Woa-wtconv-kanformer for long term time series forecasting.

Area of Science:

• Time Series Analysis
• Machine Learning
• Signal Processing

Background:

• Multivariate time series analysis is crucial for applications like traffic management and weather forecasting.
• Traditional transformer models struggle with noise interference and require extensive hyperparameter tuning.
• Existing methods often fail to adequately address non-stationarity in time series data.

Purpose of the Study:

• To propose an optimized time series prediction model, WOA-WTConv-KANformer, that addresses limitations of traditional transformer models.
• To enhance the extraction of both frequency and time domain features for improved time series analysis.
• To reduce model complexity and improve training efficiency for practical applications.

Main Methods:

• The proposed WOA-WTConv-KANformer model optimizes the i-transformer architecture.
• Wavelet transform (WTConv2d) is employed for extracting frequency and time domain features, addressing non-stationarity.
• Whale Optimization Algorithm (WOA) is used for hyperparameter optimization, and the KAN module replaces MLPs for improved performance and reduced parameters.

Main Results:

• The WOA-WTConv-KANformer model demonstrated improved performance across various prediction lengths on five public datasets.
• The integration of WTConv2d effectively handled non-stationarity and noise interference.
• The use of WOA and KAN modules led to more efficient training and reduced parameter count.

Conclusions:

• The WOA-WTConv-KANformer model offers a significant advancement in multivariate time series prediction.
• The model's ability to handle complex data characteristics and improve training efficiency highlights its potential for real-world applications.
• This research provides a robust framework for future developments in time series forecasting.