This　paper　proposes　a　self-organizing　cascade　neural　network　(SCNN)　for　nonlinear　system　modeling.　An　objective　function　based　on　orthogonal　least　squares　(OLS)　method　is　proposed　to　select　the　input　units　and　hidden　units.　After　the　new　hidden　unit　is　added　to　the　network,　its　input　weight　remains　unchanged　in　the　subsequent　training　process　and　the　output　weights　are　updated　in　an　incremental　way.　A　stop　criterion　based　on　test　error　is　proposed　to　select　the　optimal　network　structure.　Finally,　the　proposed　SCNN　is　tested　on　two　benchmark　nonlinear　systems　and　an　actual　problem.　The　experimental　results　show　that　the　proposed　algorithm　is　efficient.