基于自编码压缩与多尺度特征提取的抽水蓄能机组劣化趋势评估与预测

恶劣的运行环境为抽水蓄能机组安全运行带来严峻挑战,抽水蓄能机组劣化趋势评估与预测技术能够有效反映机组运行状况并预测机组未来劣化情况,为机组状态检修提供重要依据。然而,机组运行工况参数中存在大量冗余或干扰信息,严重影响劣化趋势评估的可靠性;此外,难以对复杂的劣化趋势序列实现准确的预测。为解决上述问题,提出一种基于自编码压缩与多尺度特征提取的抽水蓄能机组劣化评估预测模型。首先,为降低拟合误差,利用深度自编码器(DAE)凝练工况参数中的关键信息,结合多层感知机(MLP)建立健康模型;其次,根据机组运行数据与健康模型,生成机组劣化度;最后,以一维卷积神经网络(1DCNN)提取局部空间特征,以双向门控循环单元(BiGRU)提取双向全局时序特征,结合二者的优势,构建多尺度特征提取网络,实现精确的劣化趋势预测。通过某抽水蓄能机组验证了该模型的有效性。与其他模型相比,自编码压缩模型的拟合误差最低,能够生成可靠的劣化趋势;多尺度特征提取网络能够学习劣化趋势序列中的长期趋势与局部波动信息,预测精度更高。

Degradation trend assessment and prediction of pumped storage unit based on deep auto-encoder compression and multiscale feature extraction

The harsh operating environment brings challenges to the safe operation of the pumped storage unit (PSU).The degradation trend assessment and prediction of PSU can effectively reflect operation state of the unit.However,there is a large amount of redundant and interfering information in operating parameters of the unit,which seriously affects the reliability of the assessment.Besides,it is hard to achieve accurate prediction for complex degradation trend sequences.To solve above problems,a degradation trend assessment and prediction model of PSU is proposed on the basis of deep auto-encoder compression and multiscale feature extraction.Firstly,the healthy model is built using deep auto-encoder (DAE) and multilayer perceptron (MLP) to reduce the fitting error,where DAE is adopted to condense the critical information.Secondly,the degradation trend is generated based on the healthy model.Finally,a multiscale feature extraction network is constructed by combining the advantages of one-dimensional convolutional neural network (1DCNN) and bi-directional gated recurrent unit (BiGRU) for the accurate prediction.Compared with other models,the proposed healthy model achieved the lowest fitting error and the multiscale feature extraction network has the highest prediction accuracy.