【作者】 彭勇；

【导师】 周惠成；

【作者基本信息】 大连理工大学 ， 水文学及水资源， 2007， 博士

【摘要】 随着计算机技术的迅速发展和预测优化理论的深入研究，水利科学工作者几乎尝试应用了各种算法来解决水文预报和水库群的优化调度问题，并且取得了丰富的研究成果和一定的实践效果。但由于在水文中长期预报方面缺乏机理上的认知突破，在水库群优化调度方面一些新方法仍存在理论与应用上的问题，所以尚有大量的基础工作、理论研究、模型与方法改进等课题需要进一步的研究和解决。本文主要是进行模型与方法的改进工作，深入研究了模糊优选神经网络、支持向量机、小波分频预测和微粒群算法在水文预报和水库调度中的建模及其系统集成方法，主要内容如下：(1)针对传统模糊优选神经网络模型训练速度慢的缺点，模糊优选神经网络训练时采用Levenberg-Marquardt算法，其收敛速度近似二阶。Levenberg-Marquardt算法是Gauss-Newton法的改进形式，既有Gauss-Newton法的局部特性，又具有梯度法的全局特性，提供了Newton法的速度和保证收敛的梯度下降法之间的折衷，收敛的迭代次数少，能快速完成网络训练。两个实例分析结果表明，Levenberg-Marquardt算法训练速度快，并且训练得到的模糊优选神经网络模型的模拟预测结果都比较满意。(2)利用微粒群算法的全局寻优特性来进行支持向量机的参数识别。在支持向量机建模过程中选择径向基函数为核函数，在此基础上应用微粒群算法用于辨识支持向量机参数。在微粒群搜索参数前，对参数做了指数变换，使[0，1]区间和[1，∞]区间有着相同的搜索概率。微粒群算法的适应值函数以支持向量机模型的推广能力为标准，讨论了测试样本的最小误差和留一法对支持向量机学习方法推广能力的两种估计。支持向量机训练算法采用序列极小化训练算法，为了提高支持向量机模型的训练速度，训练工作集的选择利用了目标函数的二阶导数信息，使工作集中的两个乘子在当前训练迭代步中具有最大的改变量。最后以长江的宜昌站、寸滩站和漫湾水库的月径流资料为例，分别采用ARMA模型、季节性ARIMA模型、BP神经网络模型以及所建立的支持向量机模型进行模拟预测，实例结果显示了该模型有效性。(3)利用小波分析的多分辨率功能，建立了月径流序列的小波分频预测模型。该模型通过小波分解方法将月径流非平稳时间序列分解为多个细节信号序列和一个逼近信号序列，然后运用平稳时间序列的ARMA模型对各信号序列分别进行预测。以长江的宜昌站和寸滩站的月径流资料为例，分别采用ARMA模型、季节性ARIMA模型、BP神经网络模型以及本文所建立的小波分频预测模型进行模拟预测，并讨论了小波分解尺度数对小波分频预测模型的影响。最后还讨论了小波分频预测模型的实用性、缺陷和小波拓延方式对预测精度的影响。(4)对基本微粒群算法进行了改进，以进一步提高基本微粒群算法的优化性能，并将其用于求解梯级水库群优化调度问题。该算法为了提高搜索效率引进了类似遗传算法的交叉和变异算子，其中交叉是微粒在解空间中的位置以一定的概率随机进行算术交叉，变异是微粒以一定的概率随机使速度矢量的某一维分量变为0。为了加速收敛，初始微粒群生成时采用了有条件的随机自动生成方式，并利用惩罚函数法来处理边界条件和其它非等式约束。以丰满-白山梯级为计算实例，计算结果与常规动态规划法和常规微粒群算法相比，表明改进算法计算速度更快，优化调度结果比较满意。最后还研究了改进微粒群算法参数(交叉概率、变异概率、微粒个数和微粒最大速度等)的取值范围。(5)首先根据历史月径流资料，以“支持向量机方法”估计出当年各月的来水量；然后根据历史年径流资料，采用加权马尔可夫链预估当年总的来水情况；最后根据总的来水情况来修正当年主汛期各月来水。这样，各月预报来水获得后，再采用“改进微粒群算法”进行水库群的优化调度计算。同时以白山—丰满梯级水电站水库群为例，提出了径流预测与优化调度相结合的确定水电站最优运行方式的滚动预测决策方法。探讨了实际、预报、上限、下限和插值5种不同径流描述情况下水库的发电效益，计算结果表明本文提出的滚动预测决策方法在白山—丰满梯级水电站水库群是可行的。(6)首先以图论分析方法，把水库群概化成一张有向图，运用邻接表、邻接矩阵和关联矩阵进行拓扑分析解决了水库群的集成问题。并以图论为基础，利用方案关联表进行方案的拓扑分析，降低了方案管理的复杂性，便于用户使用与管理。然后研究了软件工程中的多种设计模式在水库预报调度系统中的应用，并将上述预报优化调度模型融入其中。应用桥梁模式实现水库预报调度系统的体系结构设计与数据存取、应用策略模式实现预报优化调度模型管理、应用迭代子模式和装饰模式实现预报调度方案管理、应用代理模式实现用户权限管理、应用单例模式实现数据库连接池的管理等。实现了模块的松耦合关系，增加了系统的可维护性和复用性。该系统采用Web浏览器和Java窗口两种显示界面，运行时可根据环境自动选择不同的通信方式与系统业务逻辑功能模块进行交互。最后对全文做了总结，并对有待于进一步研究的问题进行了展望。更多还原

【Abstract】 With the development of the technology of computer and the theory of optimization andforecasting, researchers of hydraulic engineering have utilized many different algorithms tosolve the problems of hydrological forecasting and optimal operation of reservoirs system. Alarge amount of results have been made and come into effect. However, there are still a lot ofproblems by using these new methods duo to lack of reorganization of the mechanics ofhydrological forecasting. Therefore, large numbers of subjects inbasic work, theory study andimprovement of models should be studied and solved further. This dissertation focuses onimproving the models and the methods for hydrological forecasting and optimal operation.Long-term hydrological forecasting and optimal operation modeling methods and itsintegration are studied in-depth using the Fuzzy Optimization Neural Network (FONN),Support Vector Machines (SVM), wavelet transform subband forecast method and particleswarm optimization algorithm (PSOA). The main task of this dissertation includes thefollowing sections:(1) A new fuzzy optimization neural network model is proposed based on theLevenberg-Marquardt (LM) algorithm on account of the disadvantages of slow convergenceof traditional fuzzy optimization neural network model. In this new model, the gradientdescent algorithm is replaced by the LM algorithm to obtain the minimum of output errorsduring network training, which changes the weights adjusting equations of the network andincreases the training speed. Two cases study is utilized to validate this new model, and theresults reveal that the new model fast training Speed and better forecasting capability.