The　construction　of　the　ubiquitous　power　internet　of　things　(UPIoT)　provides　a　new　feasible　solution　for　gas-insulated　switchgear　(GIS)　online　monitoring　and　fault　diagnosis,　but　it　also　puts　forward　greater　requirements　for　time　and　accuracy.　How　to　find　an　effective　real-time　model　that　can　be　applied　to　the　UPIoT　mobile　terminals　has　become　an　urgent　problem　needing　to　be　solved.　To　this　end,　this　study　proposes　a　lightweight　convolutional　neural　network　(LCNN)　for　GIS　partial　discharge　(PD)　pattern　recognition　using　three　lightweight　convolutional　blocks,　and　introduces　the　lowest　recognition　accuracy　of　single-class　faults　as　the　primary　indicator　for　selecting　the　optimal　model　under　the　UPIoT.　First,　three　lightweight　convolutional　blocks　are　introduced　for　constructing　an　LCNN.　Then,　the　optimal　model　constructed　by　the　lightweight　blocks　is　sought.　Next,　criteria　for　determining　the　best　model　are　introduced,　and　the　best　model　under　the　UPIoT　is　selected.　This　study　provides　a　reference　standard　for　the　construction　of　GIS　PD　pattern　recognition　under　the　UPIoT.　Meanwhile,　through　the　balance　of　evaluation　indicators,　this　study　verifies　that　the　minimum　recognition　accuracy　of　the　MnasNet　model　is　98.8%,　which　is　obviously　better　than　other　methods　and　lays　a　solid　foundation　for　GIS　PD　pattern　recognition.　©　The　Institution　of　Engineering　and　Technology　2020