The　precision　of　conventional　PV　fault　diagnostic　methods　faces　challenges　due　to　the　nonlinear　output　power　characteristics　of　photovoltaic　(PV)　arrays　and　the　implementation　of　the　maximum　power　point　tracking　(MPPT)　algorithm.　In　severe　cases,　it　may　lead　to　power　losses　and　even　arouse　safety　issues.　In　this　study,　the　variation　characteristics　of　the　sequence　waveforms　at　the　moment　of　failure　are　investigated　and　used　to　develop　a　novel　PV　fault　diagnostic　framework.　Firstly,　the　sequence　waveforms　of　string　voltages　and　currents　before　and　after　the　fault　occurred　are　collected;　the　normalized　sequence　data　of　voltages,　currents,　and　powers　are　used　as　analytic　data.　Then,　the　fault　feature　extraction　is　realized　via　a　stacked　autoencoder　(SAE)　model.　After　that,　an　improved　multi-grained　cascade　forest　(IgcForest)　is　proposed　to　diagnose　faults,　e.g.,　line-to-line　(L-L)　fault,　open-circuit　(OC)　fault,　partial-shading　of　PV　arrays,　etc.　The　advantages　of　the　proposed　method　are　that　the　SAE　method　to　extract　features　with　higher　recognition　automatically,　and　the　IgcForest　to　enhance　and　exploit　fault　features.　Particularly,　the　proposed　improvements　can　reduce　the　feature　vector　dimension　and　enhance　the　information　connectivity　between　forests　at　all　levels　for　further　improving　the　accuracy　of　diagnoses.　In　addition,　the　validity　of　the　proposed　method　is　verified　by　numerical　simulations　and　measured　data,　and　the　corresponding　accuracy　of　fault　diagnoses　for　single　failure　reach　99.33%　and　98.61%,　respectively,　which　are　superior　to　traditional　methods,　such　as　softmax,　support　vector　machines,　random　forest,　gcForest,　and　daForest.　Furthermore,　it　also　has　a　high　accuracy　of　98.83%　for　data　sets　with　the　occurrence　of　multiple　faults.