RFNet: Multivariate long sequence time-series forecasting based on recurrent representation and feature enhancement.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Science

Background:

• Multivariate time series forecasting (MLSTF) is challenging due to complex interactions and long-term dependencies.
• Existing Transformer-based models often use complex encoder-decoder architectures, increasing computational cost and neglecting spatial patterns.

Purpose of the Study:

• To propose RFNet, a lightweight model for MLSTF that addresses computational complexity and spatial information limitations.
• To improve the accuracy and efficiency of forecasting long multivariate time series.

Main Methods:

• RFNet partitions time series into subsequences to capture local temporal and cross-variable spatial patterns.
• A recurrent representation module with gate attention and memory units captures local and long-term correlations.
• A shared multi-layer perceptron (MLP) and a feature enhancement module extract global and complex spatial patterns.

Main Results:

• RFNet was validated on ten real-world datasets.
• The model demonstrated significant performance improvements over state-of-the-art MLSTF models.
• An approximate 55.3% improvement was observed compared to existing methods.

Conclusions:

• RFNet offers a computationally efficient and effective solution for MLSTF.
• The model's recurrent representation and feature enhancement approach successfully captures complex temporal and spatial patterns.
• RFNet presents a significant advantage in multivariate long sequence time series forecasting.