In　order　to　speed　up　the　construction　of　energy　saving　and　emission　reduction,　and　strengthen　the　power　consumption　management　on　the　demand　side,　the　non-intrusive　load　monitoring　has　become　a　research　hotspot　because　of　its　easy　implementation　and　reliability.　However,　the　current　research　has　some　problems,　such　as　low　identification　accuracy　of　low-frequency　data　load,　complex　extraction　of　high-frequency　data　features　and　poor　network　generalization　performance.　Therefore,　this　paper　proposes　a　non-intrusive　load　monitoring　based　on　ResNeXt　network　and　transfer　learning.　The　one-dimensional　time-series　total　power　is　converted　into　two-dimensional　image　with　time　characteristics　as　input　through　Gram　angle　field　(GAF)　algorithm,　and　the　image　is　put　into　ResNeXt　network　under　transfer　learning　for　load　identification.　This　method　uses　the　low-frequency　data　that　can　be　collected　by　the　existing　meter　as　the　input,　reduces　the　data　input　dimension　and　adds　time　characteristics.　And　then,　after　the　images　are　standardized,　the　deep　load　information　is　learned　by　stacking　the　residual　neural　network　with　deep-layers,　and　the　trained　network　model　parameters　under　ImageNet-1K　dataset　are　transferred　to　the　new　target　domain　by　transfer　learning,　so　as　to　accelerate　the　convergence　speed　of　the　network,　improve　the　accuracy　of　load　classification　and　the　generalization　of　the　network.　Finally,　this　method　is　verified　by　using　the　open　data　sets　AMPds　and　UK-DALE　to　simulate　different　power　consumption　scenarios,　and　the　accuracy　is　above　99%,　which　verifies　the　efficiency　and　generalization　of　the　proposed　method.　©　2023　Automation　of　Electric　Power　Systems　Press.　All　rights　reserved.