Thermal Fault Warning of Turbine Generators Based on Cluster Heatmap CNN-GRU-Attention Method

The outlet water temperature of a water-cooled turbine generator is a crucial indicator for early detection of generator stator winding thermal faults. This paper proposes a thermal fault warning method for stator winding based on cluster heatmaps, CNN-GRU-Attention, and sliding window technique. First, thermal cluster maps are used for visual analysis of variables collected by the DCS system, and the most important features affecting the outlet water temperature are selected to form input feature vectors, which are then inputted into the hybrid neural network prediction model. Then, the sliding window technique is employed to calculate the maximum normal residual value, which serves as the fault threshold to issue a thermal fault early warning. Finally, the proposed method is validated by using the historical DCS data collected from a 1000 MW steam turbine generator unit, and the results show that the method proposed in this paper performs well in terms of both sensitivity and reliability.