FCP-Former: Enhancing Long-Term Multivariate Time Series Forecasting with Frequency Compensation.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Time Series Analysis

Background:

• Long-term multivariate time series forecasting is vital for energy, traffic, healthcare, and finance.
• Transformer models with patch mechanisms offer computational efficiency but struggle with intra-patch temporal dependencies, limiting prediction accuracy.

Purpose of the Study:

• To propose the Frequency Compensation Patch-wise Transformer (FCP-Former) to enhance intra-patch temporal dependency capture in time series forecasting.
• To improve the accuracy and efficiency of long-term multivariate time series predictions.

Main Methods:

• Developed FCP-Former integrating a frequency compensation layer with a patching mechanism.
• Utilized Fast Fourier Transform (FFT) to extract frequency-domain features and enrich patch representations.
• Validated FCP-Former on eight benchmark datasets using PyTorch and NVIDIA RTX 4090 GPU.

Main Results:

• FCP-Former achieved 48 optimal and 17 suboptimal experiment results across all tested datasets.
• Demonstrated superior forecasting accuracy, with notable performance on ETTh1 (MSE: 0.437, MAE: 0.430) and Electricity (MSE: 0.186, MAE: 0.277) datasets.
• Showcased an enhanced ability to capture periodic and trend patterns in time series data.

Conclusions:

• FCP-Former effectively mitigates intra-patch information loss by integrating frequency-domain features.
• The proposed model offers improved accuracy and robust performance for long-term multivariate time series forecasting.
• FCP-Former represents a significant advancement in capturing complex temporal dynamics for predictive modeling.