| 引用本文: | 汤健,侯慧娟,陈洪岗,王劭菁,盛戈皞,江秀臣.基于BI-GRU改进的Seq2Seq网络的变压器油中溶解气体浓度预测方法[J].电力自动化设备,2022,42(3): |
| TANG Jian,HOU Huijuan,CHEN Honggang,WANG Shaojing,SHENG Gehao,JIANG Xiuchen.Concentration prediction method based on Seq2Seq network improved by BI-GRU for dissolved gas in transformer oil[J].Electric Power Automation Equipment,2022,42(3): |
|
| 摘要: |
| 基于门控循环单元(GRU)构建双向多层门控循环单元，并引入编码器-解码器结构搭建Seq2Seq网络模型，通过优化神经元及神经网络结构提取时序数据依赖关系。同时引入注意力机制和Scheduled Sampling算法，自动获取与当前时刻预测输出显著相关的关键输入时间点，提高长时间预测的精度。变压器正常运行状态下的气体浓度预测算例结果表明，与基于简单GRU模型及简单Seq2Seq模型的方法相比，所提方法的预测误差更低且预测的发展趋势更符合真实值；变压器异常运行状态下的气体浓度预测算例结果表明，所提方法的平均相对误差和最大相对误差相比长短期记忆(LSTM)网络方法分别降低了0.73 %和2.31%。 |
| 关键词: 电力变压器 油中溶解气体 门控循环单元 Seq2Seq 注意力机制 Scheduled Sampling算法 |
| DOI：10.16081/j.epae.202111017 |
| 分类号:TM41 |
| 基金项目:上海交通大学新进青年教师启动计划基金资助项目 |
|
| Concentration prediction method based on Seq2Seq network improved by BI-GRU for dissolved gas in transformer oil |
|
TANG Jian1, HOU Huijuan1, CHEN Honggang2, WANG Shaojing2, SHENG Gehao1, JIANG Xiuchen1
|
|
1.Department of Electrical Engineering, Shanghai Jiao Tong University, Shanghai 200240, China;2.State Grid Shanghai Electric Power Research Institute, Shanghai 200437, China
|
| Abstract: |
| Based on the GRU(Gate Recurrent Unit),the bidirectional multi-layer GRU is constructed. The encoder-decoder structure is introduced to build a Seq2Seq(Sequence to Sequence) network model. The time series data dependencies are obtained automatically by optimizing neurons and neural network structure. At the same time, the attention mechanism and the Scheduled Sampling algorithm are introduced to automatically obtain the key input time points significantly related to the prediction output at the current moment, so as to improve the accuracy of long-term prediction. In the case of gas concentration prediction under normal operating condition of transformer, compared with the methods based on simple GRU model and the simple Seq2Seq model, the proposed method has lower prediction error, and the prediction development trend is more in line with the true value. In the case of gas concentration prediction under abnormal operating condition of transformer, the average relative error and maximum relative error of the proposed model are respectively reduced by 0.73 % and 2.31% compared with the LSTM(Long Short-Term Memory) network method. |
| Key words: power transformers dissolved gas in oil gate recurrent unit Seq2Seq attention mechanism Scheduled Sampling algorithm |
