This paper proposes a multi-scale non-intrusive industrial load decomposition method considering scarce load characteristics. The total power consumption is regarded as the load consumption data of a virtual electrical equipment

and the working state of the virtual electrical equipment is taken as a load decomposition feature to increase the types of input features. To overcome the non-equilibrium problem of industrial load data

an adaptive density peak clustering algorithm is used to determine the working state of the total power sequence

and the Gram angle field method is used to transform the active power sequence and the working state sequence into two-bit image features. To improve the accuracy of load decomposition

the dilated convolution is used to construct the load decomposition model

expand the receptive field of the convolution kernel

and extract the load characteristics of different scales. Three scenarios are designed to evaluate the effectiveness of the model.