Power Plant Transient Fault Diagnostics Based on Two-Stage Neural Networks

In order to realize power plant transient fault diagnosis, a fault diagnosis approach based on two-stage neural networks is proposed. Among them, the first neural network adopts an improved Elman neural network with time-delay sequence inputs to predict the expected normal values of the fault feature variables, which is applied to calculate the normalized fault symptoms for fault classification purpose. The second neural network uses a conventional feedforward network to fulfill real-time fault diagnosis. By taking the inner leakage faults of the feedwater heaters of a 600MW supercritical unit as example investigated, an expected value prediction model of the fault feature variables is built, trained and validated by a large amount of historical operating data, including typical steady-state load conditions and load-varying transient processes. With this two-stage NN fault diagnosis method, detailed fault diagnosis simulation tests are carried out on a full-scope simulator of the given supercritical power unit. It is shown by tests that the suggested method can achieve satisfactory fault diagnosis results for load-varying transient processes.