【作者】 韦安磊；

【导师】 曾光明；

【作者基本信息】 湖南大学 ， 环境科学与工程， 2010， 博士

【摘要】 如何提高污水处理效率和过程优化控制策略是国内外污水处理研究领域普遍关注的问题。污水处理过程具有时变性、非线性和复杂性等鲜明特征,这使得污水处理系统的运行和控制极为复杂。在我国当前水环境形势下,开展污水处理过程数学模型方法研究,即具有重要的理论价值,也有紧迫的现实需要。数学模型方法的有效性和可行性是数学模型应用推广的前提。论文回顾了污水处理过程模型近50多年的发展历程,认为数学模型是污水处理管理和决策不可缺少的工具;同时由于传统数学模型方法的相对复杂性与实际操作人员的知识有限性之间的矛盾,当前污水处理过程模型推广应用存在着两个难点:一是缺少简捷有效的机理模型校核方法;二是缺少稳定性和准确性俱佳黑箱模型方法。因此,应该寻求简捷可靠、行之有效的模型方法及其关键技术,以增强污水处理过程模型的实用性和可行性。论文以提高污水处理过程模型精度和简化模型校核步骤两方面为目标,提出了一种待估参数选择方法,反向选择法。该方法通过分析模型输出均方差,并将均方差分解为方差项和偏差项:前者与测量噪声方差和所选待估参数个数成正比,后者同模型参数取值及其灵敏度有关。反向选择法最初设定全部模型参数均属于待估参数集,然后逐步行进,每一步挑选出一个对提高模型精度无益的参数,并将其排除在待估参数集之外。随着待估参数个数的减少,方差项值逐渐降低,偏项值则逐渐增加,当偏项值增加幅度大于方差项值减少幅度时,便停止选择步骤,那么仍停留在待估参数集的参数则被用于模型校核。反向选择法在计算参数灵敏度时考虑了参数的不确定性,这更符合实际情况。通过与全体参数估计法和另一种待估参数选择方法正交选择法对比,发现反向选择法能够大大降低模型校核估计参数数目,同时提供了良好的模拟和预测性能。此外,论文还通过实验室规模的SBR系统模型验证了反向选择法的有效性。机理模型与人工神经网络(ANN)模型相结合的复合模型方法能够在机理模型未校核的前提下实现满意的模拟和预测精度。论文提出了三种复合模型方法:加和式并联复合模型方法(ASM2d+ANN)、乘积式并联复合模型方法(ASM2d*ANN)与串联复合模型方法(ASM2d-ANN)。复合模型方法中,以ANN模型学习机理模型因未校核而引起的误差部分,进而完成最终模拟。基于某污水处理脱氮除磷过程测量数据,分别建立了未校核的ASM2d机理模型, ASM2d+ANN模型,ASM2d*ANN模型与ASM2d-ANN模型,采用这些模型模拟出水水质。模拟结果表明:尽管ASM2d机理模型未经校核、并存在较大误差,但其输出能展示出与出水水质相似的波动特征,这为构建复合模型奠定了基础;基于ASM2d机理模型,辅以ANN模型,三种复合模型均显示了良好的模拟和预测性能,这表明复合模型不仅具有良好的向内差值功能,而且拥有可靠地外推插值能力;三种复合模型之中,ASM2d-ANN模型性能最优,其串联结构决定了ASM2d-ANN模型能够合理地描述存在于ASM2d机理模型输出和变量测量值之间的非线性关系。因此,复合模型方法适用于污水处理过程模拟,是一种简捷有效的模型方法。黑箱模型方法是解决污水处理复杂动态过程模拟问题的有效途径。论文深入探讨了三种黑箱模型方法:适应性模糊推理系统(ANFIS)、小波网络和小波变换-模糊马尔可夫链方法。ANFIS是一种具有神经网络学习特征的模糊推理系统;论文利用高浓度废水厌氧处理过程数据,验证了ANFIS模型的有效性;并且针对ANFIS模型快速收敛的特点,提出了一种模型输入优化选择方法,潜在式最优输入选择法,通过与主成份分析法对比以及模型不确定性分析,表明基于潜在式最优输入选择法构建的ANFIS模型泛化能力强、稳定性好。小波网络是隐层神经元为小波函数的单隐层神经网络;论文通过膜滤过程膜通量测量数据验证了小波网络模型的有效性,并且发现小波网络具有网络构建方法明确、网络初始化效果好、快速收敛等显著特征,可用于污水处理过程模拟。小波变换-模糊马尔可夫链方法是一种时间序列预测方法,其首先利用小波变换将原始时间序列分解为若干子序列,然后对每个子序列分别建立模糊马尔可夫链模型,最后再通过小波重构将各子序列输出合并以预测原始时间序列;论文将此方法用于污水处理厂进水BOD预测,结果表明小波变换-模糊马尔可夫链方法适用于污水处理过程时间序列预测,且可以通过适当增加模糊划分数目与改变小波函数类型提高预测精度。通过对上述三种黑箱模型方法的分析,表明黑箱模型方法能够充分利用现有数据,准确地把握住污水处理过程中的非线性不规则特征,可以作为污水处理过程模拟方法的选择。更多还原

