CTF-former: A novel simplified multi-task learning strategy for simultaneous multivariate chaotic time series...

Area of Science:

• Chaos theory
• Time series analysis
• Machine learning

Background:

• Multivariate chaotic time series prediction is complex, especially for simultaneous multiple variable forecasting.
• Existing multi-model approaches struggle with synchronization, and multi-task learning lacks clear representation allocation principles.
• These challenges hinder precise and immediate communication between predicted values in related prediction tasks.

Purpose of the Study:

• To propose a novel, simplified multi-task learning method for precise simultaneous multiple chaotic time series prediction.
• To address synchronization issues and ambiguity in representation allocation inherent in current methods.
• To improve the accuracy and engineering applicability of multivariate chaotic time series forecasting.

Main Methods:

• Introduced a simplified multi-task learning scheme featuring a cross-convolution operator for capturing variable and sequence correlations.
• Developed an attention module utilizing non-linear transformations and convolution for sequence structure information, combining local and global dependencies.
• Devised an attention weight calculation integrating time-frequency domain features and series-channel information, with a simplified multi-task design reducing specific networks to single neurons.

Main Results:

• The proposed method demonstrated high precision in simultaneous multiple chaotic time series prediction.
• Validation on the Lorenz system showed an average 82.9% reduction in mean absolute error compared to Gated Recurrent Unit (GRU).
• Application to power consumption data resulted in a 19.83% average reduction in mean absolute error compared to GRU.

Conclusions:

• The simplified multi-task learning method effectively addresses challenges in multivariate chaotic time series prediction.
• The novel cross-convolution and attention mechanisms enhance correlation capture and information embedding.
• The proposed approach offers a precise and potentially applicable solution for complex forecasting tasks.