Intelligent fault prediction and diagnosis for wind-powered heating systems using graph neural networks

Area of Science:

• Renewable Energy Systems
• Artificial Intelligence in Engineering
• Data Science

Background:

• Wind-powered heating systems are crucial for clean energy but face reliability issues due to complex operations and harsh conditions.
• Efficient utilization of wind energy, especially in northern China, requires robust system monitoring and fault prediction.
• Existing methods struggle with multi-source data fusion and extracting critical spatiotemporal features.

Purpose of the Study:

• To develop an adaptive fault prediction and intelligent diagnosis method for wind-powered heating systems.
• To address challenges in data fusion and spatiotemporal feature extraction in these complex systems.
• To improve the reliability and operational efficiency of wind energy utilization.

Main Methods:

• A Multi-level Spatiotemporal Graph Neural Network (MSGNN) framework was developed.
• Dynamic adaptive thresholds were generated using maximum a posteriori probability estimation and interquartile range analysis.
• Graph attention networks, parallel subgraph architectures, and temporal attention modules were employed for feature extraction and dependency analysis.

Main Results:

• The proposed method achieved a comprehensive prediction accuracy of 93.5% and a fault detection Fβ-score (β=0.5) of 0.95.
• Demonstrated an 18.6% improvement over traditional methods in diagnostic capability.
• Showcased strong robustness with KL divergence of 0.09 ± 0.02 under transient conditions using 42 TB of SCADA data.

Conclusions:

• The MSGNN-based approach offers superior multi-level diagnostic capabilities for wind-powered heating systems.
• The adaptive threshold mechanism enables real-time monitoring and early fault warnings, enhancing system reliability.
• This method provides a significant advancement for the intelligent operation and maintenance of renewable energy infrastructure.