铝电解关键指标预测方法的研究与应用
文献类型：学位论文
| 作者 | 陈勇; 周中寒
|
| 学位类别 | 硕士 |
| 答辩日期 | 2018-05-17 |
| 授予单位 | 中国科学院沈阳自动化研究所 |
| 授予地点 | 沈阳 |
| 导师 | 周晓锋 |
| 关键词 | 铝生产 模糊聚类 朴素贝叶斯 累积法 增量思想 |
| 其他题名 | Research and application of prediction method for key indicators of aluminum electrolysis |
| 学位专业 | 机械制造及其自动化 |
| 中文摘要 | 为了解电解槽在日常生产中的槽况规律，维持铝电解生产的持续性，保证电解槽的物耗稳定和能耗稳定，本文通过对铝厂数据的挖掘与建模，提出一整套维持电解槽稳定的策略方法，并用于指导实际生产。通过对于相关内容的研究现状总结分析，对铝厂数据和生产过程的调研，本文的主要研究内容如下：(1) 首先对所有槽的半年数据进行融合，对融合后的数据做缺失值填补、去噪等预处理工作，得到较完整干净的数据。(2) 针对铝厂数据分布特征未知的特点，结合现有模糊聚类算法的优缺点，发现模糊C聚类算法存在一些可优化的方向，提出一种无参的自适用模糊C聚类算法（Self Applying Fuzzy C-means algorithm without reference, SANR-FCM），通过迭代自适应得到类簇个数和簇中心，算法改进了传统模糊C聚类算法的损失函数，不仅考虑了点对簇的隶属关系，而且添加了熵项来控制加权指数和隶属度，以此通过迭代过程中隶属度和类簇的实际情况自动更新隶属度和类簇个数，解决了模糊C聚类的初始化参数影响聚类结果的问题。算法在UCI数据集和铝厂数据集上皆表现出了良好的实验效果。(3) 根据聚类结果，将铝厂数据按实际意义标签化，提出一种基于距离相似性度量的动态朴素贝叶斯算法(Distance Dynamic Naive Bayes-DNB)，算法结合铝厂数据的实际特点，总结朴素贝叶斯在连续型数据上应用的缺陷，以及非参数估计在铝厂数据上计算分布函数的效果不佳等因素，改进的朴素贝叶斯算法给出了一种基于点与其余点距离作为点与簇的重要性考量的方法，通过点与点得距离描叙点对簇的重要程度，并对分类器使用增量思想，使算法动态分类准确率得到提高。使用UCI数据和铝厂数据对改进算法进行验证，实验证明改进的朴素贝叶斯算法在铝厂数据上表现较非参数估计和朴素贝叶斯算法有较稳定和准确的表现。(4) 应用单槽历史数据，结合构建的分类器，通过累积法完成当天各指标等级趋势的预测，并确定各指标下变量相对于前一天的变化量，完成相关指标的预测。实验发现，整个预测模型可完成铝电解关键指标的预测，为铝厂日常生产规划提供数据支持；提出的无参的自适用模糊C聚类算法可以在无初始化参数的同时完成聚类、基于距离相似性度量的动态朴素贝叶斯算法在UCI数据及铝厂数据上变现稳定且良好。 |
| 英文摘要 | This paper presents a strategy to maintain the stability of aluminum reduction cells through the data mining and modeling of the aluminum factory in order to maintain the continuity of aluminum electrolytic production and ensure the stability of energy consumption and the stability of energy consumption. The research on the data and production process of aluminum plant is made through the summary and analysis of the related content. The main contents of this paper are as follows: (1) First, we integrate all the six months' data of all slots, and do preprocessing of missing values filling and denoising for the merged data to get more complete and clean data. (2) In view of the unknown data distribution characteristics of aluminum plants, combined with the advantages and disadvantages of the existing fuzzy clustering algorithm, we find that there are some optimized directions in fuzzy C clustering algorithm. A self applying fuzzy C clustering algorithm without reference is proposed. The number of cluster and cluster centers are obtained by iterative adaptive algorithm. The algorithm improves the loss function of the traditional fuzzy clustering algorithm C: it is not only consider the point of cluster membership, moreover the entropy term is added to control the weighted index and membership degree, so as to automatically update membership number and cluster number through the actual situation of membership and cluster in iteration process. The problem of clustering results which is influenced by the initialization parameters of fuzzy C clustering is solved. The algorithm shows good experimental results on the data set of the UCI data set aluminum plant. (3) According to the result of clustering, the data of aluminum factory is labelled according to the actual meaning, and a dynamic naive Bayes algorithm based on distance similarity measure is proposed. Combined with the actual characteristics of aluminum plant data, this paper summarizes that naive Bayes' Application on continuous data makes defects, and the experimental results of nonparametric estimation on the data of aluminum plants fitting the distribution function are not good. The improved naive Bayes algorithm gives a method for the importance of points and clusters based on the distance between points and other points. The algorithm points that the importance of point to cluster can be described as point and point distance. The incremental idea of the classifier is used to improve the accuracy of the dynamic classification of the algorithm. UCI data and aluminum factory data are used to verify the improved algorithm. The experiment shows that the improved naive Bayes algorithm is more stable and accurate than the non parametric estimation and the naive Bayes algorithm on the aluminum plant data. (4) Using single slot historical data, combined with the constructed classifier, we can predict the trend of each index grade by accumulating method, and determine the variation of variables under each index relative to the previous day, and then complete the prediction of related indicators. It is found that the whole prediction model can predict the key indexes of aluminum electrolysis, and provide data support for the daily production planning of aluminum plant. The proposed self applicable fuzzy C clustering algorithm can complete the clustering without initialization parameters. The dynamic naive Bayes algorithm based on distance similarity measure is stable and good in UCI data and aluminum factory data. |
| 语种 | 中文 |
| 产权排序 | 1 |
| 页码 | 71页 |
| 源URL | [http://ir.sia.cn/handle/173321/21771]  |
| 专题 | 沈阳自动化研究所_工业控制网络与系统研究室 |
| 推荐引用方式 GB/T 7714 | 陈勇,周中寒. 铝电解关键指标预测方法的研究与应用[D]. 沈阳. 中国科学院沈阳自动化研究所. 2018. |
入库方式： OAI收割
来源：沈阳自动化研究所
浏览0
下载0
收藏0
其他版本
除非特别说明，本系统中所有内容都受版权保护，并保留所有权利。