Spatial-Temporal Convolutional Transformer Network for Multivariate Time Series Forecasting.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Science

Background:

• Multivariate time series forecasting is crucial but challenging due to complex spatiotemporal dependencies.
• Existing methods often fail to capture local temporal context and diverse spatial patterns.
• These limitations hinder accurate forecasting in real-world applications.

Purpose of the Study:

• To propose a novel Transformer-based model, the spatial-temporal convolutional Transformer network (STCTN), for enhanced multivariate time series forecasting.
• To address the shortcomings of existing methods in modeling local temporal context and multiple spatial patterns.
• To improve the robustness and performance of time series forecasting models.

Main Methods:

• Developed the spatial-temporal convolutional Transformer network (STCTN), a novel Transformer-based architecture.
• Introduced a local-range convolutional attention mechanism to capture both global and local temporal dependencies.
• Designed a group-range convolutional attention mechanism to model multiple spatial patterns efficiently.
• Incorporated continuous positional encoding to enhance the linkage between historical and future data points.

Main Results:

• STCTN demonstrated superior performance compared to state-of-the-art methods across six real-world datasets.
• The model effectively captures complex local temporal and multiple spatial dependencies.
• STCTN showed increased robustness, particularly with nonsmooth time series data.

Conclusions:

• The proposed STCTN effectively addresses key challenges in multivariate time series forecasting.
• The novel attention mechanisms and positional encoding contribute to improved forecasting accuracy and robustness.
• STCTN represents a significant advancement in the field of spatiotemporal forecasting.