The　fault　characteristics　of　photovoltaic　(PV)　systems　are　greatly　influenced　by　environmental　factors,　which　causes　grand　challenges　in　PV　fault　detection.　Therefore,　this　paper　proposes　an　anomaly　detection　algorithm　for　grid-connected　PV　system　via　anomaly-transformer.　Firstly,　a　PV　platform　was　built　to　carry　out　fault　experiments　under　different　meteorological　conditions,　and　a　total　of　218　sets　of　DC　voltage/current　datasets　were　constructed.　Aiming　at　the　characteristics　of　multi-dimensional　time　series　data,　the　multi-branch　anomaly-attention　mechanism　is　used　to　calculate　prior-association　and　series-association,　then　use　transformer　to　reconstruct　the　loss　values　based　on　the　obtained　data.　The　association　discrepancy　is　calculated　as　the　index　of　anomaly　detection,　so　as　to　achieve　the　goal　of　time-based　localization　of　PV　faults.　The　experimental　results　show　that　compared　with　graph　deviation　network　(GDN),　unsupervised　anomaly　detection　(USAD)　and　other　algorithms,　the　Precision　of　anomaly-transformer　reaches　76.45%　and　95.41%　respectively　in　sunny　and　cloudy　test　data　sets,　and　the　F1-score　reaches　86.65%　and　97.65%　respectively.　It　can　accurately　locate　the　fault　time,　which　provides　an　effective　method　for　PV　fault　detection.　©　Beijing　Paike　Culture　Commu.　Co.,　Ltd.　2024.