Water　quality　prediction　is　an　important　research　focus　in　smart　water　and　can　provide　the　support　to　control　and　reduce　water　pollution.　However,　existing　water　quality　prediction　models　are　mainly　data-driven　and　only　rely　on　various　sensor　data.　This　paper　proposes　a　new　water　quality　prediction　modeling　approach　integrating　data　and　knowledge.　We　develop　a　water　quality　prediction　framework　that　combines　knowledge　graph　and　deep　adversarial　networks.　The　knowledge　extraction　and　management　compound　extracts　the　water　quality　knowledge　graph　from　different　knowledge　sources　by　using　the　deep　adversarial　joint　model.　The　fusing　parameter　importance　learning　compound　calculates　the　contribution　of　parameters　in　water　quality　prediction　by　taking　into　account　both　knowledge　and　data　levels　of　correlation.　Finally,　a　water　quality　prediction　model　combining　weighted　CNN-LSTM　with　adversarial　learning　predicts　the　values　of　total　nitrogen　based　on　real-time　monitoring　data.　The　experimental　results　on　monitoring　data　from　the　Juhe　River　of　China　show　that　the　proposed　model　can　greatly　improve　the　effect　of　water　quality　prediction.