Time Evidence Fusion Network: Multi-Source View in Long-Term Time Series Forecasting.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Science

Background:

• Time series forecasting demands both accuracy and computational efficiency.
• Existing model architectures are continuously researched to meet these practical requirements.
• Uncertainty and complexity in multivariate time series data pose significant challenges.

Purpose of the Study:

• To propose a novel backbone architecture, the Time Evidence Fusion Network (TEFN), for enhanced time series forecasting.
• To address the challenges of accuracy, efficiency, and interpretability in forecasting models.
• To introduce a new approach for capturing and fusing information from multivariate time series data.

Main Methods:

• Developed the Time Evidence Fusion Network (TEFN) incorporating a Basic Probability Assignment (BPA) Module.
• Utilized evidence theory within the BPA Module to capture data uncertainty across channel and time dimensions.
• Implemented a novel multi-source information fusion method to integrate BPA outputs.

Main Results:

• TEFN achieves performance comparable to state-of-the-art methods.
• TEFN demonstrates significantly lower computational complexity and reduced training time.
• Experiments confirm TEFN's high robustness with minimal error fluctuations and strong interpretability.

Conclusions:

• TEFN offers a balanced solution for time series forecasting, excelling in accuracy, efficiency, stability, and interpretability.
• The proposed architecture effectively handles uncertainty in multivariate time series data.
• TEFN presents a desirable and practical choice for real-world forecasting applications.