STF-DKANMixer: Tri-component decomposition with KAN-MLP hybrid architecture for time series forecasting

Area of Science:

• Artificial Intelligence
• Machine Learning
• Time Series Analysis

Background:

• Long-term time series forecasting is crucial for traffic and energy systems.
• Current models struggle with multiscale patterns and nonlinear dynamics, leading to inaccuracies during abrupt changes.

Purpose of the Study:

• Introduce STF-DKANMixer, a hybrid architecture for improved long-term time series forecasting.
• Address limitations of contemporary models in capturing complex patterns and nonlinear dynamics.

Main Methods:

• Hybrid architecture combining Multi-Layer Perceptron (MLP) and Kolmogorov-Arnold Network (KAN).
• DFT-based decomposition for trends/seasonality and Haar wavelet for residuals.
• Past-Information-Mixing (PIM) with KAN and deformable feature attention (DFA).
• Future-Information-Mixing (FIM) using adaptive weighted ensemble with residual connections.

Main Results:

• STF-DKANMixer significantly outperforms state-of-the-art models.
• Reduced Mean Squared Error (MSE) by up to 36.1% and Mean Absolute Error (MAE) by up to 28.8%.
• Achieved superior results using less than half the computational resources.

Conclusions:

• STF-DKANMixer is a robust, efficient, and highly accurate solution for complex long-horizon forecasting.
• Sets a new performance standard for long-term time series forecasting challenges.