(2) Using the global searching performance of PSOA to identify the parameters of SVM.In the process of modeling SVM, radial basis function (RBF) is used as kernel function, basedon which PSOA is employed to identify the parameters of SVM. Before particle swarmsearches the parameters, the parameters are transformed into exponent, which makes interval[0, 1] and interval [1,∞] have the same searching probability. Adaptive value function ofPSOA takes the generalization capability of SVM model as the criterion; thus two kind ofestimation of the generalization capability of SVM, the minimum error of test samples andleave-one-out method, have been discussed. Sequential minimal optimization trainingalgorithm is taken as the training algorithm of SVM. In order to improve the training speed,the second derivative of objective function are employed to choose the working set for training, which makes the two multipliers in working set have the maximum variation incurrent iteration. Taken the monthly discharge of Yichang station, Cuntan station in YangtzeRiver and Manwan reservoir as examples, ARMA model, and seasonal ARMA model, BPneural network model and proposed SVM model are used to forecast the monthly discharge,respectively. The simulation and forecasting results show the effectivity of the proposedmodel.(3) Based on wavelet analysis theory, a wavelet prediction model is presented for thesimulation and prediction of monthly discharge time series. In this model, the non-stationarytime series of monthly discharge is decomposed into an approximated time series and severalstationary detail time series according to the principle of wavelet decomposition. Each one ofthe decomposed time series is predicted respectively through the ARMA model for stationarytime series. Taking the monthly discharge at Yichang station and Cuntan station of YangtzeRiver as examples, the monthly discharge is simulated by using ARMA model, seasonalARIMA model, BP artificial neural network model and the wavelet prediction modelproposed in this article, respectively. And the effect of decomposition scale for the waveletprediction model is also discussed. Meanwhile, the practicability and the shortcoming ofwavelet prediction model are analyzed. Moreover, the influence of the wavelet extensionmode on the forecasting accuracy is discussed.(4) To improve the optimization capability of traditional PSOA, a revised PSOA isproposed and used to solve the problems of optimal operation for cascade reservoirs. Forincreasing the searching efficiency, the revised PSOA introduces crossover operator andmutation operator, which are similar to the operators in genetic algorithm. The positions ofthe particles in solution space are doing the arithmetical crossover at a certain of probabilitycalls crossover, and mutation means the particles at a certain of