With　the　continuous　consumption　of　fossil　fuels　such　as　coal,　oil　and　natural　gas,　the　environmental　energy　problem　has　become　the　focus　of　attention　in　the　world.　The　utilization　of　clean　and　non-polluting　solar　energy　for　photovoltaic　(PV)　power　generation　can　effectively　utilize　renewable　energy.　However,　the　instability　of　weather　condition　makes　the　output　of　PV　power　have　strong　randomness,　fluctuations　and　intermittence.　Therefore,　reliable　PV　power　prediction　method　can　reduce　the　disadvantages　of　PV　power　generation,　which　is　of　great　significance　to　maintenance　and　repair　of　power　plants.　In　the　study,　a　novel　hybrid　prediction　model　combining　improved　K-means　clustering,　grey　relational　analysis　(GRA)　and　Elman　neural　network　(Hybrid　improved　Kmeans-GRA-Elman,　HKGE)　for　short-term　PV　power　prediction　is　proposed.　The　proposed　model　is　established　by　using　multivariate　meteorological　factors　and　historical　power　datasets　for　two　years.　The　improved　K-means　approach　is　applied　to　cluster　the　historical　power　datasets,　and　combining　the　GRA　method　to　determine　the　similarity　days　and　the　optimal　similarity　day　of　the　forecasting　day.　The　Elman　neural　network　model　is　employed　to　better　develop　the　nonlinear　relationship　between　multivariate　meteorological　factors　and　power　data.　Compared　with　the　other　eight　prediction　methods,　the　results　show　that　the　proposed　method　has　an　outstanding　performance　on　improving　the　prediction　accuracy.