In　view　of　the　difficulty　of　real-time　measurement　of　the　effluent　total　phosphorus　(TP)　for　a　wastewater　treatment　plant　(WWTP),　in　this　paper,　a　new　TP　soft　sensor　which　is　different　from　the　traditional　single　value　method　is　presented.　It　realizes　the　guaranteed　estimation　of　the　TP　concentration　by　predicting　the　upper　and　lower　bounds.　Partial　least　squares　is　used　to　obtain　the　secondary　variables　of　the　effluent　TP.　Then,　an　input-output　model　with　secondary　variables　as　the　inputs　and　the　effluent　TP　as　the　output　is　built　by　the　radial　basis　function　neural　network　(RBFNN).　Considering　the　bounded　modeling　error,　the　linear-in-parameter　set　membership　identification　algorithm　is　used　to　obtain　a　description　of　the　uncertain　set　of　the　output　weights　of　the　RBFNN.　During　the　operation　of　the　WWTP,　the　established　soft　sensor　can　predict　the　upper　and　lower　bounds　of　the　effluent　TP　concentration.　Besides,　a　bundle　of　soft　sensors　is　constructed　and　the　intersection　of　the　results　given　by　the　soft　sensors　is　used　to　reduce　the　conservativeness　caused　by　using　a　single　sensor.　The　experimental　results　show　the　effectiveness　of　the　proposed　method.