Dynamic feature extraction and recognition of flow states in vaneless space of a prototype reversible pump turbine in generating mode based on variational mode decomposition and energy index

Dynamic feature extraction of pressure fluctuation signals and recognition of flow states in vaneless space have important engineering significances to guarantee the operational stability of the prototype reversible pump turbine (RPT) in the pumped hydro energy storage power station. In the present paper, in order to carry out accurate dynamic feature extraction of the flow states in the vaneless space, a method based on variational mode decomposition (VMD), energy index (EI) and support vector machine (SVM) is proposed. During the analysis procedure of pressure fluctuation signals, VMD can greatly avoid the mode mixing phenomenon, which often occurs in conventional empirical mode decomposition (EMD). After VMD analysis, several mode components with strong physical significances and separated central frequencies are obtained. Then, the EI of each mode component based on VMD is calculated and screened out. The EIs of the first three mode components are further constructed as the eigenvector that can accurately reflect different flow states in the vaneless space of the RPT. In addition, the intelligent classifier of SVM is employed to identify three types of flow states. The above eigenvectors can be employed as the input vectors of SVM. The average recognition results of 10 times show that the correct recognition rate of the proposed VMD-EI-SVM method (98.67 %) is higher than that of the EMD-EI-SVM method (88.00 %), which is more suitable for engineering applications. •A method of dynamic feature extraction is proposed based on energy index of variational mode.•Effective eigenvector is established to reflect different flow states in vaneless space of prototype reversible pump turbine.•Recognition and classification of different flow states are successfully realized in support vector machine.•Analysis results of proposed method are better than those based on empirical mode decomposition.