Short-term　prediction　of　water　demand　provides　basic　guarantee　of　water　supply　system　operation　and　management.　In　this　study,　an　effective　model　for　daily　water　demand　forecasting　is　proposed.　Firstly,　principle　component　analysis　(PCA)　is　utilized　to　simplify　the　complexity　and　reduce　the　correlation　between　influence　variables,　and　the　score　values　of　selected　principle　components　(PCs)　turn　into　the　irrelevant　input　data　of　fuzzy　neural　network　(FNN),　which　models　the　prediction　of　water　demand.　Moreover,　an　improved　Levenberg-Marquardt　(ILM)　algorithm　is　employed　to　optimize　the　parameters　of　FNN　simultaneously.　Quassi-Hessian　and　gradient　matrices　could　be　calculated　directly　without　the　storage　and　multiplication　of　whole　Jaccobian　matrix,　therefore　the　problems　of　heavy　computing　burden　and　limited　memory　space　could　be　solved.　At　last,　contrast　experiments　are　implemented　to　demonstrate　the　fuzzy　neural　network　with　Levenberg-Marquardt　algorithm　(ILM-FNN)　has　better　prediction　performance　and　capability　to　handle　practical　issues.