基于在线增量LSSVM污水软测量建模研究与应用

发布时间：2018-05-24 13:34

本文选题：软测量 + 最小二乘支持向量机　；参考：《安徽工业大学》2017年硕士论文

【摘要】：当今社会,淡水资源的短缺和水体环境的污染时刻提醒着我们污水处理的必要性。污水处理是具有强非线性、大滞后、大惯性等特点的复杂现代工业过程,其复杂多样的工艺过程导致污水中一些反映水质指标的重要过程参数的实时检测以及根据过程变量进行预判输出变得十分困难,现有的检测仪器或因为成本、或因为精度,都不能给予很可靠的保证。然而软测量技术作为传统检测技术的延伸和发展,应用于实际的污水处理过程中具有很重要的现实意义。本文以曝气生物滤池为研究对象,建立了污水软测量模型,实现了对出水COD浓度的预测,为判断出水质量以及做出相应措施提供超前的判据。主要内容如下:首先,本文基于LSSVM相对于传统SVM所具有的优势,选取LSSVM作为本文模型的基础算法,并且分别从模型核函数和超参数两个方面进行改进,一方面采用线性加权的混合核函数取代单一核函数;另一方面采用改进的智能优化算法对参数组合进行寻优。仿真对比分析,两方面的改进对模型精确度都有一定程度的提升。然后,针对离线模型面临大规模实时更新的数据时整体预测效果会随之变差的情况,本文采用了一种改进的在线增量LSSVM算法,该算法采用预测误差进行选择性增量学习和配套的剪枝算法,即新样本添加时,根据预测误差阈值判别是否需要增量学习,针对超过阈值的样本进行增量式学习,用递归的方式不断更新支持向量,避免了标准的求逆运算,缩短了运算时间,改善了预测的效率。同时当样本规模达到一定规模时选择剪去最早样本的剪枝操作,使样本长度始终维持在一定规模。本课题的主要贡献及创新点在于:本文选择从精确度和在线性能两方面对LSSVM污水软测量模型进行改善。线性加权的混合核函数和改进的智能优化算法的结合,使模型精确度得到很大程度的提升;改进的在线增量LSSVM算法通过巧妙的迭代、阈值的选择、时间窗的设置,使模型的快速性、在线性都有了一定程度的改善和提高;两方面的改善共同实现了对出水COD精准的在线预估。
[Abstract]:Nowadays, the shortage of fresh water resources and the pollution of water environment remind us of the necessity of sewage treatment. Sewage treatment is a complex modern industrial process with the characteristics of strong nonlinearity, large lag and large inertia. Because of its complex and diverse process, it is very difficult to detect some important process parameters in wastewater and to predict the output according to the process variables. The existing testing instruments are either because of the cost or because of the accuracy. Can not give a very reliable guarantee. However, as the extension and development of traditional detection technology, soft sensing technology has important practical significance in the process of actual sewage treatment. In this paper, the biological aerated filter is taken as the research object, the soft sensing model of sewage is established, and the prediction of effluent COD concentration is realized, which provides an advanced criterion for judging the water quality and making corresponding measures. The main contents are as follows: firstly, based on the advantages of LSSVM compared with traditional SVM, this paper selects LSSVM as the basic algorithm of the model, and improves the model from two aspects: kernel function and hyperparameter. On the one hand, the linear weighted hybrid kernel function is used to replace the single kernel function; on the other hand, the improved intelligent optimization algorithm is used to optimize the parameter combination. The simulation results show that both improvements improve the accuracy of the model to a certain extent. Then, an improved on-line incremental LSSVM algorithm is used to solve the problem that the overall prediction effect will become worse when the offline model is faced with large-scale real-time update data. In this algorithm, selective incremental learning with prediction error and pruning algorithm are used, that is, when new samples are added, incremental learning is carried out according to the threshold value of prediction error to determine whether incremental learning is required for samples that exceed the threshold. The support vector is updated recursively, which avoids the standard inverse operation, shortens the operation time and improves the efficiency of prediction. At the same time, when the sample size reaches a certain scale, the pruning operation of the earliest sample is selected to keep the sample length at a certain scale. The main contributions and innovations of this paper are as follows: in this paper, we choose to improve the LSSVM sewage soft sensor model from the aspects of accuracy and on-line performance. The combination of the linear weighted hybrid kernel function and the improved intelligent optimization algorithm greatly improves the accuracy of the model. The improved online incremental LSSVM algorithm is improved by clever iteration, selection of threshold, setting of time window. The rapidity of the model has been improved to a certain extent in linearity, and the two improvements have realized the accurate on-line prediction of effluent COD.
【学位授予单位】：安徽工业大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：X832;TP18

【参考文献】