面向变压器油色谱趋势预测的深度递归信念网络

Deep Recurrent Belief Network Model for Trend Prediction of Transformer Oil Chromatography Data

油色谱数据及其变化趋势是评估变压器健康状态的重要依据。现有研究表明,深度信念网络(deep belief network,DBN)在油色谱数据预测领域已取得一定成果,为变压器的运行维护提供了参考。但在实际应用过程中,仍存在因网络结构限制导致油色谱时域相关性表述不充分的情况,其预测结果呈现显著的"时移"误差,从而使得基于该方法的设备状态预测结果与实际不符。针对此问题,提出了一种面向油色谱预测的深度递归信念网络算法(deepre current belief network,DRBN),该算法构建了具有时序关联特征的深度网络结构,使预测结果呈现的"时移"误差得以消除,更新了误差的迭代修正过程,使误差在网络层间和层内得以同时流动,从而提升了预测准确率。测试结果表明,文中所提出的方法可以有效克服"时移"误差,其预测准确率可达95.16%以上,为变压器的状态预测和故障预判提供了依据。

Oil chromatography data and their variation trend provide key basis for evaluation of transformer health state. Existing studies show that deep belief network(DBN) has achieved a few results in the field of oil chromatography data prediction, providing a reference for operation and maintenance of transformers. However, in practical application, there is still insufficient expression of time-domain correlation of oil chromatography due to the limitation of model structure. Obvious "time-shift" error could be observed in chromatography prediction results, making the equipment state prediction results based on this model inconsistent with actual situation. Aiming at this problem, a deep recurrent belief network(DRBN) model for transformer state prediction is proposed based on time series theory and oil chromatography data characteristics. A self-adaptive delay network with timing correlation features is constructed, able to eliminate "time-shift" error in the prediction results. The iterative correction process of the error is updated so that the error flows simultaneously between and within network layers, thereby improving prediction accuracy. Field case studies are performed, verifying that the model proposed in this paper can availably overcome the "time-shift" error, and its prediction accuracy can reach more than 95.16%.