面向节电生产的铝电解电流效率预测模型研究

CORC > 沈阳自动化研究所 > 中国科学院沈阳自动化研究所 > 数字工厂研究室

题名	面向节电生产的铝电解电流效率预测模型研究
作者	钟新成
学位类别	硕士
答辩日期	2016-05-25
授予单位	中国科学院沈阳自动化研究所
导师	朱军 ; 刘昶
关键词	铝电解电流效率预测模型 KFCM LSSVM copula分布估计算法
其他题名	Research on Forecasting Model of Energy Aluminum Electrolytic Current Efficiency for the Production of Energy-Saving
学位专业	模式识别与智能系统
中文摘要	铝电解是传统的高能耗过程，面对当前严峻的竞争以及能源形势，节能降耗已成为其首要目标。电解铝企业要达到节电生产的目的，归根结底是要降低吨铝直流电耗，而吨铝直流电耗和电流效率是成反比的关系，因此提高电流效率便能达到节电生产的目的，同时还可以指导出铝和下料管理。若能实现对电流效率的在线监测，便可为铝电解的生产提供决策依据。但是铝电解槽是一个十分复杂的时变系统，许多电解槽参数难以实现在线测量，电流效率便是其中之一。针对以上情况，本文提出一种基于copula分布估计算法（cEDA）优化的模糊核聚类(KFCM)最小二乘支持向量机(LSSVM)的铝电解电流效率预测模型。该方法可以简述为KFCM用来对数据进行分类、LSSVM完成非线性回归、cEDA用于对模型参数的优化。此外，在建模过程中还加入了动态学习以增强模型的自适应性。具体研究内容如下。首先，对近年来电解槽参数软测量及电流效率测量方面的研究现状进行了综述，在结合他们研究方法的优缺点并深入分析电流效率影响因素的机理的前提下，确定了模型输入的辅助变量，提出了本文关于电流效率软测量的建模方法。其次，由于电解槽存在不同的槽况，同种槽况采集到的数据会具有相似的信息。若忽视此点而用单一的模型进行预测，不仅会加大模型的复杂度还会丢失部分数据信息。针对该情况，提出用数据挖掘中的模糊核聚类算法对数据进行分类。虽然KFCM将输入变量映射到高维空间加大了样本的差别，一定程度上减弱了对变量分布情况的依赖，但在数据差别很大的情况，会出现小数据类被误分的情况。为解决这一问题，本文加入动态权值对其进行改进。即通过动态加权，自动削弱那些具有较少元素的类对分类的作用，从而达到改善分类的效果。仿真结果表明，本文提出的改进算法使数据样本的划分更加合理。再次，对聚类后的各样本子集分别建立LSSVM子模型然后得到相应的回归函数。由于LSSVM中超参数选取的好坏对预测精度的影响很大，所以本文引进cEDA对其进行参数寻优。但是cEDA在搜索过程中容易陷入局部最优，本文通过对其种群进行重新整合，大大改善了算法的寻优性能以及收敛速度。最后，在结合前面内容的情况下来构建电流效率的整体预测模型。预测过程中，当输入一组数据时，首先采用改进的KFCM计算其对于每类的模糊隶属度函数，然后利用模糊隶属度值最大的那一类的回归函数进行预测。此外，为了实现该模型的自适应性，需要对该子模型进行动态学习得到新的回归函数。仿真结果表明，本文建立的预测模型无论在预测精度还是泛化性能上都有很大改善，能为铝电解的生产提供决策依据。得到电流效率的预测值后能更好的指导铝电解的实际生产。
英文摘要	Aluminum electrolysis is a traditional energy intensive process. Considering the serious energy and competition situation, the energy saving has be promoted to be the primary target of this process. In order to achieve the purpose of energy-saving production, the final analysis is to reduce the tons of aluminum DC consumption. While the tons of aluminum DC power consumption is inversely proportional relationship with current efficiency, thus increasing the current efficiency will be able to achieve the purpose of energy-saving production. If we can achieve the current efficiency of online monitoring, it can provide the basis for decision making aluminum electrolytic production. But aluminum cell is a very complex time-varying systems, many parameters are difficult to achieve cell-line detection, the current efficiency is one of them. For the above, we propose a copula-based estimation of distribution algorithms (cEDA) optimized KFCM with squares support vector machines (LSSVM) electrolytic current efficiency forecasting model. This method can be summarized as KFCM used to classify data, LSSVM complete non-linear regression, cEDA used to optimize the model parameters. In addition, the modeling process also added to enhance the adaptive dynamic learning model. Specific studies to accommodate follows. First, the cell parameter soft measuring current status of research in recent years, and efficiency measurements were reviewed, their advantages and disadvantages in combining research methods and in-depth analysis of the mechanism of current efficiency factors of the premise, to determine the model of the auxiliary input variable modeling method proposed in this paper on the current efficiency of soft measurement. Second, due to the presence of grooves different cell conditions, the same kind of grooves status data collected will have similar properties. If you ignore this point and use a single model to predict not only will increase the complexity of the model but also will lose some data. For this case, proposed a method of fuzzy kernel clustering algorithm for data classification. Although KFCM got input variable mapped into a high dimensional space and increases the difference between samples, to a certain extent reduced reliance on variable distribution, but when the data is very different, the small class will be misclassified. To solve this problem, add a dynamic weights to improve it. Through dynamic weighting, automatically weaken those classes which has less effect on the classification of elements, so as to achieve the effect of improving the classification. Simulation results show that the improved algorithm is proposed to make a more rational division of data samples. Again, for every book set after the cluster were established LSSVM sub-model and the corresponding regression function. Since the hyper parameter selection of LSSVM has large influence on prediction accuracy, so use cEDA to optimize its parameter. But cEDA is easy to fall into local optimum in the search process, the paper re-integration of its population which has greatly improved the performance of the algorithm optimization and convergence rate. Finally, overall current efficiency forecasting models was constructed after combining with the previous contents. Forecasting process, when the input data set, the first using an improved KFCM calculated for each class of its fuzzy membership function and fuzzy membership functions return the maximum value of the kind of forecast. Furthermore, in order to achieve self-adaptability of the model, the sub-model needs to get a new function through dynamic learning. The simulation results show that the prediction model established in this paper has greatly improved in terms of predictive accuracy or generalization performance and it can provide decision making for aluminum electrolytic production. It is better to guide the actual production of electrolytic after getting a real-time predictive value of the current efficiency.
语种	中文
产权排序	1
页码	64页
内容类型	学位论文
源URL	[http://ir.sia.cn/handle/173321/19632]
专题	沈阳自动化研究所_数字工厂研究室
推荐引用方式 GB/T 7714	钟新成. 面向节电生产的铝电解电流效率预测模型研究[D]. 中国科学院沈阳自动化研究所. 2016.