【作者】 王祎；

【导师】 王鹏；

【作者基本信息】 哈尔滨工业大学 ， 环境科学与工程， 2013， 博士

【摘要】 依靠数字模拟或概念模型等手段在宏观层面上改善河流水质，不仅可以节省大量人力物力，也可以实时对目标问题进行快速高效的分析解决。针对目前我国河流水质管理中存在的主要技术问题，本文在调研国内外研究现状的基础上，以松花江哈尔滨段为研究对象，利用计算智能方法等数值模拟手段开展了入河污染源识别模式、河流水质预测模型和入河排污削减模式的研究。基于专家系统思想研究建立了入河污染源识别模式。结合河流常规水质监测，建立了基于污染超标初筛和多元数据挖掘技术的核心水质指标解析机制，可对常规监测矩阵解析出代表污染信息的污染因子及其荷载核心水质指标。根据专家知识构建了污染源类别判定树及其22类产生式规则，基于53则国家行业水污染排放标准，设计建立了针对工业点源污染因子的行业相似度和污染贡献指数。该模式可识别出对目标水域影响较大的点源污染行业并排序。对松花江哈尔滨段各类水域进行了行业污染源识别，影响非城市水域较大的点源污染行业主要有制药和铁矿采选工业；影响阿什河内复杂高污染水域的上游点源污染行业主要有制药、陶瓷、炼焦和造纸等工业，根据资料和污染源调查，验证识别结果合理。大顶子山工程竣工后，影响过城主干流水域的污染点源主要有制药工业、污水厂、铁矿采选和电镀工业等，相比竣工前，点源行业贡献指数降低，非点源污染因子数量增加且污染因子总数减少，表明工程的建立在一定程度上强化了水体的自净作用，产业结构政策调整或环保政策的下达实施也颇有成效。基于双重Bootstrap方法和小波技术，本研究建立了适用于常规水质时间序列预测及其综合不确定性评价的人工神经网络模型（BWNN）。模型以Morlet基函数作为ANN激活函数，水质监测时滞作为模型输入。BWNN对朱顺屯氨氮和DO序列的预测结果满意，准确性指标NSE分别可达0.9678和0.8556，验证表明BWNN比ARIMA和其他传统ANN等模型的预测准确性和可靠性更高，对于季节性较好的氨氮序列，模型结构带来的不确定性所占比例更多；对非季节性或序列波动较大的DO序列，数据带来的不确定性所占比例更多；北方冰封河流枯水期的平均综合预测不确定性比其他水期要大10-20%。以BWNN为基础，开展了水质时间序列预测方案优选研究。BWNN对缺失达8.3%的无插补序列预测结果仍为满意，对季节周期性较好的缺失序列，应首选时间指数平移滑动平均插补策略；对季节周期性不好的缺失序列，应首选空间多重插补策略，当同类空间区域数据不易获取时，局部平均插补可以作为优先备选策略。对有监测缺失的大顶子山断面的水质进行了预测，结果表明在未来的一段时间，COD、氨氮和TP的预测值均能达到IV类功能区标准，但预测区间上限在冬季仍有超标。基于水环境容量控制思想和计算智能方法，设计建立了用于河流水质管理的入河排污削减模式。利用Bootstrap法和ANN建立了实用性良好的排污源强-河流水质关联模型，仅需各目标断面的实际水质监测数据和相关的排污口数据，生成的预测区间可保证削减管理的安全性，适用于多种情形。根据大中型河流和小型河流的特点，分别研究构建了单功能区排污口优化削减模式和多目标多功能区排污口联合优化削减模式，并基于智能优化技术设计了两种模式的优化算法。单功能区排污口优化削减模式可直接对排污口和目标断面水质监测数据进行反向模拟与优化，在流域尺度内具有较好的普适性；多目标多功能区排污口联合削减模式可有效产生满足系统水环境容量和削减成本目标的若干个非劣解，可根据实际问题从中选择可行解。研究区域氨氮削减实例表明，该模式可为流域水质管理提供技术支撑，并有效提高流域常规水环境管理工作的系统性。更多还原

【Abstract】 Improving river water quality in management aspect by digital simulation modelsnot only can save resources const, but also effectively achieve the task goal. Tosolve the main technical problems of river water quality management, models foridentifying pollution sources, forecasting water quality and allocating the reductionof outlet loads, were studied based on computational intelligence (CI) methods,through investigating the current research situation at home and abroad, with theSonghua River-Harbin section as the study area.The model for identification pollution sources to river water was establishedbased on expert system. Through the normal water quality monitoring matrix, aninterpreting mechanism was built by preliminary screening and multivariate datamining technique, by which major pollution factors and their loading water qualityindices could be obtained.Then a decision tree model was constructed based on the22classes of produced rules by experts. According to the53national industrialwater pollution discharge standards, a similarity matching model for identifying thepotential point sources of industrial pollution was established, which can calculatethe contributing rate of a certain industry on target area. The model was applied tothe Songhua River Harbin section and for outside city group, the main point sourcesordinally were medical and pharmacy、iron ores dressing. For high pollution group,the major five were pottery and porcelain production, chemical coking, traditionalChinese medicine pharmacy and pulp and paper industry depending on thecontributing rate. The validation results were reasonable. For mainstream groupafter Dadingzishan Dam Project (DDP), the main five were pharmacy, sewage plant,iron ores dressing and electroplating industries. Compared to that before DDP, thecontribution rates of point sources were decreased and the number of non-pointsources factors increases, suggesting the DDP strengthened the self cleaningcapacity of river water body and the environmental policy was also effective.For predicting monthly water quality series and assessing the total uncertainty,a Bootstrap wavelet neural network model (BWNN) was developed based on re-Bootstrap method. The Morlet wavelet basis function (WBF) was employed asnonlinear activation function and the time-lag orders were set as the input variables. Performances of BWNN models were satisfactory and were better than ARIMA andother ANN models for NH4+-N and DO in Zhushuntun station, with respect NSE0.9678and0.8556. The uncertainty from data noise was smaller than that frommodel structure for seasonal NH4+-N series; conversely, the uncertainty from datanoise was larger for unseasonal DO series. Besides, total uncertainties in the low-flow period were10-20%bigger than other flow periods.The optimal data missing-refilling scheme was studied for the above BWNN.Performances of BWNN were still satisfactory when the missing percentage was8.3%. Temporal method was satisfactory for filling seasonal series, whereas spatialimputation was fit for unseasonal series and the local average infilling could be aneffective alternative when the spatial data were unavailable. The case study ofDadingzishan station showed that the predictions of COD, NH4+-N and TP couldmeet the national water quality standard IV, but the upper band of predictionintervals violated the standard in some winter months.Loads reducing allocation model of regional outlet discharge was establishedbased on water environment capacity control and the CI method. A linkage modelwas developed by bootstrap and ANN, which only needed water quality monitoringdata and discharge loads data. However, obtained prediction intervals couldguarantee the margin of safety in environmental management. Considering thecharacters of big and small rivers, reduction allocation models for single waterfunctional area (RMS) and multi-functional area (RMM) were developed,respectively. Then two intelligent optimizing algorithms were created for solvingeach model. The RMS could inversely simulate and optimize the allocation ofreduction rate of discharge loads and it had good applicability in watershed scale;the RMM could generate many noninferior solutions that could meet the balancebetween water environment capacity and reduction cost, then some solutions couldbe chosen as feasible ones for the practice. The results of NH4+-N case studyshowed that the two models could provide effective scientific supports forwatershed water quality management and would strengthen its systematicness.更多还原