Pressure Pulsations Prediction Framework for Load Rejection of Pumped Storage Power Station Based on Data Augmentation and One-Dimensional Convolutional Neural Network

Abstract

The safety of pumped storage power plants (PSPs) will be threatened by the extreme pressure pulsation that occurs during the load rejection transition process. Accurately forecasting pressure pulsation in extreme load rejection conditions is essential for power plant safety. This study suggests a pressure pulsations prediction framework based on data enhancement and one-dimensional convolutional neural network (1D-CNN) to precisely predict the pressure pulsations at the volute inlet (VI) and draft tube inlet (DTI) under conditions of extreme load rejection. In order to effectively achieve the augmentation of pressure pulsation data, this work first proposes a data augmentation technique based on variational mode reconstruction. The water hammer pressure (WP) is then extracted from the enhanced data using the Savitzky-Golay filter and put into the constructed 1D-CNN for training. After that, utilizing the method of characteristics, a fine simulation model of the PSP is constructed. This model is employed to calculate the WP at the VI and DTI. Finally, the calculated WP is input into the trained 1D-CNN to predict the corresponding pressure pulsations. Based on the measured load rejection data of a PSP in China, this study verifies the effectiveness of the proposed pressure pulsations prediction framework. The findings indicate that the prediction errors of maximum pressure at VI and minimum pressure at the DTI are only 2 m and 1.01 m, respectively.

Keywords: Pumped storage power station, Transient process, Pressure pulsations prediction, Data augmentation, Convolutional neural network