A　continuous　deep　belief　network　(cDBN)　with　two　hidden　layers　is　proposed　to　solve　the　problem　of　low　accuracy　of　traditional　DBN　in　modeling　continuous　data.　The　whole　process　is　to　train　the　input　data　in　an　unsupervised　way　using　continuous　version　of　transfer　function,　to　design　the　contrastive　divergence　in　hidden-layer　training　process,　and　then　to　fine-tune　the　net　by　back　propagation.　Besides,　hyper-parameters　are　analyzed　according　to　stability　analysis,　as　is　given　in　the　paper,　to　make　sure　the　network　finds　the　optimal.　Experiments　on　Lorenz,　CATS　benchmark　simulation　and　CO2　forecasting　show　a　simplified　structure,　fast　convergence　speed　and　accuracy　of　this　cDBN.　Copyright　©　2015　Acta　Automatica　Sinica.　All　rights　reserved.