probability randomly makingsome one dimension component change into zero. In order to increase the convergence speed,a stochastic generated method under certain conditions is utilize to generate the initial particleswarm, and penalty function is used to deal with the boundary conditions and other inequalityconstraints. Taking example for Fengman-Baishan cascade reservoirs to compare the revisedPSOA with the traditional dynamic programming algorithm and the traditional PSOA, theresults show that the revised PSOA has faster calculation speed and optimal operation resultsare satisfied. Furthermore, the spans of the parameters (crossover probability, mutationprobability, the number of the particles and the maximum speed of the particles) in the revisedPSOA are also studied.(5) Using the proposed SVM method and the revised PSOA into the optimal operation ofthe cascade reservoirs. Firstly, the proposed SVM is used to estimate the incoming monthlyrunoff according to the monthly discharge in history. Secondly, the weighted Markov-Chain is used to estimate the total runoff of current year according to the annual runoff in history.Finally, the total runoff is used to revise the current monthly runoff in main flood period. Inthis way, after obtaining the forecasting monthly runoff, the revised PSOA is used to get theoptimal operation of the cascade reservoirs. Taking Feng-man-Baishan cascade reservoirs asan example, a rolling prediction operation method, which combines the runoff forecasting andthe optimal operation, is put forward to determine the optimal operation manner ofhydroelectric station. Meanwhile, the benefit of electric power generation under differentrunoff situations, such as real, forecasting, upper limit, lower limit and interpolation, arediscussed. The results show that it is feasible to using the proposed rolling predictionoperation method in Fengman-Baishan cascade reservoirs.(6) Based on the graph theory, multi-reservoir is conceptualized as a digraph. Theadjacent list, adjacent matrix and incidence matrix are used to analyze topology relations andto solve the integration of multi-reservoir. And relation tables of alternatives are used toanalyze topology relations, so the complexity of alternative management is reduced for theconvenient use and management of client. Then several design patterns in softwareengineering were studied for the purpose of applying them to the forecast and operationSystem for reservoirs (FOSR). The bridge pattern is employed to design the architecture ofFOSR and the data access; the strategy pattern is used to manage the models of FOSR; theiterator pattern and the decorator pattern are employed in the alternatives management; theproxy pattern is for managing the users’ authorities; and the singleton pattern for the databaseconnection pool. The patterns used in the software achieve the loose coupling relationshipbetween different modular and increase the maintainability and reusability of the FOSR. Twotypes of display interfaces, Web Browser and Java window, are developed in this system.When FOSR is running, it can choose the communication manner automatically and interactwith the system business logic functional modular according to the operating environment.Finally, a summary is given and some problems to be further studied are discussed.更多还原