DFA_VMD_LSTM组合日径流预测模型研究

为有效提取径流序列的局部特征信息、提高神经网络径流预测模型的非线性拟合能力和预测性能,引入变分模态分解(VMD)、去趋势波动分析(DFA)方法,提出了一种基于长短时记忆(LSTM)神经网络的组合日径流预测模型(DFA_VMD_LSTM),并采用均方误差(RRMSE)、平均绝对误差(MMAE)、平均绝对百分误差(MMAPE)以及确定性系数(DDC)统计指标对模型进行评价。在三峡水库的径流预测研究中,经过与其他三种典型数据驱动模型的预测结果对比发现,DFA_VMD_LSTM组合日径流预测模型在不同评价指标上均有显著提升,说明该模型可充分挖掘径流序列组成特性,学习历史长程依赖,能有效提高径流预报精度。

Study on DFA_VMD_LSTM Hybrid Daily Runoff Forecasting Model

In order to effectively extract the local feature information of runoff series and improve the nonlinear fitting ability and prediction performance of the neural network runoff prediction model,this paper uses variational mode decomposition(VMD),detrended fluctuation analysis(DFA),and proposes a hybrid daily runoff prediction model(DFA_VMD_LSTM)based on long short-term memory(LSTM)neural network.The root mean squared error(RRMSE),mean absolute error(MMAE),mean absolute percent error(MMAPE),and determination coefficient(DDC)were used as the criteria for evaluation of models.Compared with the prediction results of the other three typical models in the study of runoff prediction for the Three Gorges Reservoir,it is found that the DFA_VMD_LSTM hybrid daily runoff prediction model has significantly improved on different evaluation indicators.The results indicate that the model can fully exploit the characteristics of runoff sequence composition,learning history long-range dependence and improve the accuracy of runoff forecasting effectively.