基于样本熵重构与RF-LSTM模型的水电机组状态趋势预测

State Trend Prediction of Hydropower Unit based on Sample EntropyReconstruction and RF-LSTM Model

针对水电机组运行状态预测问题,提出一种基于样本熵重构(sample entropy reconstruction, SER)与随机森林(random forest, RF)-长短期记忆网络(long short-term memory, LSTM)的混合预测模型。首先,利用改进自适应噪声完备集成经验模态分解(improved complete ensemble empirical mode decomposition with adaptive white noise, ICEEMDAN)方法,将复杂非线性振摆信号分解为一组本征模态(intrinsic mode functions, IMFs)分量;其次,采用SER原理重组具有相似复杂度的IMFs,得到多个重构特征分量(reconstruction feature components, RFCs);然后,利用随机森林预测样本熵最小的RFC,利用LSTM预测剩余的RFCs;最后,叠加各RFCs预测结果,实现水电机组状态趋势的准确预测。实验结果表明,所提方法具备更优的预测性能,可为实施机组预测性维护提供可靠的数据支持。

For the problem of state trend prediction of hydropower unit,a hybrid method based on sample entropy reconstruction(SER)mechanism and random forest long short-term memory(RFLSTM)model is proposed in this work.Firstly,the complex signal of unit can be decomposed into a group of intrinsic mode function(IMF)components by the improved complete ensemble empirical mode decomposition with adaptive white noise(ICEEMDAN)approach.Secondly,the IMFs with the similar complexity are reconstructed using the SER theory and a series of RFCs can be obtained.Then,the RFC component that has the smallest value of sample entropy is predicted by the random forest model and the remaining RFCs can be predicted by LSTM network.Finally,the predicted results of RFCs are accumulated to achieve the purpose of accurate forecasting of state tendency.The experiment results show that the proposed method in this paper has the superior prediction performance comparing with other approaches and provides the data foundation for predictive maintenance of hydropower unit.