A multivariate time series prediction model based on the KAN network.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Science

Background:

• Traditional deep learning models like RNNs and CNNs struggle with long-range dependencies in time series forecasting.
• Transformer models offer enhanced capabilities but face challenges with complexity and noise sensitivity.
• Existing methods often lack efficiency and interpretability for complex multivariate time series data.

Purpose of the Study:

• To introduce KANMTS, a novel model integrating Kolmogorov-Arnold Networks (KAN) and Multi-Layer Perceptrons (MLP) for improved time series forecasting.
• To enhance the capture of complex patterns and long-range dependencies in multivariate time series data.
• To provide a forecasting solution that is computationally efficient and interpretable.

Main Methods:

• Developed the KANMTS model by combining the non-linear mapping of KAN with the computational simplicity of MLP.
• Utilized KAN's ability to capture intricate data relationships and MLP's efficiency for multivariate time series analysis.
• Employed symbolic regression and visualization techniques to explore model interpretability.

Main Results:

• KANMTS demonstrated superior forecasting performance compared to existing methods across various datasets.
• The model showed significant accuracy improvements, particularly on large-scale datasets.
• KANMTS exhibited enhanced resource utilization efficiency and good generalization capabilities.

Conclusions:

• KANMTS offers a simple, efficient, and interpretable solution for multivariate time series forecasting.
• The integration of KAN and MLP effectively addresses limitations of previous deep learning models.
• This advancement holds promise for diverse applications requiring accurate and understandable time series predictions.