A novel time series forecasting model for vibration of hydroelectric turbine units under multisource information fusion

Abstract

The abundant multi-location and historical vibration monitoring data of hydroelectric generating units provide valuable information for predictive maintenance strategies and digital transformation. However, existing methods primarily focus on univariate vibration prediction and fail to construct a global state representation of generating units through physical interrelationships among multiple sensors, which limits the adaptability of models to complex operating conditions. To address these limitations, this study proposes an innovative vibration prediction model for turbines, termed MSFSP, utilizes multi-scale gated units to extract vibration features from single sensor data, then combines point-wise convolution to couple the feature information between variables such as guide bearing swing, frame vibration, and finally predicts the unit’s vibration through a fully connected layer. The results demonstrate that in four different original vibration signal prediction tasks, the proposed model outperforms existing prediction methods in three cases, achieving a maximum reduction of 9.6% in mean squared error and 5.5% in mean absolute error. This indicates that the proposed prediction model exhibits high accuracy and strong noise resistance. This study provides high-precision data support for components of generating units with incomplete monitoring data, thereby laying a theoretical foundation for predictive maintenance strategies.