A nonlinear combination forecasting model was established by using neural network and accelerating genetic algorithm (AGA) in the paper. AGA was used to optimize the network parameters as BP approach was slow with training network. Optimization results of AGA were taken as original values of BP approach, the network was trained with BP approach. Network convergence rate was increased with running BP approach and AGA alternately. Meanwhile the part least problem was improved. Examples were presented finally, as a result, the forecasting precision high in evidence.