【作者】 吕青；

【导师】 谢克明；

【作者基本信息】 太原理工大学 ， 电路与系统， 2011， 博士

【摘要】 由两种或两种以上的煤按一定比例配合炼焦的工艺称为配煤炼焦。由于煤的组成、结构及性质的复杂性,使人们对煤及其焦化过程的认识具有局限性。现有的配合煤质量和焦炭质量的预测模型都是针对某一地区煤、某一焦化企业适用,不具有通用性。对于本文研究对象的焦化厂,直接使用这些模型并不适合,根据本厂多年积累的生产数据和配煤炼焦经验,建立自己的配合煤粘结指数模型和焦炭强度预测模型,并把它们应用在配煤比优化计算中,是本论文研究的重点。论文的主要内容如下：用各单种煤的粘结指数G与配比的加和来估计配合煤G值偏差很大,配合煤的粘结指数不满足加和性。论文除了使用粘结指数G指标外,还引入了单种煤的Vdaf指标。用高斯函数分别对焦煤的Vdaf和瘦煤的Vdaf为进行了非线性变换,与配合煤G的加和值一起构成了预测配合煤G模型的3个组成部分。对于增加了Vdaf指标的非线性预测模型中的7个参数采用思维进化算法来寻找最优解,确定配合煤G的回归模型。该模型的预测效果明显提高,相对误差不超过±6%。论文采用了核主成分分析方法(KPCA),在累计贡献率达到90%时,把5项配合煤煤质检测指标和3项备煤炼焦工艺参数指标特征提取为2个主成分,以这2个主成分作为支持向量机的输入变量,分别建立了焦炭机械强度M25和M10预测模型。并与主成分分析方法(PCA)方法的提取特征效果和建模效果进行了比较。结果表明基于KPCA-SVR焦炭强度预测模型的降维效果明显,推广‘性能好,预测误差小,满足实际生产需要。使用支持向量回归机(SVR)建立焦炭强度预测模型过程中,对训练样本进行筛选,提出了一种基于粒度的支持向量回归机样本选择策略。对训练样本集在核空间不同的粒层上进行预处理,在粗粒度层次上,排除异常样本数据(噪声数据);在细粒度层次上,根据粒度的稠密不同,采用不同的样本约减策略。在合适的粒度范围内,以小规模样本子集来表征整体训练集的分布,降低学习的代价,同时获得了较好的焦炭机械强度回归模型,提高了SVR算法的效率。SVR建立焦炭机械强度模型过程中,对于ε-SVR及核函数的参数确定,采用基于思维进化(MEA)的优化方法,该方法与网格搜索方法比较,MEA方法的寻优时间大量减少,而优化结果却近似。配煤比例的最终确定离不开炼焦试验。论文设计了20 kg铁桶试验,该配煤试验将铁桶置于工业焦炉中随炉成焦,铁桶内煤的成焦与工业焦炉成焦保持了相同的环境,该方法投资小、操作方便、劳动强度低、试验调节灵活。经检验当煤饼与铁桶之间间隙为23mm,堆密度为1.2t/m3,结焦时间在24-48小时范围内情况下,20 kg铁桶试验焦炭与工业焦炉焦炭的机械强度相关性强,两者关系满足一元线性回归方程,20 kg铁桶试验可较好地预测工业焦炉的焦炭强度。论文综合考虑配合煤的价格成本和焦化产品的收益,选择了配比优化模型的目标函数,允分体现了焦化(?)的效益最大。基于给定的焦炭质量指标和配合煤质量指标、煤场现有的各大类煤种质量参数和其价格及一定的备煤和炼焦条件,采用MEA算法优化配煤比,并把论文所建立的配合煤粘结指数预测模型和焦炭机械强度预测模型应用在配比计算过程中。与标准遗传算法(GA)方法比较, MEA方法的效率要高,20 kg铁桶试验验证MEA方法所优化的配煤比例科学合理。更多还原

【Abstract】 The technology of two or more different coals in certain proportion blending coking is called Coal Blending Cokemaking. Because of the complexity of the composition, structure and nature of coal. man’s knowledge of coal and it’s coking have many limitations.The cur-rently available prediction models of blend coal quality and coke quality are suitable for a certain area coal or a certain coking enterprises, which are not common. For the researched coking plant of the paper, using these existing model doesn’t fit. So according to the factory’s production data and coking experience accumulated for many years, establishing blend caking index model and the coke mechanical strength model of it’s own, and applying the models to optimizating coal blending proportion are studied in the paper. The main content of the paper is listed as follows:The caking index G of blend coal can be incorrectly predicted with the additive value of the every single coal G and its proportion, additivity of blend G is not satisfied. Despite the indicator of caking index G. the paper also introduces the Volatile Vdaf indicator. Nonlinear transform is made for the coking-coal Vdaf and the lean-coal Vdaf by Gaussian function, which together with the blend G additive value, formed three components of predicting the blend G. The paper uses the MEA algorithms to find the optimal solution of the 7 parameters of the previous nonlinear model. After establishing the regression predicting model, the predicted effects of the blend G model is improved obviously and the relative error is not more than±6%.When the accumulative contribution rate is more 90%, KPCA extracts the 2 principal components from 5 blend quality indicators and 3 coking technology indicators. These two principal components employed as the inputs of SVM, we respectively set up the coke mechanical strength prediction models of M25and M10. Be compare with the effect of feature extraction and modeling of PCA, the coke strength prediction models based on KPCA-SVR have drastic dimension reduction. good generalization performance, less forecasting error and can meet the requirement of actual production.During the establishment of the coke strength model by SVR, A samples selection method based on G-SVR (granularity support vector regression) is proposed for filtering training samples. We pretreat the sample sets in different granulose of the kernel space. Abnormal sample data (noise data) are excluded in the coarse-granularity level and partial dense repeated samples are removed in fine-granularity level. In the reasonable range of granularity, the distribution of the original samples substituted by a small sample subset is not changed, while the learning experience is reduced, the coke prediction model is good and the efficiency of SVR algorithm is improved.In the process of building the coke mechanical strength SVR model, the optimizing method based on MEA is applied to determine the parameters ofε-SVR and Kernel function. Be compare with the mesh scanning algorithm, the optimization time of MEA decreases greatly but the optimization results are similar.The coal blending proportion can not be set without the coking test. The paper designs a 20 kg iron drum experiment that put iron drum into the industrial coke oven for coking, the coal of iron drum and the coal of industry oven are burnt in the same environment. This iron drum test method has small investment, easy operation, less intensity of labour and flexible trial adjustment. Upon examination, when the space between the iron drum and the coal pie within the iron drum is 23mm, the bulk density is 1.2t/m3 and the burned time is in 24～48 hours, There is a strong correlation between the 20 kg iron drum coke mechanical strength and the industrial oven coke mechanical strength. The relationship of the two mechanical strengths satisfies the one variable linear regression equation. The 20 kg iron drum experiment can predict the industrial oven coke mechanical strength perfectlyConsidering the cost price of blend coal and the income from coking products, the paper chooses the objective function of the coal blending ratio optimization model, which fully embodies the biggest benefits of the coking plant. Setting coke quality indicators and blend quality indicators, knowing the major existing single coals, its quality parameters, prices, and the conditions of pretreating coal and coking, a MEA algorithm is adopted to optimize coal blending ratio. The prediction models of blend G and coke mechanical strength established in the paper are applied in the process of optimizing blend ratio. The MEA method is higher retrieval efficiency than GA method and the results of the 20 kg iron drum experiment show that the blend ratio optimized by MEA method is scientific and reasonable.更多还原