基于SSA-MGF的BP神经网络多步预测模型

BP Neural Network Multi-Step Prediction Model Based on SSA-MGF

采用奇异谱分析(Singular Spectrum Analysis,SSA)方法对标准化样本序列进行准周期信号分量重建,将重建序列构造均值生成函数(Mean Generating Function,MGF)延拓矩阵作为输入因子,原样本序列作为输出因子,构建BP神经网络多步预测模型。通过实际建模并与逐步回归等方法进行对比预测试验,结果表明,基于SSA-MGF的BP神经网络多步预测模型预测效果优于其他3种模型,说明SSA的去噪及BP神经网络预报模型对于提高预测准确率是相对有效的,是一种具有较高应用价值的多步预测方法。

Based on Singular Spectrum Analysis （SSA） method the standardized sample series have been reconstructed using the quasi-period signal weight. A BP neural network multi-step prediction model with the Mean Generating Function（MGF） prolongation matrix of the reconstruction series as the input factor, and the original sample series as the output factor has been formulated. Results show that the model is superior in predictions over the other three models, indicating that denoising by SSA and constructing BP neural network prediction model are relatively effective for improving forecast accuracy. Therefore, the model is an useful one for the actual operational forecasting.