Research On Early Warning Of Streamflow In The Middle And Lower Reaches Of Chishui River Based On Machine Learning

In this thesis,a wavelet-hybrid artificial neural network(ANN)model integrated with trees-based iterative feature selection(TIFS)is evaluated for its statistical preciseness in water level warning of Chishui River,and it is then benchmarked against interative iterative input selection(IIS).Many machine learning algorithms for making predictions have been introduced in recent years.Therefore,it was attempted in this study to compare the prediction performance of several machine learning models to 3 hours and 6 hours waterlevel forecast of Chishui station,these models were:multiple linear regression(MLR),extremely randomized trees(Extra-Trees),and ANN.Then,we compare the performance of two feature selection algorithms of IIS and TIFS in the water level prediction of Chishui hydrological station,and further introduce MODWT to build the integrated wavelet decomposition model.Finally,the TIFS-MODWT-ANN model was proposed for water level early warning in the middle and lower reaches of Chishui River.(1)First,this thesis studies the effective of three kinds of machine learning algorithms including multiple regression,Extra-Trees and ANN in the application of 3 hours and 6 hours waterlevel forecast in Chishui River.(2)Then,we introduced different kinds of commonly used input selection methods,including PCIS(correlation analysis based),PMIS(information theory based)and the iterative input selection method IIS.We discussed their respective application scenarios and distinctive features,and finally decided to use IIS as the feature selection algorithm of water level forecasting model of Chishui hydrological station.In the model building process,based on the specific characteristics of the water level data in the middle and lower reaches of the Chishui river,we improved the IIS and proposed the trees-based iterative feature selection algorithm(TIFS)is proposed.(3)Furthermore,to establish robust forecasting models,iterative input selection(IIS)and tree-based iteration feature selection(TIFS)algorithm were applied to screen the best data from the predictor matrix and were integrated with the non-decimated maximum overlap discrete wavelet transform(MODWT)applied on the IIS-selected and TIFS-selected variables.This resolved the frequencies contained in predictor data while constructing a wavelet-hybrid model.(4)And then,three kinds of machine learning algorithms,including Extra-Trees,MLR and ANN were used to developing the IIS-MODWT and TIFS-MODWT-hybrid forecasting models,and forecasting ability of those wavelet-hybrid models were evaluated via root-mean-square-error(RMSE),Willmott's Index(WI),Nash-Sutcliffe Efficiency(NSE),and mean absolute error(MAE).(5)Finally,the design and implementation of the flood control decision support and early warning platform for the middle and lower reaches of the Chishui River are completed.The experimental results show that,after integrated with maximum overlap discrete wavelet transform and tree-based input feature selection algorithm,both MLR,Extra-Trees and ANN models demonstrated a dramatic performance improvement.And,of all the models discussed,the TIFS-MODWT-ANN model has obtained the optimal predictive performance,with RMSE=0.072 m,MAE=0.041 m,NSE=0.993 and WI= 0.998 for the 3 hour-ahed forecast.It is clear that the integration of the ANN model with the TIFS-MODWT scheme has produced 1.3% and 0.5% increase in NSE and WI,respectively,whereas the RMSE has decreased by 38.5% and MAE by 34.9%.And when compared with the IIS-MODWT-ANN model,reduction 20% in RMSE value and 24.1% in MAE value,while NSE increased 1.3% and 0.4% increased in WI.