混煤硫释放的BP神经网络模型预测

通过对不同混煤一维燃烧过程中H2S和SO2释放特性的有关数据,应用BP人工神经网络进行预测.通过分析和计算建立了典型的三层BP网络,输入神经元为8个,隐含层神经元个数为6个,输出层神经元个数为2个,用加入动量项的方法对传统的BP网络算法进行改进,通过样本数据训练,测试数据检验,该网络能够较为准确地预测混煤一维燃烧硫释放的情况.     

嵌入岭回归的BP算法及其在软测量中的应用

针对三层神经网络(ANN)最佳隐节点个数难以确定和随着隐节点个数增加ANN模型易出现过拟合等缺点,提出了嵌入岭回归(RR)的误差反传算法(BP).BP-RR根据样本规模自适应确定隐节点个数,并通过BP算法充分提取样本数据信息.然后,针对隐含层输出可能存在的复共线性,采用RR以预测性能为指标,通过进化算法确定最佳岭参数,进而重新确定隐含层与输出层之间最佳的权值和阈值,克服ANN过拟合,建立具有良好预测性能的模型.将BP-RR应用于建立石脑油干点软测量,结果显示,BP-RR模型具有良好的预测性能.与ANN相比,BP-RR模型鲁棒性强,预测精度高.     

细颗粒物在旋风分离器内流动特性

为了研究操作条件和颗粒特征对细粉体颗粒在旋风分离器中流动状态的影响,文中采用欧拉-拉格朗日方法对细颗粒在旋风分离器内的停留时间进行模拟计算。探究入口气速、粉体粒径和固气比对粉体在旋风分离器内的停留时间分布密度函数、平均停留时间和停留时间量纲一方差的影响。结果表明：粉体颗粒在旋风分离器内部的停留时间分布密度函数总体呈正态分布;当入口气速u=10 m/s时,粉体颗粒平均停留时间为0.40 s,停留时间量纲一方差为0.11,此时粉体颗粒平均停留时间最长,颗粒运动状态更接近平推流;在模拟工况范围内,随着粉体粒径增大,停留时间量纲一方差增大,颗粒在旋风分离器内的返混程度加剧;颗粒粒径和固气比变化对平均停留时间影响较小。     

高温旋风分离分级效率的理论计算及其分析

旋风分离器是高温除尘技术中优先考虑的预除尘设备。以平衡尘粒模型以及相似理论原理分析为基础,给出了该分离器高温旋风分离的分级效率的计算方法,对反映其在高温状态下分离特性的参数分割粒径dc50和分布指数m进行了分析。以Stairmand高效旋风分离器为例,分级效率的理论计算结果与高温旋风分离的试验值进行了比较对照。     

旋风分离器内气相流场的相似模化分析（Ⅱ）尺寸参数

基于数值模拟方法分析了几何相似旋风分离器在直径尺寸变化时的气相流场的模化特性。旋风分离器的直径变化范围为100～2100 mm，重点考察流场量纲1速度分布的相似性、尺寸参数的自模特性。模拟结果表明，旋风分离器的直径尺寸变化后，流场的形态结构不变，但量纲1切向速度分布存在一定变化，流场不具有严格的自模性。但随直径尺寸的增大，旋风分离器内流场由弱自模性向强自模性发展，当直径超过2000 mm后，流场量纲1速度基本保持相似与恒定，Euler数与Reynolds数不相关，流场处于强自模区。流场的这种自模特性反映了流场的惯性阻力损失和黏性摩擦损失之间的比例关系和大小变化。随着直径尺寸的增大，黏性摩擦损失系数减小，惯性阻力损失系数增大，但Euler数是减小的。当达到很大的直径时，Euler数恒定，Euler数与Reynolds数不相关，流场进入自模区，流场流态保持相似性。     