【Abstract】 It is important to improve wastewater treatment efficiencies and optimize controling strategies. Because wastewater treatment process is often characterized by time-varability, non-linearity and complexity, it is difficult to operate and control wastewater treatment systems normally. Now water environment is seriously threatened in China. Thus it is both theoretically critical and practically essential to study wastewater treatment process models.Mathematical models must have validity and practicability to ensure their popularization. Following a systematic review for the development of wastewater treatment process models over the last fifty years, this thesis highlighted the need for mathematical models in wastewaetr treatment management and decision making. However, there are disputes between profound model theories and limited ability of model designer. Using wastewater treatment process models are now restricted by two points. One is the absence of short-cut for mechanism model calibration, and the other is the lack of black-box models which work well with both stability and accuracy. Therefore, establishing credible and effective modeling approaches and key technologes will be potential for improvinig the validity and practicability of wastewater treatment process models.The thesis proposed a parameter selection approach for model calibration, the inverse selection approach. The proposed approach selected parameters to be estimated by decreasing the mean square errors (MSE) of model output. The MSE consists of two part, a variance item and a bias item. The variance item is proportional to the variance of measurement random errors and the number of selected parameters, and the bias item is related to the values and sensitivities of model parameters. In the first step, the selection procedure made an assumption that all parameters should be estimated. In following steps, each step removed one parameter which contributed the least to improving model simulating accuracy. Follwing a removed parameter, the variance item was reduced and the bias item was increased. Once the decreased value of the variance item was less than the increased value of the bias item, the selection procedure would be stoped, and the remained parameters would be estimated. Note that the inverse selection calculated the sensitivity with parameter uncertainty, which made the procedure more objective. The inverse selection approach was evaluated with experimental data of a lab-scale SBR equipment. The results showed that the inverse selection approach is valid during calibrating mechanism models.Three hybrid model approaches were proposed to represent the dynamic behavior of wastewater treatment processes. Each hybrid model consists of a mechanism model and an artificial neural network (ANN) model. Those models are the additive parallel model (ASM2d+ANN), the multiplying parallel model (ASM2d*ANN), and the cascade model. In the hybrid models, the ANN submodels are used to simulate the error part of the uncalibrated mechanism model. Using experimental data of a nutrient removal process, the proposed models were evaluated to predict the effluent. The results showed that all three models presented with good prediction as well as fitting. Such good performance was caused by the mechanism model which contributed good extrapolation and the black-box model which shared good adatability. The performance of ASM2d-ANN model was better than the others, which was caused by its cascade structure. The cascade hybrid structure ensure the model accurately described the nonlinear relationship between the output of mechanism model and the measurements. Therefore, the hybrid models is suitable for modeling wastewater treatment.Pure black-box model approaches could also be alternatives for modeling wastewater treatment process. Adaptive-network-based fuzzy inference system (ANFIS), the wavelet network, and the wavelet transform-fuzzy Markov chains approach were discussed here. Firstly, ANFIS models were evaluated by experimental data of wastewter anaerobic treatment process, and an potential optimal input selection procedure was proposed for ANFIS models. ANFIS models based on the potential optimal input selection procedure were found with good performance and generalization. Secondly, A wavelet network model was used to simulate the permeate flux of membrane filtration. The wavelet network is a special neural network with only single hidden layer whose neuron is a wavelet function. The results showed that the wavelet network could be constructed definitely and have an good initialization which leads to a fast convergence. Thirdly, the wavelet transform-fuzzy Markov chains approach was evaluated by predicting influent BOD time series. The time series model showed good predicting precision, and could be improved by increasing fuzzy partition and changing the wavelet type. All the above black-box models showed that the black-box model approach could sufficiently make use of the avaiable data and successfully depict the nonlinear irregularity, so it could be used to model wastewater treatment process.更多还原