基于深度学习决策融合的非侵入式负荷分类的研究

Research on non-intrusive load classification based on decision fusion of deep learning

针对负荷分类中单一特征在负荷特征相似时的局限以及不同负荷分类模型在不同特征下的适用性存在较大差异,提出一种基于多稳态特征建模和深度学习决策融合的非侵入式负荷分类方法。首先采集原始负荷数据,从中提取电流谐波(H)、有功功率(P)、无功功率(Q)和V-I轨迹图等电气特征,然后组合P、Q、H得到PQH特征。为降低PQH组合特征之间的数量级差异,利用z-score方法将PQH特征进行归一化预处理;为减小V-I轨迹图对神经网络性能的影响,使用图像二值化方法将V-I轨迹图进行预处理。处理后的PQH特征和V-I轨迹二值图分别在LSTM模型和CNN模型上进行训练,利用决策融合方法构建负荷分类模型。通过PLAID公共数据集进行模型测试,使用准确率A、精确率P、召回率R和F1值4种评价指标和混淆矩阵验证模型效果。结果表明,决策融合模型的辨识结果(平均A、P、R和F1值分别为99.32%、96.36%、96.36%、96.34%)优于LSTM模型(平均A、P、R和F1值分别为98.57%、94.04%、92.47%、92.21%)和CNN模型(平均A、P、R和F1值分别为98.45%、92.11%、91.94%、91.94%)的辨识结果,组合特征和决策融合方法能够从多维度实现负荷分类,弥补单一特征和单一算法的不足,提高负荷分类效果。

Addressing the limitation of single feature in load classification when the load features are similar,and the greater differences in applicability of different load classification model under different characteristics,this paper proposes a non-intrusive load classification method based on multi-steady-state features and decision fusion of deep learning.Firstly,the original load data are collected and the electrical characteristics of current harmonics(H),active power(P),reactive power(Q)and V-I trajectory diagrams are extracted.Secondly,by combining P,Q,H,the PQH features are obtained.Furthermore,in order to reduce the difference in order of magnitude between features of P,Q and H in the combination,the z-score method is used to normalize the PQH feature.At the same time,the image binarization method is used to preprocess the V-I trajectory diagrams to improve the neural network performance.The preprocessed PQH features and V-I trajectory binary graph features are trained on the LSTM model and the CNN model respectively,and the two models are fused using the decision fusion method to build a load classification model.Finally,the model proposed above is tested on the PLAID public dataset,and its effectiveness is verified using confusion matrix and four existing evaluation indicators,accuracy rate(A),precision ratio(P),recall rate(R)and F1 value(F1).The results show that the identification results of the decision fusion model(the average A,P,R and F1 values are 99.32%,96.36%,96.36%,and 96.34%respectively)are better than those of the CNN model(the average A,P,R and F1 values are 98.45%,92.11%,91.94%,and 91.94%,respectively),and better than those of the LSTM model(the average A,P,R and F1 values are 98.57%,94.04%,92.47%,and 92.21%respectively).The results show that the combined feature and decision fusion method can help realize the multi-dimensional distinction of load,thus making up for the shortcomings of single feature and single algorithm and improving the performance of load classification algorithm.