The Theory Of Machine Learning And Its Applications In The Hydrological Forecasting

Hydrological forecasting is always one of the main subjects in the realm of hydrology and water resources,which can provide very important decision making information for various water-related activities including the prevention and control of hydrometeorological hazards,the safe and economic operation of watershed hydropower energy system,the scientific and rational allocation of watershed water resources,and the sustainable development of human society.In recent years,due to the strong influence of global and reginal climate abnormal change,large-scale development of hydropower,and anthropogenic disturbances,the evolution law of water circulation system is undergoing profound changes,meanwhile,the spatial and temporal heterogeneity of water resources is undergoing aggravated,which produces more frequent natural hazarzs including flooding and drought,and makes more difficulties for basin water resources allocation.Thereby,higher requirements and new challenges are put forward for the research of watershed hydrological forecasting.Therefore,it is very urgent to study and propose novel theories and methodogies for hydrological forecasting to fully extract the effective and implicit information in hydro-meteorological series and further improve the accuracy and reliability of runoff forecasting.Focus on the key scientific and technical problems in disaster prevention and mitigation of the basin water circulation system under changing environment,this thesis takes the upper reaches of the Yangtze River as study area,adopts advanced machine learning methods,feature selection and intelligent optimization techniques,exploit the advanced theory,techniques and methods of non-linear hydrological forecasting.The current research has important theoretical significance and engineering practical value for reducing the loss of major natural disasters and realizing the safe and economic operation of hydropower energy system.Relevant research results can further promote the development of hydrological forecasting in practical engineering application,also can be utilized for reference by river basin authorities.The main contents and innovative achievements of this thesis include:(1)According to historical hydro-meteorological data in 1961?2010,many methodologies including exploratory data analysis,mathematical statistics and time-frequency multi-scale analysis were adopted to comprehensively analyze the seasonal variations and inter-annual rules of the key hydro-meteorological factors including runoff and precipitation of the Jinsha River Basin over the past 50 years.Additionally,slope changing ratio of cumulative quantity was applied to quantify the contribution of climate change and anthropogenic disturbances to runoff variation.The experimental results show that the annual precipitation and annual runoff in Jinsha river basin had an increasing trend,and the influence of climate change on runoff is more significant than that of human activities in this period.(2)The flood forecasting based on machine learning has attract much more attentions because it does not need to deeply understand the physical mechanisms of all processes in water circulation system,and it has strong non-linear fitting ability,less data demand and simple model construction.However,it is difficult for a single machine learning method to fully characterize the intrinsic characteristics of runoff at different scales and the external influence of meteorological factors and human activities on runoff.In this regard,this paper studied the theory of non-linear intelligent flood forecasting based on intelligent optimization algorithm and machine learning methodologies,designed two strategies of orthogonal initialization population and adaptive variation scale coefficient to improve the standard backtracking search algorithm,and then established the flood forecasting model called ELM-IBSA in which the improved BSA was applied to optimize the parameters of ELM.Case study shows that the ELM-IBSA model has better stability and generalization ability than the traditional GRNN,the single ELM and the ELM-BSA,as well as it can provide more reliable flood forecasting information for real-time flood control operation of reservoir group.(3)Most of the decomposition-ensemble streamflow forecasting models(DESFMs)are established according to one-time decomposition strategy.In their modelling processes,some future information,which is actually unknown at the present moment,is used.So,the one-time decomposition-based forecasting models are not suitable for practical applications.In this regards,we designed an adaptive dynamic decomposition strategy to solve the above problem,and then proposed a hybrid forecasting model which can be adaptively and dynamically updated when new data is added.Comprehensively comparisons show that the developed hybrid model is successfully be implemented as a promising alternative for mid-long term streamflow forecasting not only from the perspective of reliability but also from the view of overall skills.(4)Due to the deterministic streamflow forecasting can just provide a single point value of the targeted variable in the future time.Meanwhile the streamflow forecasting method based on machine learning usually do not consider the runoff generation mechanism.Therefore,a new streamflow interval prediction method based on Gaussian Process Regression(GPR)and Random Forest(RF)was proposed.In this method,the RF was applied to compute the variable importance measure(VIM)of alternative input vectors which consist of previous measured rainfall,streamflow and teleconnection climatic factors.Then rank input vectors according to their VIM values.According to the forward search strategy,the optimal input set can be determined.After that the best input variable set were plugged into GPR to do streamflow forecasting and provide uncertainty interval of the forecasting errors at the same time.Results indicate that adding climate factors can improve forecasting accuracy and the main influential factors in short forecast period are the previous streamflow,while in long forecast period are the teleconnection climatic factors such as NINO indexes.