Solar　power　generation　(SPG)　is　essentially　dependent　on　spatial　and　meteorological　characteristics　which　makes　the　planning　and　operation　of　power　systems　difficult.　To　promote　the　integration　of　solar　power　into　electric　power　grid,　accurate　prediction　of　geographically　distributed　SPG　is　needed.　In　this　paper,　we　present　a　combined　method　for　day-ahead　SPG　prediction　of　multi-region　photovoltaic　(PV)　plants.　First,　automatic　machine　learning　(AML)　is　applied　to　generate　the　most　suitable　ensemble　prediction　model　with　optimal　parameters　and　then　an　improved　genetic　algorithm　(GA)　is　implemented　which　processes　the　candidate　features　by　assigning　appropriate　operators.　To　achieve　more　accurate　forecast　results　as　well　as　mine　the　interpretable　relationship　between　SPG　and　related　weather　or　PV　system　factors,　the　SPG　physical　model　is　taken　into　account.　The　method　performance　is　evaluated　by　the　real　SPG　data　along　with　meteorological　variables　of　multi-region　PV　plants　in　Hokkaido　from　2016　to　2018.　Results　indicate　that　the　combined　method　provides　acceptable　accuracy　and　outperforms　several　baselines　and　other　methods　used　for　comparison.　(c)　2021　Elsevier　Ltd.　All　rights　reserved.