Multi-scale convolution enhanced transformer for multivariate long-term time series forecasting.

Area of Science:

• Data Science
• Machine Learning
• Deep Learning

Background:

• Transformer models show promise for time series forecasting but face challenges with computational complexity and modeling local/cross-dimensional dependencies.
• Existing methods struggle with efficiency and accurately capturing intricate relationships in multivariate long-term time series data.

Purpose of the Study:

• To propose the Multi-Scale Convolution Enhanced Transformer (MSCformer) model for improved multivariate long-term time series forecasting.
• To address the limitations of standard Transformer models in terms of computational complexity and dependency modeling.

Main Methods:

• A novel multi-scale segmentation strategy is employed to process time series into segments of varying lengths.
• A Multi-Dependency Aggregation module is introduced to capture cross-temporal and cross-dimensional dependencies, enhancing local feature representation.
• The MSCformer model synthesizes dependency information from multi-scale segments to reconstruct future time series.

Main Results:

• The multi-scale segmentation reduces the computational complexity of the attention mechanism.
• The Multi-Dependency Aggregation module effectively captures complex dependencies and local features.
• MSCformer demonstrates superior forecasting accuracy compared to existing methods across diverse domains.

Conclusions:

• MSCformer offers a more efficient and accurate approach to multivariate long-term time series forecasting.
• The model's ability to leverage multi-scale information and complex dependencies provides a significant advancement in the field.