Research On Runoff Prediction In Han River Basin Based On Combined Machine Learning Model

Influenced by climate,environment and human activities,the formation mechanism and evolution pattern of river runoff have changed,posing a great challenge to the existing runoff prediction models.Most of the previous studies focus on the validation of the model performance,not bring the prediction model closer to the actual problem.The rise of cross-disciplines and the development of combined machine learning models provide a new paradigm for runoff prediction problems.In this paper,the measured monthly runoff from four hydrological stations in the Han River basin,namely,Wuhou Town,Yangxian,Huangjiagang and Huangzhuang,is used as the study object.With a balance between model prediction accuracy and practical application,a combined machine learning model approach for runoff prediction is proposed.The main research results obtained in this paper are as follows:(1)A set of runoff data pre-processing methods based on big data analysis and signal decomposition methods are proposed to deeply mine runoff data in terms of missing values,training set test set decomposition,normalization,etc.A "decomposition-prediction" strategy is proposed to transform the measured runoff data into predictors available to the model.(2)Machine learning and deep learning techniques are used to integrate different algorithmic models,and VMD-CNN-LSTM(VCL),VMD-LSTM(VL),VMD-PSO-SVR(VPR),and VMD-GSA-SVR(VGR)integration models are established and evaluated in comparison with the integration models under both EMD and EEMD methods.The empirical validation shows that the four integrated models have the best prediction performance compared to the comparison models,and are suitable integrated models for runoff prediction,which can be used as the base model for combined model prediction.(3)Based on the four base models of VCL,VL,VPR,and VGR,four combined models of VCL-VGR,VCL-VPR,VL-VGR,and VL-VPR are established,and a new training set and test set are generated according to the stacked generalization(SG)strategy,which are applied to the prediction of monthly runoff at Yangxian station.The results show that both VCL-VGR and VCL-VPR combined models have a certain degree of improvement in prediction performance compared to their base models.By analyzing the two constraints of the base model screening combination,the threshold value of the base model screening combination based on the difference of Nash coefficients(NSE)is set,which effectively reduces the computational cost.(4)A set of machine learning combined model rolling prediction method is proposed by synthesizing previous research methods,and the measured monthly runoff from Huangzhuang,Huangjiagang and Wuhou town stations in the Han River basin is used for example validation to compare the prediction performance of the combined model with that of the corresponding base model,and the feasibility and adaptability of the set of methods for runoff prediction in the Han River basin.