Abstract
With the rapid global transition towards clean energy, wind-powered heating systems have emerged as a critical solution for efficient wind energy utilization, particularly in the northern regions of China. However, these systems face significant reliability challenges due to complex spatiotemporal couplings and harsh operating conditions. This paper presents an adaptive fault prediction and intelligent diagnosis method based on a Multi-level Spatiotemporal Graph Neural Network to address the challenges of multi-source data fusion difficulties and inadequate spatiotemporal feature extraction. The proposed framework establishes a dynamic adaptive threshold generation mechanism by integrating maximum a posteriori probability estimation with interquartile range analysis, enabling real-time system state monitoring and early fault warning. The methodology incorporates graph attention networks, seven-branch parallel subgraph architectures, and multi-head attention mechanisms to capture topological evolution patterns through dynamic graph neural networks, while temporal attention modules are employed to enhance sequential dependencies of critical parameters. Experimental validation was conducted using 42 TB of SCADA data from China Guoneng Group's 200 MW wind-heat cogeneration project. The results demonstrate the model's superior multi-level diagnostic capability, achieving a comprehensive prediction accuracy of 93.5% and a fault detection Fβ-score (β = 0.5) of 0.95-representing an 18.6% improvement over traditional approaches-while maintaining strong robustness (KL divergence 0.09 ± 0.02) under transient operating conditions.