基于图像分类网络的非侵入式负荷辨识算法的运算成本优化

Computational cost optimization of non-intrusive load identification algorithm based on image classification network

目前基于图像分类网络的非侵入式负荷辨识算法可达到较高的辨识准确率,但存在较严重的参数冗余,引发了不必要的运算成本。对此类算法的运算成本进行优化,提出一种基于灰色编码的设备特征组合方法,以减少算法中设备特征的参数冗余;然后使用轻量级图像分类网络ZFNet构建设备辨识模型,并引入Inception模块来减少模型中卷积层输出的参数冗余,同时基于仿真实验结果对模型中全连接层的结构和参数进行适应性调整,以减少模型的参数冗余,最后使用PLAID数据集进行算例分析。结果表明:相比于同类算法,所提算法在设备特征的参数量上减少了66.7%~67.5%,在模型的参数量上减少了90%~97.1%,在整体运算量上的变动为-91.7%~6.1%。

At present,the non-intrusive load identification algorithm based on image classification network can achieve high identification accuracy,but there is serious parameter redundancy,which leads to unneces⁃sary calculation costs.The computational cost of this kind of algorithm is optimized.Firstly,a device fea⁃ture combination method based on grey coding is proposed to reduce the parameter redundancy in device features.Then,the lightweight image classification network ZFNet is used to construct the device identifica⁃tion model,and inception submodule is introduced to reduce the parameter redundancy in the output of the convolutional layers in the model.At the same time,based on the simulative results,the structure and parameters of the fully connection layers in the model are adjusted adaptively to reduce the parameter redundancy of the model.Finally,PLAID data set is used for example analysis,and the results show that,compared with similar algorithms,the proposed algorithm reduces the parameter number of equipment fea⁃tures by 66.7%~67.5%,the parameter number of the model by 90%~97.1%,and the change of the overall calculation amount is-91.7%~6.1%.