基于国家粮食安全预警系统的开发项目,针对我国粮食年产量预测中精度差和波动大的问题,分析了逐步回归、BP神经网络和GM(1,N)灰色系统3种常用预测方法的预测能力。根据能够计量和具有农学意义2个原则。选择了粮食作物播种面积、化肥施用量、粮食作物有效灌溉面积等12个重要的粮食年产量影响因子,用上述3种方法构建预测模型。在建模样本相同的情况下,结果显示,BP神经网络方法5年期拟合平均相对误差为1.44%,连续5年逐年预测平均相对误差可达到2.89%,这2个性能均优于其他2种方法,可以较好地应用于粮食安全预警系统.笔者最后探讨了对BP神经网络进一步优化的方法。
Based on the project of National Grain Warning System, aiming at predicting the grain output of China, this paper has compared and analyzed forecasting performances of three methods, namely step regression, BP neural network and GM(1, N) gray system. According to the principle of calculable and having agricultural significance, we chose twelve important effecting factors, and established respective forecasting model with the above three methods. Results showed that the average error of the method of BPNN was 1.44% and its average forecast error on five years could reach 2.89%, which is better than the other two methods in performances. It can be used in the project of National Grain Warning System. Finally the paper lists feasible methods to optimize the BPNN farther.