基于CNN-BiLSTM模型的机车变压器油中溶解气体浓度预测方法

Prediction Method of Dissolved Gas Concentration in Transformer Oil Based on CNN-BiLSTM Model

针对现有变压器油中溶解气体浓度预测方法未考虑气体参量之间存在的关联关系,以及模型训练时未考虑数据的时序关联性,从而导致预测精度不高的问题,提出一种基于卷积神经网络-双向长短期记忆模型的变压器油中溶解气体浓度预测方法。利用卷积神经网络提取变压器油中溶解气体参量之间的关联关系,利用双向长短期记忆网络提取变压器油中溶解气体浓度数据中的时序特征,通过全连接层实现变压器油中溶解气体浓度预测输出,同时引入Dropout方法进行过拟合处理,提高泛化能力。以某机务段HXD_(1C)型电力机车变压器最近2年油中溶解气体检测数据进行试验,结果表明,利用本文提出的方法对油中溶解气体浓度进行预测,平均绝对百分比误差降到了2.41%,比CNN-LSTM方法降少1.55%。

For the problem of low accuracy about existing method for predicting the concentration of dissolved gas in transformer oil,which didn’t consider the relationship between the gas parameters and the time series correlation of the data did not consider when training the model,this paper proposed a method for predicting the concentration of dissolved gas in transformer oil based on the convolutional neural network-bidirectional long short-term memory model.Firstly,the correlation between the parameters of the dissolved gas in the transformer oil were extracted by using convolutional neural network.Secondly,the time series features in the dissolved gas concentration data in transformer oil were extracted by using bidirectional long short-term memory network.Finally,the prediction of dissolved gas concentration in transformer oil were predicted through the fully connected layer.at the same time,Dropout was used to prevent overfitting.The experiment was done based on the detection data of dissolved gas in transformer oil of HXD_(1C)electric locomotive in a locomotive depot in recent 2 years,the results show that using the method proposed in this paper to predict the concentration of dissolved gas in oil,Mean absolute percentage error is reduced to 2.41%,and it is 1.55% lower than the convolutional neural network-bidirectional long short-term memory method.