The　basic　concept　of　RBF(Radial　Basis　Function)　ANN　model　based　on　OLS(Orthogonal　Least　Squares)　method　is　presented　and　the　sensitivity　of　the　traditional　OLS-RBF　model　to　the　initial　width　of　basic　function　is　analyzed.　The　gradient　descent　method　is　applied　to　adjust　and　decide　the　initial　width　of　basic　function,　which　effectively　reduces　its　impact　on　network.　With　the　wind　speed　and　environment　temperature　as　inputs,　the　power　output　of　a　coastal　wind　farm　in　Fujian　is　forecasted　by　the　improved　model　and　traditional　model　respectively,　which　shows　that,　the　improved　model　has　better　precision　and　accuracy.