Lstm Vibration Prediction Based on Wavelet Analysis and Pso for the Pumped Storage Hydropower Unit

Against the backdrop of the global energy transition, achieving carbon peaking and carbon neutrality has become a common goal of the international community. Pumped storage, as a pivotal energy storage technology, plays a crucial role in the regulation of power grids. This study concentrates on the vibration prediction problem of pump-turbine units in pumped storage power plants, proposes a long short-term memory (LSTM) neural network model that integrates wavelet analysis (WA) and particle swarm optimization (PSO), namely WA-PSO-LSTM. The model applies wavelet transform for multi-scale decomposition of vibration signals to identify and extract different frequency components of the signals, and then predicts the decomposed signal subsets using the LSTM model, while the hyperparameters of the LSTM model are optimized using the PSO algorithm, and finally the prediction value of the original signal is obtained by synthesising the prediction results. The study employs the vibration data of unit No. 2 of a specific power plant under pump condition as the experimental subject. A comparative analysis of the traditional LSTM model and the LSTM-PSO model revealed that the WA-PSO-LSTM model demonstrated superior performance in several evaluation indicators, thereby substantiating the model's notable advantages in terms of prediction accuracy and stability. Furthermore, the study assesses the precise influence of distinct wavelet functions on the predictive performance. The experimental findings indicate that the DB2 wavelet function yielded the most optimal prediction in this particular case. The proposal of the WA-PSO-LSTM model not only provides an effective tool for condition monitoring and health management of pump-turbine units but also provides a feasible research method for signal prediction of other complex engineering equipment, which has important engineering application value and theoretical research significance.