【作者】 蒋茹；

【导师】 曾光明； 黄国和；

【作者基本信息】 湖南大学 ， 环境工程， 2007， 博士

【摘要】 城市污水处理厂的规划、设计和运行是涉及到多变量的、受多种不确定性影响的一系列决策过程。传统的优化决策理论与方法不能有效地考虑系统固有不确定性的影响,存在一定的片面性和局限性,使模型的实际应用受到限制。本文应用不确定性理论与方法对城市污水处理厂优化决策进行研究,内容包括费用模型的建立、设计阶段的工艺方案优选、设计参数对优化设计的影响分析和不确定性优化设计模型的求解,以及影响运行效果的关键参数污水量的非线性动力学特性研究及其短时预测。这对优化污水处理厂的设计和运行,具有重要的指导意义。污水处理费用模型的非线性、基础数据收集和统计中的不确定性以及费用影响因子内相关等特点,都限制了传统方法的应用。考虑到神经网络自适应、自学习能力以及理论上逼近任意非线性连续映射的优点,本文采用BP神经网络对污水处理厂的费用模型进行探索性的研究。研究以台湾地区26组污水处理厂数据为基础,选取设计水量、处理程度、入流BOD5浓度和集水区域面积作为网络输入,总造价和工厂建设费为输出。采用动量-自适应学习率调整算法、“留一法”和“提前停止”来克服传统BP神经网络收敛速度慢、易陷入局部极小值的缺点,以及样本容量有限、过拟合等问题。结果表明,建立的BP神经网络模型学习能力强,能够较好地预测台湾地区污水处理厂的经济费用,比多元线性回归模型有了很大的改进,预测值与样本之间的相对误差显著降低,预测精度提高。城市污水处理厂的设计首先要涉及到污水处理工艺方案的选择。针对污水处理工艺方案优化决策多目标、多指标、确定性与多种不确定性共存的特点,本文建立了多层次模糊灰色耦合模型和多属性集对分析模型,以弥补现有优化方法单纯考虑费用因素,受人为主观影响较大的不足,为决策提供科学的依据。多层次模糊灰色耦合模型结合了灰色关联分析和层次分析法的优点,采用多层次多指标评判,较常用的单层次评判更能反映方案的实际特性;综合关联度作为优劣评判的不确定度量,具有全面性和提高方案可比性的优点。多属性集对分析模型考虑了确定性和不确定性共存的特点,拓展了结构函数——同异反联系度的概念;独特的稳定性分析通过基序分析其稳定区域,判定扩展序的存在,保证了决策的准确性和可靠性。实例研究验证了两个模型的有效性。某污水处理厂提供了四种候选方案,即A2/O法、三沟式氧化沟、厌氧-单沟式氧化沟和SBR法。考察的指标有项目投资、经营成本、占地面积、氮磷去除效果、污泥处理效果、运行稳定性、工艺成熟性和操作难易程度等8项。研究结果表明,对该污水处理厂而言,厌氧-单沟式氧化沟为最佳方案,其综合效益最高。污水处理厂传统的优化设计很少考虑模型不确定性参数的影响,故一般情况下产生的设计方案往往不是最优的。本文采用典型的Lawrence-McCarty模式(即泥龄法)对活性污泥系统进行优化设计,在这个基础上对模型参数分别开展了不确定性分析和灵敏度分析,考察各参数对系统优化设计的影响。对优化设计影响较大的参数有:进水水量和水质、与氧转移效率相关的校正参数以及二沉池的浓缩常数。这对于建立低灵敏度系统,对提高设计方案理论最优解的工程实用性有重要的指导意义。城市污水处理厂优化设计模型经过40多年的发展,在理论上渐趋完善,但由于模型本身的可解性差而使其实用性变差,尤其是不具有对设计方案进行动态优化调控的能力。区间数优化设计模型的表达和求解均以区间数的形式直接反映了系统的不确定性信息。决策者在决策时,可结合实际情况和各种新的信息,根据个人或集体经验、偏好在这一行为区间中确定具体行为方案,与传统优化模型相比更科学更实用,具有更大的灵活性和可操作性。但是由于模型的高维非线性,利用传统优化方法计算繁琐,且解的质量和计算效率对迭代初始点比较敏感。本文尝试将遗传算法与传统的步长搜索法相结合,对城市污水处理厂区间数优化设计模型的求解进行了探索。实例验证表明,这种新方法不涉及求导等复杂的数学运算,同时将大大提高区间解的搜索速度。与常规数值解法所得结果相比,目标函数总费用改善了20%左右,具有显著的经济效益。城市污水处理厂的污水量对生产合理调配具有很大的参考价值。它是系统内在和外在随机因素共同作用的结果,其预测准确与否不仅仅受外在随机因素的影响,更重要的是由系统内在动力学特征所决定。本文运用相空间重构理论,通过计算最大Lyapunov指数,判断城市污水水量时间序列存在混沌特性,可以短时预测。对长沙市第二污水处理厂污水量时间序列的相空间重构,确定延迟时间为2天,嵌入维数为9,最大Lyapunov指数为0.0274。据此确定神经网络的结构,输入层、输出层和隐含层节点分别为9,1和16。该法避开了传统BP网络结构设计依赖经验,易出现偏差的缺点。将整个时间序列10%的数据(36个)作为验证数据,其余90%(329个)作为训练数据。结果表明,建立的污水量短时预测混沌神经网络模型效果良好,最大绝对预测误差为0.2449,预测的平均误差为0.077。更多还原

