Uncertainty Analysis Of Rainfall-runoff Models And Risk Assessment Of Reservoir Flood Control

Water resources security has been the key problem restricting the development of economics and society in the 21st century. Hydrological forecasting is the basis of water resources secure utilization, in terms of flood control operation decision-making, ecological environment protection, and water resources development, management and utilization in the water resources and hydropower engineerings. It is related to the development of hydrology research, and plays a key role in flood management. Uncertainty exists in the hydrological process widely.This study focuses on the runoff forecasting uncertainty analysis in the form of long-term and short-term and continuous simulation. Within the GLUE(Generalized Likelihood Uncertainty Estimation) framework, it studied the present problems of uncertainty analysis of hydrological models and proposed the multi-criteria and a limits of acceptability approach to the calibration of hydrological models based on extending observation error. And then it addressed the uncertainty analysis method into the risk assessment of reservoir flood control, built an integrated risk analysis methodology. The research mainly includes the following parts:(1) Based on the fuzzy optimization neural networks model, it studied the two issues of long-term and medium-term runoff forecasting as following. Firstly, within the fuzzy optimization neural network model, the methods of selecting forecasting factors were studied in annual runoff forecasting. The linear correlation coefficient is not appropriate to describe the correlations between factors and objective in long-term prediction. A compound non-linear correlation formula is proposed to solve this problem. Secondly, aiming at the inconsistency in simulation and prediction of this model, an integrated effective coefficient approach is presented. The main parameter—the iteration error—is calibrated.(2) A nash-cascade lag linear channel routing method is proposed, which focuses on the routing problem of nested catchment and different conditions of subcatchment in runoff generation. The TOPMODEL was introduced, and its application was discussed in Biliuhe reservoir catchment of China. According to the nested catchment—Atter catchment in Luxembourg, the nash-cascade linear channel routing method was applied in terms of continuous flow simulation. The Dynamic TOPMODEL system was more integrated, and the model run became more efficient in time costing. (3) Multi-criteria likelihood measures were addressed in the GLUE framework for uncertainty analysis of hydrological model. The likelihood measure is critical to the estimation of model performance in GLUE framework. How to select an appropriate likelihood measure is still an open question so far. Usually the Nash-Sutcliffe coefficient is a traditional likelihood measure, but this measure focuses on the global performance without a special assessment in point condition especially for flood peak. A multi-criteria likelihood measure is presented within GLUE methodology. Based on the DHF(Da Huofang)model, it was applied to Biliuhe catchment. The results show the multi-criteria likelihood measure has got improvement compared with the original measure in describing the uncertainty of model. The uncertainty assessment indicates the uncertainty problem is significant in Biliuhe catchment using DHF model, which needs further study on measures (such as input error and model structure error reducing) for uncertainty reduction.(4) A limits of acceptability approach to the calibration of hydrological models was developed by extending observation error. GLUE has been strongly criticized in the past for not using formal statistical likelihood measures and that threshold has usually been chosen subjectively on the scale of some summary goodness of fit index. It is a typical error in model calibration that rejecting a good model because of other sources of uncertainty. Within the GLUE methodology, there are a number of advantages of taking a limits of acceptability approach to model evaluation for non-ideal applications where the strong assumptions of statistical identification might be difficult to justify. The compensation for model structural error by modifying the inputs was avoided. In this way, the input error could be handled implicitly. A novel method for identifying behavioural models in an extended GLUE methodology is developed. It was applied to an application of Dynamic TOPMODEL to the Attert catchment in Luxemburg with semi-distributed inputs to nested sub-catchments incorporating the rating curves. The results raise some important issues about testing model structures as hypotheses of catchment responses.(5) The Monte Carlo stochastic simulation theory was introduced to the risk analysis of reservoir flood control and a novel risk analysis method was proposed. The traditional risk analysis method of reservoir flood control in single factor (e.g, flood forecasting error) has more errors due to the several times of interpolation deducing the point risk. Meanwhile, there is no effective method for assemble risk analysis. This study proposed a risk analysis framework for single and combined factors with introducing the Monte Carlo stochastic simulation theory and Latin hypercube sampling technique. It was applied to Biliuhe reservoir flood control and over-topping risk analysis.Finally, the conclusions are drawn and the problems to be further studied are discussed.