Aiming at the problems of insufficient mining of load characteristic information and large scale of identification model in current non-intrusive load identification methods, a non-intrusive load identification method based on improved V-I trajectory is proposed. Firstly, the active current, instantaneous power, and V-I_f trajectory are fused into new load features by using the Gramian angular field (GAF) and color encoding techniques. Then, the convolutional neural network (CNN) model framework is optimized through the depthwise separable convolution (DSC) module and the hybrid dilated convolution (HDC) module to construct a lightweight load identification model. Finally, experiments are conducted using public datasets for analysis. The results show that the F1 score of the proposed method is 0.953, which can further improve the identification accuracy of electrical loads while reducing the occupation of software and hardware resources.