【Abstract】 The planning, design and operation of municipal wastewater treatment plant is a series of decision-making processes with characteristics of multivariables and uncertain. Traditional optimization theories and methods failed to deal with intrinsic uncertainty effectively. Consequently optimization models were unilateral and limitative to a certain extent, so that it was difficult to popularize actual application. In the paper, uncartainty theories and methods were applied to optimization decision-making of municipal wastewater treatment plant, including establishment of cost function, optimal selection of wastewater treatment schemes, effect analysis of design parameter on optimal design, and solution of uncertain optimal design model, and nonlinear dynamic characteristic analysis and short-time forecast of influent flow, one of the most influencing parameter on operation of wastewater treatment plant. This research will of great significance for the design, operation and management for municipal wastrewater treatment plant.Not only the assumption of linearity for cost function but also the uncertainty resulting from collection and statistical analysis process of basic data, together with the inter-relationships among cost related factors, limited the application of traditional approaches in literatures. Considering the advantages of self-organizing, self-learning, and approximating arbitrary nonlinear continuous maps in theory, BP neural network was introduced into cost function of municipal wastewater treatment plant based on 26 data sets of Taiwan region. Design flow rate, influent BOD5 concentration, treatment degree and collection area were chosen as input variables of network, and total construction cost and plant construction cost as the output. The momentum constant and adaptive learning rate training algotithm was adopted to train network to overcome the problems of conventional BP algorithm such as falling into the local minima easily and converging slowly. Due to finite samples, leave-one-out method was applied. Early-stopping strategy was implemented to avoid over-learning. The comparative research revealed that the built BP neural network with strong learning capability outperformed multiple linear regression and can estimate cost of wastewater treatment effectively by reducing relative prediction error and improving the accuracy greatly.One important issue before design and construction of any municipal wastewater treatment plant is optimal selection of wastewater treatment schemes. Aimed at the multi-index, multi-objective and coexistence of certainty and various uncertainties involved in the optimization decision-making, this paper was intended to develop two novel approaches, that is hierarchy grey relational analysis and multi-attribute set pair analysis, to offset the shortages of conventional approaches such as considering the sole objective of minimizing system costs, and affected greatly by subjective factors, and to provide scientific basis for decision-making. The hierarchy grey relational analysis combined grey relational analysis with the idea of the hierarchy of analytic hierarchy process. It allowed for more effective reflection of actual characteristics of the problem as compared to mono level-based evaluation. In addition, the quantified evaluating scale, namely integrated grey relational grade, made wastewater treatment alternative selection more comparable and comprehensive. The multi-attribute set pair analysis developed the concept of identity - discrepancy - contrary connection degree, which should be regarded as structural function. Its unique stability analysis for basic ranking to determine other extra ranking could ensure the veracity and stability of decision results. The effectiveness of these two approaches was verified through case study. For some a municipal wastewater treatment plant, four alternatives (A2/O, triple oxidation ditch, anaerobic single oxidation ditch and SBR) were evaluated and compared against multiple economic, technical and administrative performance criteria, including capital cost, operation and maintenance cost, land area, removal of nitrogenous and phosphorous pollutants, sludge disposal effect, stability of plant operation, maturity of technology and professional skills required for operation and maintenance. The results showed that anaerobic single oxidation ditch was the optimal scheme and would obtain the maximum general benefits.Traditional optimal design of activated sludge system rarely took the parameter uncertainty into account, so that the obtained optimal solution was by no means the best for actual situation. In the paper, the typical Lawrence-McCarty mode, that is, sludge age-based design mode, was selected for optimal design of activated sludge system. Based on this, sensitivity analysis and uncertainty analysis were carried through to examine the influence of uncertain parameters on optimal design. Results indicated that influent flow and strength, parameters related with oxygen transfer efficiency, and thickening parameters of final clarifier had the most significantly influence on activated sludge system optimal design. This research was of great significance for the establishment of low-sensitivity system and the improvement of engineering practicability of theoretic optimal design solution.With the development of optimization theories during the past more than forty years, optimal design models of activated sludge system trended towards integrity. But it was difficult to obtain optimal solutions, as well as to provide dynamic modulation for design schemes. Both disadvantages weakened the practicability of optimal models. Interval optimization model can directly reflect the uncertainties that exist in actual systems, and a group of result intervals can be obtained from the solution of the model. According to personal or collective experience and prejudice, decision makers could determine detailed schemes in the result intervals combining with some other actual conditions. Obviously, interval optimization model was more scientific, applicable and operable than other optimization models. However, due to high-dimensional nonlinearity of interval optimization model, the solution process was laborious and time-consuming. In addition, the quality of ultimate solutions and calculation efficiency were sensitive to initial point for iteration. Genetic algorithm combined with classic step-search method was introduced to solve the interval optimization model for wastewater treatment plant. The verification by case study indicated that this synthetical optimization algorithm could improve searching efficiency distinctly without complicated mathematical operation such as derivative calculation. The objective, i.e. total cost, was reduced by 20 percent compared with that of classical numerical solution, which would bring great economic benefit.Influent inflow is important for the operation and control of municipal wastewater treatment plant, which is codetermined by both intrinsic factors and extrinsic stochastic factors. Its prediction accuracy not only depends on extrinsic stochastic factors, but greatly on intrinsic dynamic character. Phase space reconstruction theory was applied to calculate the largest Lyapunov exponents, based on which to recognized the chaos in influent flow time series. If there is chaos in time series, then the influent flow can be short-time forecasted. For influent flow of the No. 2 Changsha wastewater treatment plant, delay times were 2 days, and embedding windows were 9, and largest Lyapunov exponent was 0.0274. According to above results, structure of neural network was determined. The nodes of input, output and hidden layer were 9, 1, 16, respectively. This approach to determine the structure of neural network could avoid the disadvantages of dependency on experience and easy deviation. The data were divided into 36 validation data and 329 training data. Result showed that the built short-term forecasting chaos neural network for influent inflow performed well, and the largest absolute error was 0.2449 and the average prediction error was 0.077.更多还原