Abstract: This paper proposes a non-intrusive cross-domain load identification method comprising the complete process of event detection, feature extraction, domain adaptation, and cross-domain load identification. First, an event detection algorithm is employed to identify user load switching events from high-frequency voltage and current signals. Then, steady-state features, harmonic features, V-I trajectory shape features, and transient features are extracted to build a comprehensive feature library for each load. A one-dimensional deep subdomain adaptive network is constructed, in which the labeled load feature data serve as the source domain and the unlabeled data of the loads to be identified serve as the target domain. The distribution differences between the domains are quantified using local maximum mean discrepancy, and parameter alignment is performed during training to achieve effective cross-domain load identification.