The　measurement　data　of　power　systems　are　often　mixed　with　a　lot　of　noise　due　to　the　interference　of　the　external　environment.　In　order　to　eliminate　the　effect　of　noise,　it　is　significant　to　denoise　the　noisy　data　to　obtain　the　real　state　measurements.　In　order　to　deal　with　the　problem　of　insufficient　interpretability　in　existing　data-driven　denoising　methods,　a　hybrid　physical-data　driven　denoising　model　(PDDM)　based　on　the　stacked　denoising　autoencoder　(SDAE)　is　proposed.　First,　the　previous　knowledge　is　extracted　from　the　physical　model　of　the　generator.　Physical　constraints　are　designed　based　on　the　inherent　relationships　between　rotor　angle,　angular　frequency,　and　power.　Second,　based　on　SDAE　deep-learning　(DL)　model,　physical　constraints　are　embedded　into　the　loss　function　to　guide　the　training　of　a　neural　network.　The　derivatives　of　denoised　data　are　leveraged　in　anticipation　of　satisfying　the　differential-algebraic　equations.　The　physical　process　is　directly　approximated　by　the　neural　network　in　this　method,　making　the　outputs　satisfy　the　physical　laws.　The　reliability　and　interpretability　of　the　denoising　results　are　improved.　Meanwhile,　the　dependence　on　datasets　is　reduced　by　virtue　of　the　hybrid　physical-data　driven　mode.　The　robustness　is　still　maintained.　Finally,　it　is　verified　in　the　39-bus　New　England　system　and　a　realistic　regional　power　system.　The　real　noisy　data　are　also　taken　into　account　in　testing　to　verify　its　extensibility.　The　test　results　show　that　the　method　proposed　can　achieve　a　satisfactory　effect　in　both　denoising　accuracy　and　generalization　capability.