Study On Performance Evaluation And Prediction Of Hydropower Units Under Complex Operating Conditions

As the core equipment for the development and utilization of hydropower resources,hydropower units(HU)undertake the critical tasks of peak regulation,frequency regulation phase regulation and emergency backup of the power system.Therefore,it is difficult to evaluate the state of the HU due to the complex and changeable operating conditions.In this dissertation,to ensure the safety and the efficiency of the HU,advanced technology such as fluid-structure coupling simulation,clustering algorithm,artificial intelligence and intelligent optimization is introduced as the theoretical background.The physics driven simulation and the data driven machine learning are fully utilized.From the perspectives of numerical simulation,monitoring data enhancement,performance state evaluation and degradation trend prediction,several key scientific problems in the performance evaluation and the degradation trend prediction of the HU under complex working conditions are studied.This study can provide technical support for the intelligent maintenance and statebased maintenance of the HU.The main innovative achievements are as follows:(1)The numerical simulation model of the whole flow passage is established first.The influence of the power and the water head on flow phenomena such as vortex and cavitation are explored.The frequency spectrum of the pressure fluctuation in the flow passage are analyzed.Furthermore,the fluid-structure coupling simulation model is constructed to solve the natural vibration mode of the runner under the action of multi-source external forces.It is determined that the possibility of the runner resonance caused by fluid-induced vibration is relatively low.In addition,through solving the stress and strain distribution,the vulnerable part of the runner is discovered at the outflow edge of the blades.The theoretical service life of the runner is estimated,which provides a reference for the result interpretation of the data-diven performance evaluation of the HU.(2)Focusing on the actual engineering problems of abnormal and missing values in the measured data set of the HU,the association between operating conditions and monitoring data is fully considered,and a low-quality data cleaning method based on density clustering algorithm is proposed to adaptively remove the anomaly samples.Furthermore,the Wasserstein generative adversarial network is taken as the basic framework of data generation,and the gated recurrent unit with decay mechanism network is taken as the critical component of missing information mining,the missing value imputation model for the monitoring data is constructed.It improves the accuracy of missing data imputation,enhances the utilization value of low quality monitoring data and provides a high quality data basis for the performance evaluation and degradation trend prediction of the HU.(3)The traditional performance evaluation methods with fixed thresholds ignore the influence of the variable operating conditions on the monitoring data,and the existing deterministic mapping health model is greatly affected by the signal distribution range.In this dissertation,considering the distribution features’ strong robustness towards signal random fluctuation and the low sampling rates,based on the Dirichlet process-Gaussian mixture model,the healthy probability model of critical measuring points is established,which improved the stability of the performance evaluation.Then,based on the selfadjusting analytic hierarchy process,the comprehensive performance deterioration curve of the HU is constructed by integrating the degradation information of several key measuring points,which can accurately describe the performance deterioration process of the HU,and help the decision maker to understand the performance of the HU in real time.(4)Aiming at the problem that hyperparameters of machine learning prediction model are difficult to be efficiently tuned,the detection mechanisms of constraint,significance,and training epoch are designed,and the improved Hyperband algorithm is proposed.Based on this,the hyperparameters of the temporal pattern attention long-short term memory network are automatically tuned,and the self-optimization point prediction model of degradation trend is constructed,which improves the prediction accuracy.Further,using non-dominated sorting and elitist selection strategy,taking both the reliability and certainty as the objectives,the kernel parameters of the Gaussian process regression is optimized,the self-optimization probability prediction model of degradation trend is constructed,which solves the problems including the uncertainty description and single objective optimization of traditional prediction methods and provides decision support for the state-based maintenance strategy of the HU.