Inversion Of Roughness Parameter Of River Network Based On Machine Learning

Hydraulic technology has been mature since 1970s, but parameter in calculation, such as roughness parameter is hard to define automatically in project, so that automatic inversion of roughness parameter demands prompt solution. Present research is mainly based on optimization method and there exist three problems: Conventional optimization methods easily gets stuck at a local optimum;Lacking of research on optimal layout of observe stations and lacking of inversion of roughness parameter in river network through the time sequence observations on few hydrology stations; Lacking of test of inversion solution and evaluation on uniqueness and optimization.The author completed a whole framework of inversion of roughness parameter in river network based on machine learning after researching on the whole process from the calculation of roughness inversion to the test of inversion solution. The achievements of the dissertation are as follows:1. A new direct inversion of roughness parameter in river network is proposed by means of data mining based on machine learning theory. It can inverse roughness parameters of all the channels in the whole river network since it inherited the traditional idea of direct inversion and is introduced into the latest development of artificial Intelligence. Through a few calculations, it is turned out to be highly accuracy to meet the actual demand of project.2. The dissertation explored the techniques of date collecting and inversion of roughness parameter in all the channels by using the time sequence observations on few observe stations in the river network .The calculation showed that time sequence observations on few observation points can be used to inverse the roughness parameters of all the channels only by full utilization of prior information. Meanwhile, reasonable quantities and installation position of observation points were studied for placement optimization of hydrological stations in the river network engineering.3. It introduced a new GA-RBF method of inversion of roughness parameter. Compared with the traditional BP method, the new method is of faster calculating velocity and highly calculating accuracy. Compared with classic RBF network, it determines the reasonable topological structure of the network automatically with input and output information.4. Considering the inherent ill-posed characteristics, Bayesian method was introduced in inversion in order to test the result of inversion effectively. It cannot obtain determined inversed value but a posterior probability distribution of awaiting inversion parameter by using this method. Multiplicity of inversion and optimality judgment of solutions can be obtained through the analysis of histograms of posterior distribution of roughness parameter so as to test uniqueness and optimality of inversion result.5. By using Bayesian method, it realized the test on verifying spread of solution depended on the ergodicity of Bayesian method, with the disadvantage of long computing time. The key point of calculating speed is detected through analysis of calculation time structure. The machine learning is introduced in the calculation of likehood function and the ANN substituted plenty of repetitive forward modelings, so as to optimize the traditional Bayesian method and increase the efficiency of calculation.6. Case study shows that the application of the new direct inversion method based on machine learning to inverse roughness parameter of complicated river network is feasible.