DecompKAN: Decomposed Patch-KAN for Long-Term Time Series Forecasting

📰 ArXiv cs.AI

Learn how DecompKAN, a novel architecture, achieves accurate long-term time series forecasting with transparency, and apply its techniques to your own forecasting tasks

advanced Published 28 Apr 2026

Action Steps

Apply trend-residual decomposition to your time series data to separate trends and residuals
Use channel-wise patching to extract local patterns from the decomposed data
Implement learned instance normalization to normalize the patched data
Configure a B-spline Kolmogorov-Arnold Network (KAN) with edge functions to model complex relationships
Test DecompKAN on your long-term time series forecasting task and compare its performance to existing methods

Who Needs to Know This

Data scientists and researchers working on time series forecasting tasks can benefit from DecompKAN's innovative approach, which combines decomposition, patching, and learned normalization to improve prediction accuracy and model interpretability

Key Insight

💡 DecompKAN's combination of decomposition, patching, and learned normalization enables accurate and interpretable time series forecasting

Key Takeaways

Learn how DecompKAN, a novel architecture, achieves accurate long-term time series forecasting with transparency, and apply its techniques to your own forecasting tasks

Full Article

Title: DecompKAN: Decomposed Patch-KAN for Long-Term Time Series Forecasting

Abstract:
arXiv:2604.23968v1 Announce Type: cross Abstract: Accurate time series forecasting in scientific domains such as climate modeling, physiological monitoring, and energy systems benefits from both competitive predictions and model transparency. This work proposes DecompKAN, a lightweight attention-free architecture that combines trend-residual decomposition, channel-wise patching, learned instance normalization, and B-spline Kolmogorov-Arnold Network (KAN) edge functions. Each KAN edge learns an e

Read full paper → ← Back to Reads