Study Of Intelligent Neural Network Model For Dam Safety Monitoring And Safety Evaluation

Based on the previous dam safety monitoring models and dam safety evaluationmethods, the dissertation is devoted to combining Back-Propagation(BP)neural networkwith some novel intelligent algorithms to form a series of intelligent neural network modelsto overcome shortcomings of slow convergent rate, local minimum, poor stability andinfluence of initial values of BP neural network model. And then, the intelligent neuralnetwork models are applied to dam safety monitoring prediction and dam safety evaluation.The main research works of the dissertation are as follows:Aimed at the slow convergent rate of BP neural network, Levenberg-Marquardtalgorithm was adopted to replace the gradient descend method in training neural network.And then two kinds of dam safety monitoring models named seriatim Levenberg-Marquardtmodel and whole Levenberg-Marquardt model are established. Case study shows that thewhole Levenberg-Marquardt model and the seriatim Levenberg-Marquardt model aresuperior to the fast BP neural network in prediction accuracy and training speed, and theseriatim Levenberg-Marquardt model are superior to the whole Levenberg-Marquardt model.A dam safety monitoring prediction model based on radial basis function neural networkis established to reduce the forecasting period. Case study shows that the radial basis functionneural network model is much better than the fast BP neural network model in generatingability, prediction accuracy and training speed.Considering influence of initial values, poor stability, slow convergent rate and the localminimum problem of BP neural network model, Genetic Algorithm (GA) is combined withBP neural network based on Levenberg-Marquardt algorithm to form GA neural network,and then two kinds of dam safety monitoring models named seriatim hybrid geneticalgorithm model and whole genetic algorithm model are established. Case study shows thatthe seriatim hybrid genetic algorithm model and the whole genetic algorithm model aresuperior to the seriatim Levenberg-Marquardt model and the whole Levenberg-Marquardtmodel respectively, and the seriatim hybrid genetic algorithm model and the seriatimLevenberg-Marquardt model are superior to the whole genetic algorithm model and thewhole Levenberg-Marquardt model respectively in prediction accuracy. With accumulationof samples, the prediction accuracy of the seriatim hybrid genetic algorithm model can beimproved remarkablely in less training time. Thus, it is very effective in real-time predictionof dam safety monitoring. In order to improve the local search ability of genetic algorithm in the latter period ofevolution, simplex method is combined with genetic neural network to form geneticalgorithm simplex method neural network, and then a dam safety monitoring model based ongenetic simplex neural network is established. Case study shows that the genetic simplexneural network model owns good convergent rate, high prediction accuracy, fast trainingspeed and superior generating ability compared with the whole genetic algorithm model andthe whole Levenberg-Marquardt model. Thus, the method is feasible and effective for damsafety monitoring prediction.Considering the local minimum problem, slow convergent rate and influence of initialvalues of BP neural network model, Particle Swarm Optimization(PSO)is combined with BPneural network based on Levenberg-Marquardt algorithm to form PSO neural network, andthen two kinds of dam safety monitoring models named seriatim PSO neural network modeland whole PSO neural network model are put forward. Case study shows that the seriatimprediction models are obviously superior to those whole models in forecasting precision.Furthermore, The PSO neural network models exhibit faster convergence rate and betterforecasting precision than BP neural network models based on Levenberg-Marquardtalgorithm. Especially, the seriatim PSO neural network model is more efficient in precisionand training rate to satisfy the demand of real-time prediction of dam monitoring.In order to overcome shortcomings in traditional weighted arithmetic averagingcombination forecasting method and improve prediction accuracy, the concepts of inducedordered weighted averaging(IOWA) operator and induced ordered weighted geometricaveraging (IOWGA) operator are introduced, and then two kinds of dam safety monitoringcombination forecasting models based on IOWA operator and IOWGA operator are putforward respectively. Case study shows that the proposed combination models are superior totraditional ones and each single model in prediction accuracy and are effective and credibleas well in prediction of dam safety monitoring.Aimed at shortcomings of slow convergent rate and poor stability of BP neural networkmodel, the radial basis function neural network is attempted to apply to dam safetyevaluation based on existing evaluation methods and theories. The evaluation example of tentypical dam sections of Fengman dam testifies the validity of the new method.The main contributions are summarized and further works are suggested at the end ofthe dissertation.