量得到显著
提升。但是，技术上，目前数值预报方法已经发展到了极致阶段。随着气象观测
技术的快速发展，气象行业积累了海量的气象大数据，为构建新型的数据驱动
的气象预测模型提供了机遇。与此同时，深度学习方法在视觉理解、自然语言处
理、语音识别和时间序列预测等多类型任务中得到成功应用。目前，许多研究人
员试图将深度学习方法引入至气象预测领域之中。基于深度学习的气象预测逐
渐成为一个热门的研究方向。然而，由于气象数据所存在的长时依赖关系和大范
围空间关联关系，加之多模态气象要素间复杂的跨模态耦合关系，基于深度学习
的气象预测是一个具有挑战性的研究课题。直接应用现有深度学习模型难以进
一步得到性能提升。比如，气象模式所存在的局部时空差异往往并不满足经典的
卷积循环神经网络中的归纳偏置（即有关时空共享的卷积核的假定）。为此，基
于多模态气象要素数据，在卷积循环神经网络和 Transformer 的框架下，本文开
展面向气象预测的深层神经网络模型构建、模型训练和模型验证工作，通过改进
现有模型来提升预测精度。
本文主要工作及贡献如下：
1. 提出一种新型的卷积循环神经单元——时空自适应卷积门控循环单元。
该模块将时空自适应卷积操作嵌入至门控循环单元之中。其核心思想是，在更新
卷积循环单元的内部状态时，引入额外的卷积层来同时学习卷积核的采样位置
和权重。因此，内部的自适应卷积可以根据时空信息自适应地选卷积计算位置并
调整权重，以满足气象数据局部的变化模式。在等气压层温度、相对湿度、风速
以及雷达反射回波等多个气象数据集上的进行了对比实验，实验结果证明了所
提模型的有效性和优越性。
2. 提出一种基于 Transformer 的气象预测深度学习模型。具体地，首先，提
出利用门控机制来实现等气压层温度、相对湿度、风速等多个模态的加权融合；
其次，引入 Transformer 机制，提出以并行时空轴向注意力代替传统的注意力机
制，从而有效地学习长时依赖关系和大范围空间关联关系。整体结构上，本文采
用了基于 Transformer 编码器-解码器框架。在 ERA5 再分析数据集（子区域）上进行了对比实验，实验结果表明了所提方法在温度、相对湿度、风速等预测任务上的有效性和优越性。

Due to the spatio-temporal complexity and correlation of atmospheric motion, meteorological forecasting is an extremely challenging task. Over the past few decades, with the continuous development of traditional numerical prediction methods based on the physical model, the forecasting quality has been significantly improved. But, technically, the numerical prediction methods have reached the extreme stage. Thanks to
the rapid development of meteorological observation technology, the meteorological industry has accumulated massive meteorological data, which provides an opportunity to build new data-driven meteorological forecasting methods. Meanwhile, the deep learning methods have been successfully applied in various areas, such as visual understanding, natural language processing, speech recognition, and time series prediction,
and so on. In the past years, many researchers have tried to introduce deep learning methods into meteorological forecasting. Nowadays, meteorological forecasting based on deep learning has gradually become a popular research issue. However, due to the long-term dependence and large-scale spatial correlation hidden in meteorological data,
and additionally due to the complex coupling relationship between different modalities, Actually, it is difficult to further improve the performance by directly applying the existing deep learning models. For example, the spatio-temporal variance of local meteorological patterns does not meet the inductive bias of typical convolutional recurrent neural network (namely, the hypothesis about spatio-temporal sharing kernels).
Therefore, based on the multi-modal meteorological data, under the framework of convolutional recurrent neural network and Transformer, this thesis carries out the model construction, model training and model validation of deep neural networks to improve the prediction accuracy for meteorological forecasting.
The main contributions of this thesis are described as follows:
1. We proposed a new convolutional recurrent neural unit, namely, Spatiol-Temporal Adaptive Convolutional Gated Recurrent Unit (STAConvGRU). Technically, the spatio-temporal adaptive convolution operation is embedded into the Gated gated recurrent unit. The key motivation behind STAConvGRU is to develop additional convolution
layers to learn simultaneously the sampling positions and weights of convolutional kernels, when updating the internal state of the convolution recurrent unit. As a result, the adaptive convolution could select the positions and adjust the weights according to the spatio-temporal information to the meteorological, which matches well the local meteorological transformation patterns. Comparative experiments are conducted on four types of meteorological datasets, including temperature, isobaric relative humidity, wind, and radar echo. The experimental results demonstrate the effectiveness and superiority of our proposed model.
2. We proposed a deep learning model based on Transformer for meteorological forecasting. Specifically, gating mechanism is developed to achieve the multi-mode fusion of isobaric temperature, relative humidity and wind speed; Transformer mechanism is implemented by spatio-temporal axial attention to effectively learn long-term dependence and large-scale spatial correlation. Architecturally, the Transformer encoder-decoder framework is employed as the overall framework. Extensive comparative experiments have been conducted on the regional ERA5 reanalysis dataset, demonstrating that the proposed model method is effective and superior in the prediction of temperature, relative humidity and wind.