To　address　the　issue　of　insufficient　harmonic　measurements　on　the　low-voltage　side　of　distribution　transformers,　this　study　proposes　a　deep　learning-based　method　for　estimating　harmonics　on　the　low-voltage　side　of　distribution　transformers　by　combining　short-term　test　data　and　long-term　power　data.　Initially,　a　method　combining　Fisher＇s　optimal　segmentation　and　the　derivative　dynamic　time　warping　algorithm　is　used　to　identify　the　harmonic-dominant　users.　Subsequently,　a　data　transformation　method　combining　variational　mode　decomposition　and　Gramian　angular　field　is　proposed　to　convert　the　power　signal　of　harmonic-dominant　users　and　the　harmonic　signal　on　the　low-voltage　side　of　distribution　transformers　into　pseudo-color　Gramian　power　images　and　gray　Gramian　harmonic　images.　Finally,　these　two　types　of　images　are　input　into　an　improved　PSRGAN　(pix2pix　-Super-resolution　generative　adversarial　network)　model　for　training　to　learn　the　mapping　relationship　between　the　power　data　of　the　harmonic　source　user　and　the　harmonic　data　on　the　low-voltage　side　of　distribution　transformers,　enabling　the　generation　of　long-term　monitoring　data　for　harmonics　on　the　low-voltage　side　of　distribution　transformers.　The　accuracy　of　the　proposed　method　is　validated　through　simulation　models　and　real　measurement　cases,　and　the　required　data　are　easily　obtainable,　demonstrating　the　practicality　of　the　method　for　engineering.　©　2025　Chinese　Society　for　Electrical　Engineering.　All　rights　reserved.