Scholarlink

<p>Time series forecasting involves identifying patterns and trends in historical time series data to estimate future values, while spatio-temporal forecasting incorporates spatial information to predict future states across both time and space. In recent years, these forecasting tasks have found widespread applications in key domains such as finance, meteorology, energy, transportation, and healthcare. As data volume and model complexity continue to increase, deep learning models have significantly improved predictive accuracy, but also raised the demand for computational resources. In particular, tasks such as large-scale graph modeling, long-range dependency learning, and uncertainty estimation increasingly rely on high-performance computing (HPC), GPU acceleration, and distributed processing. This paper provides a comprehensive review of recent deep learning models for time series and spatio-temporal forecasting. We analyze the characteristics, advantages, and limitations of various models, with a focus on representative approaches based on Transformer architectures and hybrid designs. In addition, we introduce common evaluation metrics, benchmark datasets, and typical application domains. The paper further discusses the computational cost and scalability of these models, especially their adaptability to HPC environments, and offers practical guidelines for model selection tailored to different application scenarios.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

A review of time series forecasting and spatio-temporal series forecasting in deep learning

  • 期刊论文

  • 发表日期：

    2025年7月15日

查看全文
  • Qianqian Yu,
  • Guang Yang,
  • Xiao Wang,
  • Yaxin Shi,
  • Ying Feng,
  • Ang Liu

摘要

Time series forecasting involves identifying patterns and trends in historical time series data to estimate future values, while spatio-temporal forecasting incorporates spatial information to predict future states across both time and space. In recent years, these forecasting tasks have found widespread applications in key domains such as finance, meteorology, energy, transportation, and healthcare. As data volume and model complexity continue to increase, deep learning models have significantly improved predictive accuracy, but also raised the demand for computational resources. In particular, tasks such as large-scale graph modeling, long-range dependency learning, and uncertainty estimation increasingly rely on high-performance computing (HPC), GPU acceleration, and distributed processing. This paper provides a comprehensive review of recent deep learning models for time series and spatio-temporal forecasting. We analyze the characteristics, advantages, and limitations of various models, with a focus on representative approaches based on Transformer architectures and hybrid designs. In addition, we introduce common evaluation metrics, benchmark datasets, and typical application domains. The paper further discusses the computational cost and scalability of these models, especially their adaptability to HPC environments, and offers practical guidelines for model selection tailored to different application scenarios.