水电能源科学

为解决非线性径流序列及单一预测模型稳定性差导致的预测精度下降问题，在“分解—预测”模型的基础上，提出了一种“优选—组合—修正”模型构建策略，首先选择DNN、SVM、LSTM、TCN、GBRT 5种模型，建立了基于EMD、CEEMDAN、VMD的15种耦合模型，并对模型进行优选；其次将优选出的模型作为基础模型，并将基础模型各时期的预测结果进行处理，输入多层感知机，构建新的组合模型；然后针对组合模型的测试期构建残差修正方程，进一步提高模型预测精度；最后将其应用于渭河流域华县站和汉江流域洋县站测试研究中。结果表明，多层感知机构建的组合模型相比单一模型有更高的预测精度，并能够结合其他模型的优点，提高模型泛化能力。通过残差修正技术的模型在各方面均优于组合模型，尤其在径流峰值拟合效果上进一步提高了预测精度。

To solve the problem of decreasing prediction accuracy caused by nonlinear runoff sequence and instability of single prediction model, this paper proposes a "selection-combination-correction" modeling strategy based on the "decomposition-prediction" model. Firstly, five models including DNN, SVM, LSTM, TCN, and GBRT are used to establish 15 coupled models based on EMD, CEEMDAN, and VMD, and the models are selected. Then, the selected model is used as the base model, and the predicted results of each period of the base model are processed and input into a multi-layer perceptron to construct a new combination model. A residual correction equation is constructed for the test period of the combination model to further improve the prediction accuracy. Finally, the method is applied to the test studies of Huaxian Station in Weihe River Basin and Yangxian Station in Hanjiang River Basin. The results show that the combination model constructed by the multi-layer perceptron has higher prediction accuracy than the single model, and can integrate the advantages of other models to improve the model's generalization ability. The model with residual correction technology is superior to the combination model in all aspects, especially in the fitting of peak discharge, further improving the prediction accuracy.