The　grid-connected　photovoltaic　(PV)　power　stations　are　instability　and　volatility　due　to　meteorological　factors.　A　way　to　improve　this　problem　is　PV　power　forecasting.　This　paper　proposed　an　improved　short-term　PV　power　prediction　model　that　combines　an　extreme　learning　machine　(ELM)　neural　network　and　similar　day　method.　Firstly,　a　narrow-band　Internet　of　Things　(NB-IoT)　intelligent　combiner　box　data　monitoring　system　is　designed　to　collect　multivariate　meteorological　factors　and　original　PV　output　power　datasets　in　different　seasons.　Secondly,　the　corresponding　training　set　is　selected　according　to　the　season　type　of　the　forecast　day,　and　the　Euclidean　distance　(ED)　between　the　training　set　and　the　forecasting　day　is　calculated,　and　the　M-day　with　a　small　Euclidean　distance　is　selected.　Then,　the　N-day　similar　day　data　is　divided　among　the　M　days　as　the　new　training　set　input,　and　the　P-day　optimal　similar　day　data　and　the　multivariate　meteorological　of　the　prediction　day　are　divided　as　test　set　inputs.　Finally,　the　ELM　neural　network　prediction　model　is　used　to　predict　the　output　power　of　the　predicted　day.　The　experimental　results　show　that　the　proposed　method　has　the　highest　prediction　accuracy　in　contrast　to　other　two　prediction　models.　©　Published　under　licence　by　IOP　Publishing　Ltd.