Time-series forecasting with deep learning: a survey

Area of Science:

• Machine Learning
• Time-Series Analysis
• Artificial Intelligence

Background:

• Diverse time-series datasets necessitate specialized deep learning architectures.
• Existing models often focus on either statistical methods or neural networks.

Purpose of the Study:

• To survey common encoder and decoder designs in time-series forecasting.
• To highlight advancements in hybrid deep learning models.
• To explore deep learning applications in time-series decision support.

Main Methods:

• Review of deep learning architectures for time-series forecasting.
• Analysis of encoder-decoder designs for one-step-ahead and multi-horizon forecasting.
• Examination of hybrid models combining statistical and neural network approaches.

Main Results:

• Common encoder-decoder architectures effectively incorporate temporal information.
• Hybrid models show potential for improving upon pure statistical or deep learning methods.
• Deep learning offers new avenues for time-series data-driven decision support.

Conclusions:

• Deep learning architectures are crucial for handling diverse time-series data.
• Hybrid models represent a promising direction for enhanced forecasting accuracy.
• The application of deep learning extends beyond forecasting to decision support systems.