A comprehensive survey of deep learning for time series forecasting: architectural diversity and open challenges

TL;DR

A comprehensive survey paper that systematically organizes the architectural diversity and open challenges of deep learning methodologies for time series forecasting.

Problem

Time series forecasting is crucial in various fields such as finance, weather, and energy, but the rapid development of deep learning models has led to fragmented methodologies and difficulty in comparison. Existing surveys are biased toward specific architectures or applications and fail to provide the overall picture.

Approach

Classify major deep learning architectures such as CNN, RNN, Transformer, and hybrid models, and analyze their strengths, weaknesses, and application cases. Also address common issues such as data preprocessing, evaluation metrics, and interpretability, and identify open challenges including long-term forecasting, uncertainty quantification, and distribution shift.

Results & Contribution

Provide an overall landscape of deep learning-based time series forecasting, helping researchers select methodologies and set future research directions. In particular, the architecture selection guidelines and list of unresolved issues have high practical value.