Volatility forecasting attracts extensive attentions in both finance and computation areas． However，high frequency CNY exchange rates with main stream currencies have not been thoroughly studied due to the lack of the dedicated forecasting model that can capture the dynamics of CNY rates． This paper fills the knowledge gap by，firstly，proposing a two-component hybrid volatility model based on a neural network，which is composed of a low-pass filter，the machine-learning algorithm，and the traditional autoregressive model，and secondly，studying the forecasting performance thoroughly using the one-hour and one-day realized volatility constructed from high frequency rates of six major rates: GBP /CNY，USD/CNY，AUD/CNY，EUＲ/CNY，JPY/CNY，and CHF /CNY． The predicting results are compared with component GAＲCH，EGAＲCH and neural network only models． The experimental evaluations show that our proposed model outperforms the traditional models in CNY forecasting volatility significantly and consistently across all forecasting horizons．