MoEKAN: Multi-Scale Transformer-Based Gating KAN Experts Network for Time Series Forecasting.

Area of Science:

• Electrical Engineering
• Artificial Intelligence
• Signal Processing

Background:

• On-load tap changers (OLTCs) are vital power transformer components.
• Vibration signals from OLTCs offer insights into equipment health and potential faults.
• Accurate forecasting of these signals is crucial for predictive maintenance.

Purpose of the Study:

• To introduce a novel MoEKAN model for enhanced time series forecasting of OLTC vibration signals.
• To improve the accuracy and reliability of condition monitoring for OLTCs.

Main Methods:

• Developed a Mixture-of-Experts-based Kolmogorov-Arnold Network (MoEKAN) model.
• Integrated reversible instance normalization (RevIN) for data distribution adaptability.
• Employed a transformer-based gating mechanism for dynamic expert integration.
• Utilized multi-scale signal processing to capture complex signal patterns.

Main Results:

• The MoEKAN model demonstrated superior forecasting performance on real OLTC operational data.
• Achieved low forecasting errors: MSE of 133.4579, MAE of 7.2801, and MAPE of 4.4272.
• Outperformed all benchmarked comparison models in accuracy.

Conclusions:

• The MoEKAN model is a practical and effective tool for OLTC vibration signal forecasting.
• Confirms the model's potential for reliable condition monitoring and predictive maintenance of OLTCs.
• Highlights the significance of advanced AI models in power system diagnostics.