人民长江

• 论文 • 上一篇    下一篇

基于CEEMD-GRNN组合模型的月径流预测方法

赵雪花,桑宇婷,祝雪萍   

  • 出版日期:2019-04-28 发布日期:2019-04-28

Monthly runoff prediction based on CEEMD and GRNN hybrid model

ZHAO Xuehua,SANG Yuting ,ZHU Xueping   

  • Online:2019-04-28 Published:2019-04-28

摘要: 针对径流序列的噪声因素与非线性特性,采用互补集合经验模态分解法(Complete Ensemble Empirical Mode Decomposition, CEEMD)与广义回归神经网络(Generalized Regression Neural Networks, GRNN)的组合模型,对汾河上游上静游站、汾河水库站、寨上站、兰村站1958~2000年的月径流序列进行实例研究,探究3种不同建模方式下的组合模型对预测精度的影响,其中组合模型1使用加权平均集成法将各分量预测结果相加,组合模型2去除高频分量后再使用加权平均集成法将剩余分量预测结果相加,组合模型3去除高频分量后将剩余分量预测结果直接相加;再将组合模型与单一GRNN模型进行对比。结果表明:各模型的确定性系数(NS)均大于0.5,预测结果均具有可信度;不同的月径流资料适用不同的建模方法,对于极差较小的月径流序列,组合模型1预测效果最好,与另外两种组合模型及单一模型相比,平均绝对误差(MAE)、平均绝对百分比误差(MAPE)、均方根误差(RMSE)分别平均减少26%,17%,23%;对于极差较大的径流序列,组合模型2预测效果最好,与另外两种组合模型及单一模型相比,MAE,MAPE,RMSE分别平均减少30%,28%,33%。组合模型2预测误差总小于组合模型3,即加权平均集成法对提高预测精度有一定作用。三种建模过程的CEEMD-GRNN组合模型预测误差均比单一GRNN模型小,说明组合模型较单一模型更适用于月径流预测。

关键词: 月径流预测, CEEMD模型, GRNN模型, 加权平均集成法, 汾河上游

Abstract: Runoff sequence has noise factors and nonlinear characteristics. To solve these problems, a combined model was developed based on Complete Ensemble Empirical Mode Decomposition (CEEMD) and Generalized Regression Neural Networks(GRNN) and it was applied to forecast monthly runoff of the Shangjingyou station, the Fenhe reservoir station, the Zhaishang station and the Lancun station in the upper reaches of the Fenhe River from 1958 to 2000. We explored the influence of three different modeling methods on forecasting accuracy. The combined model I is to sum the predicted results of each component by weighted average integration method, the combined model II is to sum the predicted results of the residual components after removing the high-frequency components by the weighted average integration method, and the combined model III is to directly sum the component prediction results after the high frequency component is removed. The combination model is compared with a single GRNN model. The results show that the deterministic coefficient (NS value) of each model is greater than 0.5, and the predicted results are credible. Different monthly runoff data are suitable for different modeling methods. For the runoff sequences with max-min value, the combined model I has the best prediction effect and compared with the other two combined models and the single GRNN model, the mean absolute error, mean absolute percentage error, and mean square root error respectively decrease by an average of 26%, 17%, and 23%. For the runoff sequences with large max-min value, the combined model II has the best prediction effect and compared with the other two combined models and the single GRNN model, the mean absolute error, mean absolute percentage error, and mean square root error respectively decrease by an average of 30%, 28%, and 33%. The prediction error of the combined model II is always small than that of the combined model III, that is, the weighted average integration method has a certain role in improving the prediction accuracy. The model prediction error of CEEMD and GRNN combination of the three modeling processes are smaller than that of a single GRNN model, indicating the combined model is more suitable for monthly runoff prediction than single model.

Key words: Complete Ensemble Empirical Mode Decomposition, Generalized Regression Neural Networks, weighted average integration method, monthly runoff prediction, upper reaches of the Fenhe River