With　the　rapid　growth　of　installed　capacity　of　photovoltaic　power　systems,　status　monitoring　and　fault　diagnosis　of　PV　arrays　becomes　increasingly　important　for　improving　the　energy　conversion　and　maintenance　efficiency.　In　recent　years,　many　machine　learning　algorithms　were　successfully　applied　to　automatically　build　fault　diagnosis　models　using　the　fault　data　samples.　However,　most　of　them　suffer　overfitting　problem　and　the　generalization　performance　is　still　limited.　In　this　paper,　the　random　forest　(RF)　ensemble　learning　algorithm　is　explored　for　the　detection　and　diagnosis　of　PV　arrays　early　faults　(including　line-line　faults,　degradation,　open　circuit,　and　partial　shading),　which　combines　multiple　learning　algorithms　to　achieve　a　superior　diagnostic　performance.　The　proposed　RF　based　fault　diagnosis　model　only　takes　the　real-time　operating　voltage　and　string　currents　of　the　PV　arrays　as　the　fault　features,　which　is　irrelevant　to　the　environment　conditions.　In　addition,　a　grid-search　method　is　used　to　optimize　the　parameters　of　the　RF　model　by　minimizing　the　out-of-bag　error　estimation,　so　as　to　improve　the　fault　diagnosis　model.　In　order　to　obtain　sufficient　fault　data　samples,　comprehensive　fault　experiments　are　conducted　on　both　a　Simulink　based　simulated　PV　system　and　a　laboratory　PV　system.　The　simulation　and　experimental　results　both　demonstrate　that　the　optimized　RF　based　fault　diagnosis　model　can　achieve　a　high　overall　detection　and　diagnosis　performance.　Moreover,　the　comparison　results　indicate　that　the　generalization　performance　of　the　proposed　RF　based　model　is　better　than　the　one　of　the　decision　tree　based　model.　Therefore,　the　proposed　optimal　RF　based　method　is　an　effective　and　efficient　alternative　to　detect　and　classify　the　faults　of　PV　arrays.　Furthermore,　the　proposed　RF　based　fault　diagnosis　model　is　successfully　integrated　into　a　Matlab　based　real-time　monitoring　system　prototype　developed　for　the　laboratory　PV　system,　which　validates　the　practicability　as　well.　©　2018　Elsevier　Ltd