基于局部加权周期趋势分解算法和注意力机制的变压器顶层油温多步预测  

Multi-step Prediction of Transformers Top Oil Temperature Based on STL and Attention Mechanism

作　　者：王德文[1] 吕哲 WANG Dewen;LV Zhe(School of Control and Computer Engineering,North China Electric Power University,Baoding 071003,China)

机构地区：[1]华北电力大学控制与计算机工程学院,河北保定071003

出　　处：《电力科学与工程》2022年第11期1-8,共8页Electric Power Science and Engineering

基　　金：河北省自然科学基金(F2021502013)。

摘　　要：首先,应用局部加权周期趋势分解算法(seasonal and trend decomposition procedure based on loess,STL),将变压器顶层油温分解成趋势、周期和残差分量;然后,使用一维卷积网络和编码器–解码器提取特征,生成特征矩阵;最后,引入注意力机制挖掘特征矩阵中对当前预测结果产生显著影响的信息,并随预测时间更新,最终得到多步预测结果。算例分析表明,与传统预测方法相比,该方法能够有效提取顶层油温数据特征并缓解预测时间增长带来的预测误差累积,具有更高的多步预测精度。First,the top oil temperature of transformer is decomposed into trend,period and residual components by using the seasonal and trend decomposition procedure based on loess decomposition algorithm.Then one-dimensional convolutional network and encoder-decoder are used to extract features and generate a state matrix.Finally,the attention mechanism is introduced to mine the information in the state matrix which has a significant impact on the current prediction results,and the multi-step prediction results are obtained by updating with the prediction time.The example analysis shows that compared with the traditional prediction methods,this method can effectively extract the top oil temperature data features and alleviate the accumulation of forecasting errors caused by the increase of prediction time,and has higher multi-step prediction accuracy.

关 键 词：电力变压器 顶层油温 局部加权周期趋势分解 注意力机制 编码器–解码器 多步预测 

分 类 号：TM411[电气工程—电器]