旋风分离器内气相流场的相似模化分析（Ⅰ）流动参数

基于数值模拟方法分析了流动参数变化对旋风分离器内气相旋转流流场模化特性的影响。选择直径φ300 mm的旋风分离器为计算模型,在入口气速0.5～30 m·s^-1、气体温度293～1273 K和操作压力0.1～6.5 MPa范围内,分析了流动参数变化对流场量纲1速度分布相似性的影响,考察了常规参数模型与高参数原型之间的流场模化关系。模拟结果表明,流动参数变化后,虽然旋风分离器内旋转流流场仍呈现准自由涡与强制涡分布形态,但流场的量纲1速度分布不能保持完全恒定和相似,存在着一定的变化。然而当入口速度高于20 m·s^-1,或压力高于3.0 MPa时,流场量纲1速度可以基本保持恒定,不再随流动参数变化而改变,Euler数与Reynolds数不相关,近似处于流场的自模区;当温度超过1000 K后,Euler数基本不受Reynolds数变化的影响,流场也可以保持相似性。最后从能量损失的观点,讨论了Reynolds数和Euler数之间的相互关系,分析了流动参数对旋风分离器气相旋转流流场自模特性作用的机理。     

化肥厂并联布置的旋风分离器组的优化研究

通过研究旋风分离器的特性,了解筒体直径和压降是影响旋风分离器成本的主要因素,建立了并联布置的旋风分离器组的成本模型。取某一化肥厂的实际参数进行实例计算,得出通过减少旋风分离器的个数、合理增大旋风分离器的筒体直径、改变旋风分离器的局部结构,可在一定程度上降低成本。分析成本模型的特殊性,通过数值计算,可以得到基于总成本最小情况下的最优旋风分离器筒体直径,为旋风分离器组的优化设计提供参考。     

GA-BP神经网络的钛合金微弧氧化膜层厚度预测模型的建立

利用正交试验法获得的TC4钛合金微弧氧化实验数据建立了基于4-11-1(即4个输入神经元,11个隐含层节点,1个输出神经元)结构的BP神经网络预测膜层厚度的模型,并引入遗传算法(GA)对其权值和阈值进行优化。以微弧氧化工艺参数中的电流密度、脉冲频率、占空比和氧化时间作为网络的输入向量,氧化膜层厚度作为网络的输出向量,对比和分析了BP与GA-BP模型的预测结果。与BP网络模型相比,GA-BP网络模型稳定性能较好,并能高精度预测膜层的厚度,GA-BP网络模型预测值的平均误差为0.015,最大误差仅为0.036,而BP模型预测结果的平均误差为0.064,最大误差为0.099。     

旋风分离器压降及分离效率计算模型

介绍了旋风分离器的基本工作原理和主要用途,以及旋风分离器的主要改进与发展方向,叙述了旋风分离器压降和分离效率的理论计算模型,并对不同模型进行了相应的分析及比较。对未来旋风分离器的应用和性能进行了展望,认为未来旋风分离器的改良将向着在标准旋风分离器上添加额外部件的方向发展,改进型旋风分离器将打破旋风分离器技术不能有效分离5μm以下粒径颗粒的传统限制。     

旋风分离器锥体局部磨损对内流场及性能的影响

为了探究局部磨损对旋风分离器性能的影响,采用Oka磨损方程以及计算流体动力学(CFD)对旋风分离器壁面磨损以及内流场特性进行数值模拟,考察旋风分离器内流场等参数随磨损厚度增大的变化规律。结果表明,旋风分离器壁在锥体底部排尘口附近壁面局部磨损严重,形成螺旋形冲蚀磨损带。磨损引起设备几何结构的改变会导致切向速度降低,颗粒所受离心力降低,锥体内局部涡强度及影响范围增大,涡核旋进(PVC)的影响加大,不利于主流的稳定与固体颗粒的分离。与未磨损时相比,当局部磨损厚度为20 mm时,粒径为4μm的颗粒分离效率由100%降低至93%,分割粒径由1.3增大至1.9μm,设备压降降低了约40%。研究工作对保障旋风分离器的长期安全稳定运行具有重要理论指导意义。     

DIMENSIONLESS REPRESENTATION OF PARTICLE SEPARATION CHARACTERISTIC OF CYCLONES

In particular the collection efficiencies were measured as a function of flow rate, cyclone dimensions and particle size. For this purpose a fast, accurate and problem adapted measuring technique has been used, which enables the determination of grade efficiency curves by measuring the size distributions in the cyclone up- and downstream with optical particle counters. The extended experimental data from this parameter study were analysed by the methods of dimensional analysis and theory of models. An evaluation of all measuring results for two cyclone designs has been resulted in an empirical, nondimensional correlation of the collection characteristic, a dimensionless grade efficiency curve. Deviating from geometric similarity this correlation includes a variation of cyclone outlet diameter. Grade efficiencies of the cyclones are a definite function of the dimensionless numbers Stokes and Reynolds number and of the dimensionless cyclone outlet diameter. Analysis of own and published data has shown that this experimental correlation includes the influence of the temperature and that cyclone body diameter do not influence efficiency. The influence of cyclone height on flow behaviour and collection characteristic could be quantified as well. The range, in which prediction of collection efficiencies is possible, is marked in a state diagram Reynolds number versus dimensionless cyclone height.     

