Mirror arrays have been applied to indoor visible light communication (VLC) as a passive the reconfigurable intelligent surface (RIS) which has no signal processing to solve the problem of indoor visible light line-of-sight obstruction, however, after channel modeling, it is found that the reflected channel in this scenario has a serious multipath effect, to this end, we introduce deep learning techniques into channel estimation of VLC systems with mirror arrays for the first time, and propose a hybrid new model of Transformer and the bidirectional longshort-term memory model (Transformer-B