Quantum-enhanced dual-layer graph attention network for time-series forecasting.

Area of Science:

• Artificial Intelligence
• Quantum Computing
• Data Science

Background:

• Time-series forecasting is crucial for traffic, finance, and energy, but faces challenges from complex data patterns.
• Existing methods struggle with intertwined trends, seasonality, and latent structures, impacting forecasting accuracy.

Purpose of the Study:

• To propose QFreqFormer, a novel quantum-enhanced deep learning model for advanced time-series forecasting.
• To leverage quantum parallelism for efficient frequency decomposition and graph attention for temporal dependency modeling.

Main Methods:

• The Quantum Fourier Transform (QFT) decomposes time-series into frequency components.
• A Quantum Frequency Decomposition-Reconstruction (Q-FR-Q) module separates high- and low-frequency patterns.
• A Dual-Layer Graph Attention Network (D-PAD) models temporal dependencies across frequencies.

Main Results:

• QFreqFormer consistently outperforms state-of-the-art methods on benchmark datasets for Mean Squared Error (MSE) and Mean Absolute Error (MAE).
• The model demonstrates robust transfer learning capabilities across diverse forecasting scenarios.
• Achieved improvements in both forecasting accuracy and computational efficiency.

Conclusions:

• QFreqFormer offers a powerful quantum-enhanced deep learning framework for complex time-series forecasting.
• The model's ability to handle multi-frequency patterns and temporal dependencies enhances its generalizability.
• This approach presents practical advantages for real-world applications requiring accurate and efficient forecasting.