Time-Series Forecasting Method Based on Hierarchical Spatio-Temporal Attention Mechanism.

Area of Science:

• Intelligent decision-making
• Data science
• Machine learning

Background:

• Sensor data is crucial for physical-digital interaction in intelligent decision-making.
• Traditional time-series methods struggle with spatio-temporal coupling and long-range dependencies.
• Challenges exist in feature decoupling and multi-scale modeling for complex sensor data.

Purpose of the Study:

• To propose an innovative network, the Spatio-Temporal Attention-Enhanced Network (TSEBG), for enhanced time-series prediction.
• To address the limitations of traditional methods in handling spatio-temporal coupling and long-range dependencies.
• To improve feature decoupling and multi-scale modeling capabilities for sensor-based time-series data.

Main Methods:

• Reconstruction of Temporal Convolutional Network (TCN) layers using Squeeze-and-Excitation Network (SENet) for improved feature expression.
• Development of a Bidirectional Gated Recurrent Unit (BiGRU) with a global attention mechanism to capture cross-period dependencies and mitigate gradient disappearance.
• Implementation of a hierarchical feature fusion architecture with residual connections and dynamic attention for multi-dimensional alignment and semantic representation.

Main Results:

• The TSEBG model demonstrates superior performance in time-series single-step prediction tasks compared to dominant existing models.
• Achieved high accuracy and performance with excellent generalization stability, evidenced by a cross-dataset R² standard deviation of only 3.7%.
• Effectively addresses redundancy in local pattern capture and alleviates gradient disappearance in RNN-like models.

Conclusions:

• TSEBG offers a novel theoretical framework for analyzing complex time-series data, particularly in feature decoupling and multi-scale modeling.
• The proposed network provides a robust solution for critical applications requiring accurate sensor data analysis, such as industrial monitoring and intelligent transportation.
• The study highlights the effectiveness of integrating attention mechanisms and hierarchical fusion for advanced time-series forecasting.