What question did this study set out to answer?

To provide a comprehensive, theoretically grounded synthesis and taxonomy of Transformer-based models, applications, and performance limitations in time-series learning.

December 19, 2025Computer Science ReviewOpen Access

A survey of transformer networks for time series forecasting

View Full Paper

Ask AI

Bookmark

Authors

JZJingyuan Zhao FCFulei Chu LXLili Xie

Discussion

Member takes

Overview

Systematic review synthesizes transformer architectures for time series modeling across diverse operational domains, highlighting key trade-offs in computational scalability and real-world efficiency.

Key Points

To provide a comprehensive, theoretically grounded synthesis and taxonomy of Transformer-based models, applications, and performance limitations in time-series learning.
Conducted a systematic review analyzing literature from IEEE Xplore, ACM Digital Library, and Scopus published between 2020 and 2025.
Categorized Transformer variants using an architecture-centered and task-aware taxonomy covering forecasting, representation learning, anomaly detection, and multimodal fusion across finance, healthcare, energy, and transportation.
Identified that Transformer self-attention mechanisms effectively capture long-range temporal dependencies and scale to high-volume data better than traditional recurrent and convolutional architectures.
Identified critical operational bottlenecks, including high computational complexity, memory overhead during long-sequence processing, susceptibility to overfitting, and limited interpretability.
Outlined future research priorities, emphasizing physics-informed architecture design, hybrid mechanistic modeling, and lightweight inference for real-time edge computing.

Cite This Study

Zhao et al. (2025) studied this question.

synapsesocial.com/papers/69dab36eed2e131d3c6843c3 https://doi.org/10.1016/j.cosrev.2025.100883

View Full Paper

Ask AI

Bookmark

December 19, 2025Computer Science ReviewOpen Access

A survey of transformer networks for time series forecasting

View Full Paper

Ask AI

Bookmark

Authors

JZJingyuan Zhao FCFulei Chu LXLili Xie

Discussion

Member takes

Overview

Key Points

To provide a comprehensive, theoretically grounded synthesis and taxonomy of Transformer-based models, applications, and performance limitations in time-series learning.
Conducted a systematic review analyzing literature from IEEE Xplore, ACM Digital Library, and Scopus published between 2020 and 2025.
Categorized Transformer variants using an architecture-centered and task-aware taxonomy covering forecasting, representation learning, anomaly detection, and multimodal fusion across finance, healthcare, energy, and transportation.
Identified that Transformer self-attention mechanisms effectively capture long-range temporal dependencies and scale to high-volume data better than traditional recurrent and convolutional architectures.
Identified critical operational bottlenecks, including high computational complexity, memory overhead during long-sequence processing, susceptibility to overfitting, and limited interpretability.
Outlined future research priorities, emphasizing physics-informed architecture design, hybrid mechanistic modeling, and lightweight inference for real-time edge computing.

Cite This Study

Zhao et al. (2025) studied this question.

synapsesocial.com/papers/69dab36eed2e131d3c6843c3 https://doi.org/10.1016/j.cosrev.2025.100883

View Full Paper

Ask AI

Bookmark