Effect of Operating Conditions on PV Performance of PVA Membranes: Experimental and Neural Network Modeling

Poly (vinyl alcohol) membranes were prepared by in-situ crosslinking of poly(vinyl alcohol) with glutaraldehyde as crosslinking agent and hydrochloric acid as catalyst and used for dehydration of IPA mixtures. Effects of feed composition, operating temperature, vacuum pressure, and Reynolds number on the permeation performance of the membranes were evaluated. Eighty-nine experimental data was applied to investigate ANN modeling. A multi layered feedforward neural network was applied to model the PV membranes. Two major training algorithms and optimum number of neurons and layers were investigated. As a result, Bayesian regularization successfully predicted experimental data. Different network structures were optimized, using multi object genetic optimization algorithm. The results concluded that the network with structure composing two hidden layers performs better than the other with one hidden layer, and also there is an excellent compatibility between the experimental data and the predicted values of optimum network structure (4:3:2:2). Furthermore, the optimum network was applied to predict extrapolation data and the results showed that this network can extrapolate data as well as interpolating.     

Artificial neural network-based modeling of pressure drop coefficient for cyclone separators

In order to build the complex relationships between cyclone pressure drop coefficient (PDC) and geometrical dimensions, representative artificial neural networks (ANNs), including back propagation neural network (BPNN), radial basic functions neural network (RBFNN) and generalized regression neural network (GRNN), are developed and employed to model PDC for cyclone separators. The optimal parameters for ANNs are configured by a dynamically optimized search technique with cross-validation. According to predicted accuracy of PDC, performance of configured ANN models is compared and evaluated. It is found that, all ANN models can successfully produce the approximate results for training sample. Further, the RBFNN provides the higher generalization performance than the BPNN and GRNN as well as the conventional PDC models, with the mean squared error of 5.84 × 10^?4 and CPU time of 120.15 s. The result also demonstrates that ANN can offer an alternative technique to model cyclone pressure drop.     

Development of a Multi-Stage Axial Flow Cyclone

A prototype multi-stage cyclone system consisting of an impaction inlet and five axial flow cyclone stages has been developed to classify simulants of Lunar and Martian dusts for various research and development needs of NASA's space exploration missions. Individual axial flow cyclone stages can be either independently operated with an inline connection to other particle devices or cascaded together for particle separation and collection. The impaction inlet and first three cyclone stages were designed to operate at the flowrate of 50 lpm under pressure close to ambient. The last two cyclone stages were designed to operate under low pressure conditions to separate particles with diameters less than 200 nm. Due to the limited vacuum capacity of the pump used, the flowrates of last two cyclone stages were restricted to 11.0 and 1.0 lpm when operating the assembled prototype. The impaction inlet and each cyclone stage of the prototype were experimentally calibrated, and the cutoff particle sizes were 11.3 μm, 0.97 μm, 550 nm, 255 nm, 109 nm, and 40 nm.

It was further found that in general the flow Reynolds (Re) and particle Stokes numbers (StK) were critical parameters to characterize the performance of the axial flow cyclone stages, and the relationship between Re and the dimensionless cutoff size (√StK was established. In addition, the collection efficiency curves are shifted to a smaller size range with a decrease of the cyclone pressure. However, using √StK as the abscissa and keeping the same Re, the particle collection curves at different pressures can be merged into one. This study also found that the upstream pressure should be used to calculate StK instead of the average of upstream and downstream pressures of the test cyclone stage.     

