In clustered device scenarios, effective energy management is constrained by inadequate load disaggregation accuracy and unclear device operating state acquisition. To address this, this paper proposes a hierarchical Transformer model integrating the multi-head self-attention mechanism for clustered electrical equipment load disaggregation. The model comprises three core components: an input embedding layer transforming raw electrical data into high-dimensional features, a hierarchical Transformer encoder capturing multi-scale local transient patterns and global operational trends via 4 parallel attention heads and layered progressive extraction, and a classification output layer generating precise device-state probabilities. Experiments on real household datasets show strong performance: 1.56% mean absolute percentage error, 90.1% of power measurements identified within 20W, 100% accuracy for binary-state devices, and 90.24% for complex multi-state devices. It provides robust technical support for energy management systems demanding fine-grained load monitoring and rapid demand response.