The　high　accuracy　prediction　model　is　the　basis　to　investigate　the　organic　Rankine　cycle　(ORC)　system　performance.　Compared　with　the　traditional　thermodynamic　model,　the　data-driven　model　of　ORC　system　based　on　artificial　neural　network　(ANN)　has　obvious　advantages　in　reflecting　the　strong　coupling　characteristics　of　the　system.　The　accuracy　of　ORC　system　prediction　model　depends　on　the　training　data,　but　the　outlier　removal　from　the　training　data　has　not　been　fully　studied.　This　paper　proposes　an　unsupervised　learning　approach　for　outlier　removal　in　ORC　system.　Based　on　this　approach,　the　nonlinear　variation　relationship　between　operating　parameters　and　system　performance　is　analyzed.　The　approach　is　further　compared　with　the　common　outliers　removal　criteria.　In　addition,　reasonable　selection　of　input　variables　is　the　basis　for　the　construction　of　ORC　system　prediction　model,　but　commonly　used　selection　process　cannot　effectively　filter　out　the　redundant　and　irrelevant　features.　A　hybrid　feature　selection　algorithm　is　presented　based　on　Fourier　transform　and　partial　mutual　information.　The　effectiveness　of　the　proposed　algorithm　is　compared　with　principal　component　analysis.　A　framework　for　ORC　system　outlier　removal　and　feature　dimensionality　reduction　is　proposed.　The　results　show　that　the　use　of　this　framework　can　significantly　improve　the　prediction　accuracy　of　the　model.　The　MAPE　and　MSE　of　the　model　are　6.4　x　10(-3)%　and　3.53　x　10(-11),　respectively.　This　framework　can　provide　a　direct　reference　for　the　construction　of　data-driven　ORC　prediction　model.　(C)　2022　Elsevier　Ltd.　All　rights　reserved.