| Under the changing environment,the non-stationarity and complexity of runoff series are enhanced,which improves the difficulty of runoff prediction work.The thesis explores the applicability of time-frequency decomposition and ensemble learning model for medium-and long-term runoff prediction in the basin of karst area,taking the basin above the Pingtang hydrological station of Chengbi River as the research object,and the main research results are as follows:(1)Time-frequency characteristics analysis of runoff from Pingtang hydrological station,intra-annual characteristics analysis,inter-annual characteristics analysis,runoff trend analysis,runoff analysis and runoff cycle analysis of monthly runoff,respectively.The results show that:the intra-annual distribution is extremely heterogeneous;the proportion of inter-annual monthly runoff to annual runoff basically varies little;the trend of runoff in March,June and July is increasing,and the runoff in the remaining months shows different degrees of decrease;the Pettitt method is used to analyze the abrupt changes of monthly runoff month by month,and it is found that the runoff has abrupt changes of different significance degrees,and there are 2-5 scale cycles of runoff in each month;the runoff has complex time-frequency characteristics.(2)Five ensemble learning single models:random forest(RF)model,gradient boosting decision tree(GBDT)model,adaptive boosting(Adaboost)model,extreme gradient boosting tree model(XGBoost)and light gradient boosting machine(Light GBM)model were constructed for monthly scale runoff prediction.The results show that the prediction grade of all five models is C.The best prediction is Light GBM model,however,the DC value is only0.53.The best prediction is Light GBM model,the three evaluation indexes DC,RMSE and MAE are:0.53,34.02 m~3/s and 22.28 m~3/s,respectively;The worst prediction is the GBDT model with three evaluation indexes DC,RMSE and MAE of 0.50,35.21m~3/s and 22.85 m~3/s,respectively.(3)In view of the unsatisfactory prediction effect of the five ensemble learning models,three time-frequency decomposition methods,empirical modal decomposition(EMD),complementary ensemble empirical modal decomposition(CEEMD)and variational modal decomposition(VMD),were first used to decompose the runoff series;then five ensemble learning models were used to make predictions,and a combined model based on one time-frequency decomposition was constructed.The results show that:the prediction grade of the 15 combined models is B;the time-frequency decomposition method helps to improve the prediction effect of the models,while the decomposition effect of the VMD method is found to be better than that of the CEEMD method,which is better than that of the VMD method;the VMD-Light GBM model with the best prediction effect,the three evaluation indexes DC,RMSE and MAE are:0.87,17.86m~3/s and 12.77 m~3/s;the worst prediction effect is the EMD-GBDT model,with three evaluation indexes DC,RMSE and MAE of 0.72,26.11 m~3/s and 19.20 m~3/s,respectively.(4)On the basis of the combined model,the sub-series components with poor prediction effect of EMD and CEEMD are continued to be decomposed by VMD decomposition method,and the combined model based on two-stage time-frequency decomposition is established.The results show that the prediction accuracy of the combined model based on two-stage time-frequency decomposition is higher than that of the combined model based on one-stage time-frequency decomposition,and the accuracy improvement rates of DC value,RMSE value and MAE value range from 16.69%-24.35%,30.65%-40.95%and 28.32%-42.74%;the accuracy of the five models is grade B,and the range of DC value is 0.87-0.91;the best prediction is the CEEMD-VMD-Light GBM model,the three evaluation indexes DC,RMSE,MAE values are the best,respectively:0.91,14.60m~3/s and 11.69m~3/s;the worst prediction is the EMD-VMD-GBDT model,the three evaluation indexes DC,RMSE,MAE values are the most inferior,which are:0.87,18.11 m~3/s and13.47 m~3/s,respectively. |
PDF Full Text Request