Load　forecasting　is　important　to　the　efficiency　and　reliability　of　the　energy　management　systems　in　buildings.　In　general,　the　more　data　users　have,　the　greater　performance　of　load　forecasting　will　be.　However,　collecting　sufficient　data　for　load　forecasting　takes　a　lot　of　time　which　can　hardly　be　tolerated　by　users.　To　solve　this　problem,　in　this　paper,　we　propose　to　derive　the　load　forecasting　model　based　on　the　Federated　Learning　for　the　building　which　has　small　and　private　data.　The　data　are　collected　from　the　campus　energy　conservation　supervision　platform　in　Fuzhou　University.　Then　the　linear　regression　is　used　to　study　the　best　set　of　features　for　each　building.　The　experimental　results　show　that　federated　learning　can　improve　the　accuracy　of　load　forecasting,　while　the　privacy　of　each　building　is　guaranteed.　©　2021　IEEE.