Bayesian　extreme　learning　machine　(BELM)　has　the　characteristics　of　making　full　use　of　the　prior　information　of　data　and　self-adaptive　estimation　of　model　parameters.　However,　when　the　sample　size　increases,　the　computational　efficiency　will　be　reduced　if　BELM　training　is　repeated　every　time.　To　solve　this　problem,　a　dynamic　bayesian　extreme　learning　machine　(DBELM)　method　is　proposed　for　real-time　prediction　of　deformation　monitoring　data.　This　method　takes　BELM　training　model　parameters　as　initial　values.　According　to　the　new　sample　information,　the　initial　model　parameters　can　be　updated　dynamically,　and　the　relevant　calculation　formula　is　deduced　theoretically.　The　detailed　analysis　of　simulation　data　and　actual　deformation　data　show　that　the　prediction　accuracy　of　DBELM　method　is　better　than　that　of　BELM,　RELM　and　ELM.Especially　in　the　long　term　continuous　forecast,　its　forecasting　performance　has　obvious　advantages　over　the　other　three　methods.This　fully　demonstrates　the　feasibility　and　validity　of　the　proposed　method　in　the　field　of　deformation　monitoring　data　prediction.　©　2019,　Surveying　and　Mapping　Press.　All　right　reserved.