基于构造性神经网络的时间序列混合预测模型 *

针对传统时间序列预测模型不适应非线性预测而适应非线性预测的BP算法存在收敛速度慢,且容易陷入局部极小等问题,提出一种基于构造性神经网络的时间序列混合预测模型。采用构造性神经网络模型（覆盖算法）得出的类别值对统计时间序列模型的预测值进行修正,建立一种同时考虑时间序列自身周期变化和外生变量因子对时间序列未来变化趋势影响的混合预测模型,涵盖了实际问题的线性和非线性两方面,提高了预测精度。将该模型应用到粮食产量的预测中,取得了较好的预测效果。

Time-series mixed prediction model based on constructive neural networks

Traditional times-series prediction models are not adapted to nonlinear time-series prediction, and BP algorithm .which fits nonlinear time-series prediction has some trouble with slow convergence rate and easy getting into local minimum. This paper put forward a time-series mixed prediction model based on constructive neural networks. The predictions of statistical times-series models were corrected based on the different types which were calculated by constructive neural networks models(covering algorithm) . This mixed model considered both periodic changes of times-series and the influence of external variable factors on the times-series in the future. The prediction accuracies could be improved because the model were constructed fromthe nonlinear and linear aspects. The experimental results show that using this model to forecast and analysis wheat yield is effective.