Root cause localization for wind turbines using physics guided multivariate graphical modeling and fault propagation analysis

Under the background of widespread utilization of wind energy, the improvement of power generation efficiency and the reduction of operation and maintenance costs have increased the demand for anomaly detection and localization of wind turbines through fault propagation analysis, while the large-scale data collected by SCADA systems provide sufficient data support and promising assistance for this purpose. It is an effective and cost-saving solution to this demand to construct a high-precision wind turbine digital twin model under normal operating conditions based on multivariate monitoring data, and to realize root cause localization based on prediction deviation and backtracking analysis. This paper proposes a root cause localization framework for wind turbines based on explicit-implicit knowledge fusion and multivariate time-series graph neural networks, fully utilizing the inter-variable dependencies mining capability of graph structure learning and the temporal prediction advantage of time-series graph neural networks. Five types of explicit knowledge based on explicit rules are constructed to describe the explicit relationships between monitoring variables. The fused knowledge (i.e., graph adjacency matrix) is obtained by fusing the explicit knowledge with the implicit knowledge hidden in the data through a quasi-hard-attention mechanism based on the idea of graph structure learning. A high-precision digital twin model for wind turbines based on fused knowledge is constructed using the multivariate time-series graph neural network. The anomaly degree of monitoring variables’ state is defined based on the prediction deviation, with which the propagation of fault's state over time can be accurately described, and the root cause localization through backtracking analysis can be achieved. Case studies are conducted using field fault data from a wind farm, a sugar factory and a thermal power plant. The results show that the proposed framework can provide a solution for studying the fault propagation process and root cause localization, also demonstrating certain robustness and generalizability.