The　physical　equation　is　less　integrated　into　the　data-driven　model　in　existing　frequency　prediction　methods.　A　minimum　frequency　prediction　model　based　on　physical-data　driven　is　proposed　in　this　paper.　Firstly,　the　physical　information　embedded　layer　is　constructed.　The　minimum　frequency　expression　of　the　system　frequency　response　(SFR)　model　is　transformed　into　the　input　layer　of　the　long　short-term　memory　(LSTM)　network.　The　physical　knowledge　in　the　SFR　model　can　be　explored　through　neural　network　training.　And　the　physical　knowledge　is　used　to　improve　the　prediction　accuracy　of　the　LSTM　network.　Then,　the　steady-state　frequency　constraint　is　added　to　the　loss　function.　The　search　space　of　the　embedded　layer　parameters　is　directly　constrained.　With　the　constrain,　the　parameter　drift　problem　can　be　effectively　addressed.　Finally,　the　simulation　verification　is　performed　in　the　IEEE　39-node　system.　The　results　show　that　the　proposed　model　has　higher　prediction　accuracy　than　common　fusion　models.　When　the　data　contains　noise,　the　proposed　model　shows　good　noise　immunity.　©　2023　IEEE.