Swing Trend Prediction of Main Guide Bearing in Hydropower Units Based on MFS-DCGNN

Area of Science:

• Engineering
• Data Science
• Predictive Maintenance

Background:

• Hydropower units are critical for power generation, requiring robust fault prediction and health management.
• Accurate swing trend prediction is essential for ensuring the safety, stability, and reliability of hydropower operations.
• Existing methods often struggle with spatial dependencies and graph structure definition in complex systems.

Purpose of the Study:

• To develop an advanced algorithm for predicting the swing trends of main guide bearings in hydropower units.
• To improve the accuracy and effectiveness of fault prediction and health management systems for hydropower stations.
• To address limitations in existing Graph Neural Network (GNN) methods regarding predefined graph structures and over-smoothing.

Main Methods:

• Feature selection using multi-index algorithms and Pearson/distance correlation coefficients to identify key state variables.
• Implementation of a Dilated Convolution Graph Neural Network (DCGNN) model for swing trend prediction.
• DCGNN utilizes dilated convolutions and graph learning to capture spatial dependencies without explicit graph structures, mitigating over-smoothing.

Main Results:

• The DCGNN algorithm demonstrated superior performance compared to RNN-GRU, LSTNet, and TAP-LSTM.
• DCGNN achieved lower Mean Absolute Errors (MAEs) by up to 6.32% and Root Mean Square Errors (RMSEs) by up to 9.21%.
• Correlation (CORR) values were improved by up to 1.05%, indicating significantly enhanced prediction accuracy.

Conclusions:

• The DCGNN algorithm effectively predicts hydropower unit swing trends, offering a significant improvement over existing methods.
• This approach enhances fault prediction and health management, contributing to safer and more reliable hydropower operations.
• The DCGNN's ability to learn spatial dependencies without predefined graphs makes it a powerful tool for complex time-series analysis.