To　meet　the　requirements　of　the　daily　management　of　water　supply　systems　for　short-term　water　demand　prediction　timeliness,　a　kernel-based　extreme　learning　machine　model　(KELM)　was　established,　which　requires　short　training　time.　From　the　perspective　of　improving　the　prediction　accuracy,　a　residual　correction　module　based　on　the　Fourier　series　(FS)　was　constructed,　which　was　used　to　model　the　difference　between　the　initial　predicted　value　and　the　observed　value　of　water　demand,　and　the　residual　correction　of　the　initial　predicted　value　was　completed.　The　module　was　superimposed　on　the　KELM　model　to　form　the　hybrid　prediction　model　(KELM+FS).　The　performance　of　the　models　was　tested　using　real　water　demand　data.　Experimental　results　show　that　the　KELM　model　could　produce　similar　prediction　accuracy　as　the　artificial　neural　network　model　and　the　support　vector　regression　model,　but　the　prediction　time　was　only　about　5%　of　the　average　time　of　the　two　models.　Compared　with　the　KELM　model,　the　relative　prediction　accuracy　of　the　hybrid　model　KELM+FS　was　improved　by　about　12%　without　significantly　increasing　the　prediction　time.　Therefore,　when　applied　to　short-term　water　demand　prediction,　both　the　single　model　KELM　and　the　hybrid　model　KELM+FS　could　achieve　the　goal　of　improving　the　prediction　efficiency.　Copyright　©2022　Journal　of　Harbin　Institute　of　Technology.All　rights　reserved.