CASIA OpenIR  > 毕业生  > 硕士学位论文
复杂时空数据预测方法研究
许宝文
2024-05-16
Pages95
Subtype硕士
Abstract

随着传感器等数据采集技术的不断进步,产生了海量且高质量的时空数据,迫切需要探索有效的时空数据建模方法。面向复杂时空数据的预测任务是时空数据挖掘领域的核心研究任务之一,在气象、交通、灾害评估、城市规划以及物流等领域得到了广泛应用。传统统计模型和机器学习模型在复杂时空相关性建模方面存在局限性,难以有效捕捉时空数据中的关键模式,如相关性、周期性和异质性等,导致它们的预测结果无法达到实际应用的准确性要求,而深层神经网络模型凭借其强大的通用函数逼近能力,能够有效地对复杂的时空数据进行建模。因此,本文针对具有显式空间信息的时空网格、时空图和时空轨迹数据和具有隐式空间信息的多元时间序列这四类时空数据提出了一系列基于深层神经网络模型的创新性预测模型,并在相应的时空数据类型下的典型应用进行了验证,证明了提出模型的有效性。主要的研究工作和贡献归纳如下:
1、针对空间拓扑结构是静态且均匀分布的时空网格数据的建模问题,提出了两种创新的U 型网络模型:分层U-Net(HU-Net)和注意力脉冲U-Net(ASUNet)。HU-Net 通过其分层设计,旨在充分挖掘时空网格数据中的全局与局部空间相关性,并采用了重参数技巧,以提升模型在推断时的速度。而ASU-Net,是融合了注意力机制的脉冲U 型网络,旨在降低人工神经网络在训练和推断过程中的高能耗。通过在天气预测应用上的实验验证,HU-Net 模型显著提高了预测的准确率;ASU-Net 模型则实现了在较低能耗下保持高预测精度的目标。
2、针对空间拓扑结构是静态且非均匀分布的时空图数据的建模问题,提出了两种不同的建模方法:基于图卷积的双通道小波变换神经网络(DSTWave)和基于Transformer 的时空Transformer(ST-Transformer)。DSTWave 模型由解纠缠长短期时间模式的小波变换网络和堆叠的双通道的时空图卷积组成。ST-Transformer模型引入了空间Transformer 和时间Transformer,分别用于建模全局空间依赖性和长期时间依赖性。通过在交通预测应用上的实验验证,DSTWave和ST-Transformer 能够达到最优的预测效果,从而证明了所提出方法的有效性。
3、针对空间拓扑结构是动态且非均匀分布的时空轨迹数据的建模问题,提出了基于条件变分自编码器的模型Social-CVAE。Social-CVAE 的核心是一种条件变分自动编码器架构,该架构通过以观察到的过去轨迹为条件,利用随机潜在变量来学习智能体的未来轨迹的多模态分布。通过在行人轨迹预测应用上的实验验证,提出的Social-CVAE 实现了最佳的结果,验证了该算法的有效性。
4、针对具有隐式空间信息的多元时间序列的长期建模问题,提出了非平稳条件变分自编码器(N-CVAE)和傅里叶U 型网络(F-UNet)。N-CVAE 通过引入了
噪声和设计对时间序列数据的非平稳性建模的组件,使模型能够更好地适应不同
的输入分布,缓解分布漂移问题,以实现更准确的长期时间序列预测。F-UNet 由低的时间复杂度的神经网络组件组成,以期望超越目前最优的基Transformer的高时间复杂度的长期时序预测模型。通过在真实的工业现场和多个公开基准
多元时间序列数据集上的实验验证,F-UNet 能够以低时间复杂度、高计算效率
优于目前基于Transformer 的模型,N-CVE 也取得了最佳的预测结果,证明了所
提出模型的有效性。

