Study On Soft-computing Technology And Its Application In Parameter Identification Of Complicated Environmental Model

Parameter identification plays an important role in mathematical modelapplication. The use of large and complex mathematical models is nowcommon practice in the environmental sciences, which has been madepossible through the development of increasingly deep research on processesof advection, dispersion, degradation and boundary exchange ofenvironmental pollutants. Difficulties of parameter unidentifiability ensuefrom model complication. Therefore, probing into effective and efficientparameter identification methods has great theoretical and realisticimportance in the application of complicated environmental model.The hard computational technologies refer to the traditional numericalaccurate solution methods of partial differential equations, whereas the softcomputational technologies include some new computational technologies,which pursue approximate solution effectively. As a kind of commonly usedapproaches, soft computing methods have very simple structures, and cansearch global solutions efficiently. Based on soft computing technologies, anapproach of parameter identification for complicated environmental modelwas proposed in the Dissertation. Bayesian inference was the theoretical basisof the approach, and Markov Chain Monte Carlo (MCMC) method was itskernel and other processes, such as parameter sensitivity analysis, parametercorrelation analysis, and parameter global searching, were auxiliary ways toacquire more prior information. In order to study the performance andefficiency of soft-computing technologies in application to parameteridentification of complicated environmental model, water quality simulationof the Miyun Reservoir using WASP (Water Quality Analysis SimulationProgram Modeling System) was selected as a case study.Results of both artificial data and realistic data numerical experimentsindicated that the proposed approach of parameter identification forcomplicated environmental model could acquire posterior distributions ofparameters efficiently and reliably. MCMC method was a significant tool tosampling posterior distribution. Firstly, sampling series could converge toposterior distribution of parameter, and Gelman convergence diagnosis indexwas definite and very easy to use. Secondly, MCMC sampling series hadMarkov Chain characteristics, so sampling process was random no longer. Incomparison with Monte Carlo method it could reduce computationalcomplication greatly. Eventually, MCMC method could solve the problem ofparameter sensitivity and correlation well. Besides, results indicated thatglobal searching algorithms could quickly find the high probability regions ofparameters, at the same time they further assisted to know more propertyinformation of parameter sensitivity and correlation.In general, the efficiency of using GA in practical problem depends onthe proper selection of genetic algorithm (GA) operators and controlparameters. Based on the feature of parameter identification, orthogonal testmethod was proposed for reviewing effects of different GA control parameterson the performance of model parameter optimization. The result indicated thatorthogonal method could identify key factors, and also could provide possibleoptimized parameter combination in only limited computation. The hybridalgorithm mingles different search mechanisms of different methods, so it hasstrong abilities of exploration and exploitation in large real search spaces.Two typical hybrid algorithms, such as SMSA (algorithm of SimulatedAnnealing and Simplex Method) and GASM (Genetic Algorithm and SimplexMethod), were researched. Results indicated that SMSA and GASM couldacquire the global optimizing solutions efficiently and reliably either in thecondition of non-disturbing data or in the condition of disturbing data.