Integrated codec decomposed Transformer for long-term series forecasting.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Science

Background:

• Transformer and Multilayer Perceptron (MLP) architectures are key in time series forecasting.
• Series decomposition can improve temporal pattern recognition in forecasting models.
• Existing Transformer-based decomposed models may overlook trend information and lose temporal order.

Purpose of the Study:

• To propose a novel time series forecasting model that addresses limitations in existing Transformer-based decomposed approaches.
• To enhance the perception of temporal patterns by effectively utilizing trend and seasonal information.
• To improve the handling of temporal dependencies and reduce information loss in forecasting.

Main Methods:

• Utilizing attention mechanisms for multivariate correlations in trends and MLP for seasonal patterns.
• Introducing an integrated codec for consistent multivariate relationship representation in encoding and decoding.
• Implementing a trend enhancement module to stabilize trends and improve attention-based feature representation.

Main Results:

• The proposed model demonstrates state-of-the-art prediction performance on large-scale datasets.
• The model effectively captures complex temporal patterns by integrating trend and seasonal components.
• The trend enhancement module mitigates the loss of sequentiality in attention mechanisms.

Conclusions:

• The novel approach offers significant improvements in time series forecasting accuracy.
• Decomposition combined with specialized component handling (attention for trends, MLP for seasonality) is effective.
• The model provides a robust framework for capturing intricate temporal dynamics.