Swirling vortex rope in draft tube (DT) is a typical hydraulic instability of a pump turbine (PT) in the pumped storage plant (PSP). In view of the potential hazards of the vortex rope, accurate recognition of its intensity is of great significance to maintain the stable operation of the PT. Due to the limitations of shallow learning algorithms during intelligent recognition, an adaptive deep learning framework is innovatively proposed in this study. Firstly, the measured high-precision pressure fluctuation signals based on the prototype PT in a Chinese PSP that can reflect different intensities of vortex ropes in the DT are utilized as the input data. Secondly, a preliminary deep learning framework that integrates convolutional neural network (CNN), bidirectional long short-term memory (BiLSTM) and multi-head self-attention mechanism (MHSAM) is constructed. Then, the Bayesian optimization algorithm (BOA) is utilized to adaptively determine several hyperparameters of the framework. And an adaptive BOA-CNN-BiLSTM-MHSAM framework is established to recognize different intensities of vortex ropes in the DT. Finally, the recognition performance of the proposed framework is demonstrated through comparing with other deep learning frameworks. And the recognition results illustrate that the proposed BOA-CNN-BiLSTM-MHSAM framework can be utilized to effectively recognize different intensities of vortex ropes in the DT. It will be a good technical reserve to improve the intelligent level of the monitoring system of the PSP.
Read full abstract