Other Abstract

With the continuous advancement of data collection technologies such as sensors, a massive and high-quality amount of spatio-temporal data has been generated, and there is an urgent need to explore effective methods for modeling spatio-temporal data. The prediction task for complex spatio-temporal data is one of the core research tasks in the field of spatio-temporal data mining, and has been widely applied in fields such as meteorology, transportation, disaster assessment, urban planning, and logistics. Traditional statistical models and machine learning models have limitations in modeling complex spatio-temporal correlations, making it difficult to effectively capture key patterns in spatio-temporal data, such as correlation, periodicity, and heterogeneity. As a result, their prediction results cannot meet the accuracy requirements of practical applications.
However, deep neural network models, with their powerful universal function approximation ability, can effectively model complex spatio-temporal data. Therefore, this thesis proposes a series of innovative prediction models based on deep neural network models for four types of spatio-temporal data: spatio-temporal raster, spatio-temporal graph, spatio-temporal trajectory data with explicit spatial information, and multivariate time series with implicit spatial information. These models have been validated in typical applications for the corresponding types of spatio-temporal data, demonstrating the effectiveness of the proposed models. The main research work and contributions are summarized as follows:
1. Addressing the modeling issue of spatio-temporal grid data, where the spatial topological structure is static and uniformly distributed, two innovative U-shaped network models were proposed: the Hierarchical U-Net (HU-Net) and the Attention Spiking U-Net (ASU-Net). HU-Net, with its hierarchical design, aims to fully explore both global and local spatial correlations within spatio-temporal grid data and employs reparameterization techniques to enhance the model’s speed during inference. ASU-Net, on the other hand, is a spiking U-shaped network integrated with attention mechanisms, designed to reduce the high energy consumption of artificial neural networks during training and inference processes. Through experimental validation in weather forecasting applications, the HU-Net model significantly improved prediction accuracy; meanwhile, the ASU-Net model achieved the goal of maintaining high predictive precision at a lower energy consumption.
2. For the modeling problem of spatio-temporal graph data with a static but non-uniformly distributed spatial topological structure, two different modeling methods were proposed: the Dual-channel Spatio-temporal Wavelet Transform Neural Network (DSTWave) based on graph convolution and the Spatio-temporal Transformer (STTransformer) based on Transformer. The DSTWave model consists of a wavelet transform
network for disentangling long and short-term temporal patterns and a stacked dual-channel spatio-temporal graph convolution. The ST-Transformer model introduces a spatial Transformer and a temporal Transformer to model global spatial dependencies and long-term temporal dependencies, respectively. Through experimental
validation in traffic prediction applications, DSTWave and ST-Transformer achieved optimal predictive performance, thereby proving the effectiveness of the proposed methods.
3. Addressing the modeling problem of spatio-temporal trajectory data with a dynamic and non-uniformly distributed spatial topological structure, the Social-CVAE model based on Conditional Variational Autoencoders was proposed. The core of Social-CVAE is a conditional variational autoencoder architecture that learns the multimodal
distribution of an agent’s future trajectory using stochastic latent variables, conditioned on the observed past trajectories. Through experimental validation in pedestrian trajectory prediction applications, the proposed Social-CVAE achieved optimal results, confirming the effectiveness of the algorithm.
4. To address the long-term modeling problem of multivariate time series with implicit spatial information, the Non-stationary Conditional Variational Autoencoder (NCVAE) and the Fourier U-Net (F-UNet) were proposed. N-CVAE, by introducing noise and components designed for modeling the non-stationarity of time series data, enables the model to better adapt to different input distributions, alleviating the problem of distribution drift for more accurate long-term time series prediction. F-UNet is composed of neural network components with low time complexity, aiming to surpass the current state-of-the-art long-term time series prediction models based on Transformers, which have high time complexity. Through experimental validation on real industrial sites and
multiple public benchmark multivariate time series datasets, F-UNet was able to outperform current Transformer-based models in terms of lower time complexity and higher computational efficiency. N-CVAE also achieved optimal prediction results, proving the effectiveness of the proposed models.

Keyword时空数据预测 时空网格数据预测 时空图数据预测 时空轨迹数据预测 多元时间序列预测
Language中文
Document Type学位论文
Identifierhttp://ir.ia.ac.cn/handle/173211/57061
Collection毕业生_硕士学位论文
Recommended Citation
GB/T 7714
许宝文. 复杂时空数据预测方法研究[D],2024.
Files in This Item:
File Name/Size DocType Version Access License
复杂时空数据预测方法研究.pdf(7933KB)学位论文 限制开放CC BY-NC-SA
Related Services
Recommend this item
Bookmark
Usage statistics
Export to Endnote
Google Scholar
Similar articles in Google Scholar
[许宝文]'s Articles
Baidu academic
Similar articles in Baidu academic
[许宝文]'s Articles
Bing Scholar
Similar articles in Bing Scholar
[许宝文]'s Articles
Terms of Use
No data!
Social Bookmark/Share
All comments (0)
No comment.
 

Items in the repository are protected by copyright, with all rights reserved, unless otherwise indicated.