Intermittent demand is characterized by infrequent demand arrivals and variable demand sizes, which results in the difficult of demand forecasting. To solve this problem, a new approach was developed to forecast spare parts demand. The methodology provided mechanism to forecast the demand arrivals couple with the demand values when demand occurs. It firstly used the method of modulation to transform the 0-1 demand arrival time series into the continuous time series, and then adopted the neural network model to forecast the processed time series. Next, it applied the method of time aggregation to forecast the real demand time series, and took the rolling forecasting method into disaggregating, then got the predictive demand values. Applying this approach in forecasting the spare parts of the nuclear power equipment, the experimental results showed that the prediction accuracy was superior to Croston's method, exponential smoothing and BP neural network, which proved the methodology to be effective and accurate.