Broad learning system (BLS) has been proposed for a few years. It demonstrates an effective learning capability for many classification and regression problems. However, BLS and its improved versions are mainly used to deal with unsupervised, supervised and semi-supervised learning problems in a single domain. As far as we know, a little attention is paid to the cross-domain learning ability of BLS. Therefore, we introduce BLS into the field of transfer learning and propose a novel algorithm called domain adaptation broad learning system based on locally linear embedding (DABLS-LLE). The proposed algorithm can learn a robust classification model by using a small part of labeled data from the target domain and all labeled data from the source domain. The proposed algorithm inherits the computational efficiency and learning capability of BLS. Experiments on benchmark dataset (Office-Caltech-10) verify the effectiveness of our approach. The results show that our approach can get better classification accuracy with less running time than many existing transfer learning approaches. It shows that our approach can bring a new superiority for BLS.