基于运行效率分布差异的水电机组劣化状态趋势预测

Deteriorated State Trend Prediction of Hydropower Units Based on Difference of Operation Efficiency Distribution

为解决水电机组劣化状态难以刻画及预测精度低的问题,需深入探究不同机组状态下运行效率的分布差异特性,提出了一种基于运行效率分布差异的水电机组劣化状态趋势预测方法。首先,综合考虑水电机组工况(水头、流量)与效率之间映射关系和状态监测数据随机性,利用高斯混合模型拟合多工况下机组健康状态运行效率的概率分布特性;在此基础上,计算观测样本在机组健康状态分布下的负对数似然概率,并以此作为水电机组劣化状态指标,表征观测样本与机组健康状态标准分布之间的偏差;进一步采用非因果原理和高斯误差线性单元,分别改进时间卷积网络(TCN)的膨胀卷积模块和残差模块,并融合门控循环单元(GRU),设计并构建水电机组劣化状态预测模型;最后,利用某水电站#6机组实际运行监测数据开展方法验证。结果表明,所提方法能有效提升机组劣化状态趋势预测精度。

In order to solve the problems of difficulty in characterizing the deteriorated state of hydropower units and low prediction accuracy,it is necessary to deeply explore the distribution difference characteristics of operating efficiency under different states.This paper presents a deteriorated state trend prediction of hydropower units based on the difference of operating efficiency distribution.Firstly,considering the mapping relationship between the operating conditions(head and flow)and efficiency of hydropower units,and the randomness of the state monitoring data,Gaussian mixture model is used to fit the probability distribution characteristics of the units operating efficiency under multiple operating conditions.On this basis,the negative log-likelihood probability of the observed samples under the units health state distribution is calculated,which is used as an index of the deteriorated state of the hydropower units to characterize the deviation between the observed samples and the standard distribution of the units health state.Furthermore,the expansion convolution module and the residual module of the time convolution network are improved respectively by using the noncausal principle and the Gaussian error linear units,and the gate recurrent units is fused to design and build the deteriorated state prediction model of the hydropower units.Finally,the proposed method is verified by using the actual monitoring data of Unit#6 in a hydropower station.The results show that the proposed method can effectively improve the trend prediction accuracy of deteriorated state.