基于双重注意力机制的变压器油中溶解气体预测模型

A Prediction Method for Dissolved Gas in Power Transformer Oil Based on Dual-stage Attention Mechanism

对电力变压器油中溶解气体进行预测分析可以有效掌握设备状态和发展趋势。该文以长短时记忆网络时序分析模型为基础,引入特征和时序双重注意力机制,利用参量间关联关系和时序信息依赖关系提升气体的预测准确率。特征注意力机制可以自动提取待预测气体与其他状态信息、环境和运行数据等特征之间的关联关系,为预测提供辅助信息进行适当修正,并不受传统关联规则算法的预设阈值限制;同时,利用时序注意力机制自主选取历史信息关键时间点,在LSTM时序模型基础上进一步增强关键时间点的信息表达,提升较长时间段预测效果的稳定性。在对某正常状态运行变压器甲烷气体数值预测实验中,该方法在基础LSTM模型上降低最大相对误差3%;在对某缺陷变压器的发展趋势预测中,能准确给出关键气体异常上升警戒的参考信息;具有更准确和更稳定的变压器油中溶解气体预测效果。

Predictive analysis of dissolved gas in oil could provide support for state evaluation and development trend of power transformers.A dual-stage attention mechanism was introduced to long short-term memory network as the predictive model,to extract association rules between input features and time dependencies between history time points.The feature attention mechanism was used for mining the relationship between the target parameter and other state information,environment and operation data automatically to correct prediction result appropriately,which also overcomes the limitation by the preset threshold in traditional association rule mining algorithm.The temporal attention mechanism selects the key time points in history for information enhancement over basic LSTM model,stabilizing model performance over different prediction time period.In the 100-day prediction on methane concentration of a transformer in normal status,the proposed method reduces the maximal absolute error of 3%compared to basic LSTM model.It could also provide the abnormal ascending trend of typical gasses for warning reference,if the transformer is with potential defects.A more accurate and stable result on predictions of dissolved gasses in power transformer oil was given by proposed method.