Correctformer: A transformer architecture for correcting periodic drift in time-series forecasting

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Science

Background:

• Transformer architectures excel at long-range dependencies in time-series forecasting.
• Existing models often neglect inherent periodic patterns, leading to performance degradation.
• Attention mechanisms can cause periodic blurring and drift, hindering accurate time-series analysis.

Purpose of the Study:

• To propose Correctformer, a Transformer-based architecture designed to improve time-series forecasting.
• To enhance the capture of periodic features within time-series data.
• To address the limitations of current attention mechanisms in handling periodicity.

Main Methods:

• Introduced periodic embedding to encode structural information of time-series periodicity.
• Implemented periodic correction to dynamically adjust and stabilize periodic attributes.
• Integrated these components into a novel Transformer architecture, Correctformer.

Main Results:

• Correctformer significantly improves the capture of periodic features in time-series data.
• The model demonstrates superior performance on datasets with complex periodic characteristics.
• Periodic embedding and correction effectively mitigate periodic blurring and drift.

Conclusions:

• Correctformer offers a more suitable Transformer-based architecture for time-series modeling, especially for periodic data.
• The proposed methods enhance the model's ability to learn true dynamic patterns.
• This approach advances the field of time-series forecasting by addressing periodicity challenges.