To enhance the accuracy of load identification for multi-energy devices in non-invasive integrated energy systems, a method based on an improved sliding window bilateral CUSUM (cumulative sum) , GBSSL (graph semi-supervised learning) , and ICapsNet(improved capsule network) was proposed, considering the spatiotemporal coupling characteristics of multi-energy loads. Initially, an adaptive noise value selection method was introduced to enhance the sliding window bilateral CUSUM algorithm, and the improved algorithm was used for event detection, followed by labeling unlabeled samples through GBSSL. Subsequently, on the basis of CapsNet, the similarity and weighted sum calculation methods were improved, the residual block structure convolutional network replaces the original convolutional module, and a polarization self-attention block was introduced into the main capsule module to construct ICapsNet. Finally, different non-invasive load identification methods are used to identify the loads of 10 150 integrated energy load data collected, verifying the superiority of the proposed method. Experimental results show that compared to mainstream noninvasive load identification methods such as BI-GRU, Bagging EL, and DNN, the P_{recisi on}, R_ecall, F_macro, and BA metrics are improved by an average of 1. 77%, 2. 14% , 1. 94% , and 1. 26% , respectively. The results indicates that the proposed method can accurately identify the loads of multi-energy devices in non-invasive integrated energy systems, with good computational efficiency and versatility.