TC-KAN: Time-Conditioned Kolmogorov-Arnold Networks with Time-Dependent Activations for Long-Term Time Series...

Area of Science:

• Artificial Intelligence
• Machine Learning
• Time Series Analysis

Background:

• Traditional long-term time series forecasting (LTSF) models use static activation functions, failing to capture real-world data's regime-dependent dynamics.
• Power systems and energy management require accurate forecasting that accounts for temporal variations like seasonal peaks and daily patterns.

Purpose of the Study:

• To introduce TC-KAN (Time-Conditioned Kolmogorov-Arnold Network), the first forecasting architecture using position-aware KAN activations for improved LTSF.
• To develop a computationally efficient and parameter-light model for accurate long-term load forecasting.

Main Methods:

• Augmenting Kolmogorov-Arnold Network (KAN) activations with position-conditioned coefficients generated by a lightweight MLP.
• Integrating a dual-pathway block with depthwise convolution and time-conditioned KAN layers within a channel-independent framework.
• Utilizing Reversible Instance Normalisation for enhanced stability and performance.

Main Results:

• TC-KAN achieved superior or competitive accuracy on ETT and Weather datasets with only 51K parameters.
• Demonstrated significant Mean Squared Error reduction (up to 61.4% on ETTh2) compared to DLinear.
• Matched state-of-the-art performance on ETTm2 with substantially lower computational cost.

Conclusions:

• TC-KAN offers a highly practical and parameter-efficient solution for LTSF, outperforming existing models.
• The architecture effectively addresses the limitations of stationary activation functions in time series forecasting.
• TC-KAN is well-suited for resource-constrained edge deployments in smart grids and IoT devices.