A CFD study of the effect of cone dimensions on sampling aerocyclones performance and hydrodynamics

This work presents a Computational Fluid Dynamics calculation to evaluate the effects of cone dimensions on the performance, hydrodynamics and centrifugal forces of sampling aerocyclones (gas cyclones). The problem of modeling highly swirling flow is overcome by means of an algebraic turbulence model. The axial and tangential velocities in a cyclone are successfully simulated. The refined mesh on the cyclone cone was also applied to ensure a better prediction on the effect of cone tip diameter to its performance, centrifugal forces and hydrodynamics. The pressure drop, grade efficiency and cut-off size of a cyclone of different cone dimensions was predicted very well with average deviation of about 2.9%, 5% and 2.1% respectively from experimental data presented in the literature. The findings suggest that the higher peak of tangential and axial velocity in a cyclone of a small cone lead to a higher collection efficiency and pressure drop. This helps to assess the benefit of enlarging or reducing the cone of a given cyclone. Results obtained from the computer modeling have demonstrated that CFD is suitable for modeling an effect of cyclone dimension on its performance.     

Artificial neural network-based prediction of hydrogen content of coal in power station boilers

Artificial neural networks (ANN) are powerful tools that can be used to model and investigate various highly complex and non-linear phenomena. This paper describes the development and training of a feed-forward back-propagation artificial neural network (BPNN), which is used to predict the hydrogen content in coal from proximate analysis. The ultimate objective is to enhance the performance of the combustion control system with the aid of regularly obtained knowledge of the elemental content of coal.In the present work, network modelling was performed using MATLAB with the Levenberg–Marquardt algorithm. Nine-hundred and three sets of data from a diverse range of coals have been used to develop the neural network architecture and topology. Trials were performed using one or two hidden layers with the number of neurons varied from 4 to 30. Validation data has been adopted to evaluate each trial and better model structure is determined to combat the over-fitting problem. As a result, it was found that a 4-12-1 or 4-8-4-1 network could give the most accurate prediction for this particular study. The regression analysis of the model tested gave a 0.937 correlation coefficient and the mean squared error of 0.0087. The average relative error is 5.46%. This has demonstrated that artificial neural networks have good potential for predicting elemental content of coal from frequently available proximate analysis data in power utilities.     

Process parameter optimization for ionic liquids as gas separation membranes based on a GWO–BPNN model

Gas separation membranes offer a cost-effective solution for capturing greenhouse gases, mitigating the global greenhouse effect. Ionic liquids (ILs) have emerged as one of the promising materials for greenhouse gas separation due to their strong affinity for CO₂. In this study, we propose a laboratory-scale method for preparing IL–PVDF blend membranes with high CO₂/N₂ selectivity. The separation performance of the membranes was evaluated using a custom gas permeation measurement system. The effects of casting solution composition, solidification method, and film-forming processes on separation performance were experimental investigated, and the obtained experimental data were used to train a back propagation neural network (BPNN) optimized by the Gray Wolf Optimizer (GWO) algorithm. This hybrid GWO–BPNN model was utilized to predict separation membrane efficiency, optimize the film-forming process, and identify the optimal range of process parameters. Notably, the GWO–BPNN model demonstrated a 2.76% higher prediction accuracy compared to a standalone BPNN. The results indicated that the GWO–BPNN algorithm has a great potential to accurately predict membrane separation efficiency and apply in optimal membrane process design (OMPD), and this method can significantly reduce the number of experimental trials required to achieve OMPD.     

Artificial neural network modeling for predicting melt‐blowing processing

An artificial neural network (ANN) model is established for predicting the fiber diameter of melt‐blown nonwoven fabrics from the processing parameters. An attempt is made to study the effect of the number of the hidden layers and the hidden layer neurons to minimize the prediction error. The artificial neural network with three hidden layers (5, 2, and 3 neurons in the first, second, and third hidden layer, respectively) yields the minimum prediction error, and thus, is determined as the preferred network. The square of correlation coefficient of measured and predicted fiber diameters shows the good performance of the model. Using the established ANN model, computer simulations of the effects of the processing parameter on the fiber diameter are carried out. The results show great prospects for this research in the field of computer‐assisted design of melt‐blowing technology. © 2006 Wiley Periodicals, Inc. J Appl Polym Sci 101: 4275–4280, 2006     

Artificial neural network modeling for predicting melt blowing processing

An artificial neural network model is established for predicting the fiber diameter of melt blown nonwoven fabrics from the processing parameters. An attempt is made to study the effect of the number of hidden layers and hidden layer neurons to minimize the prediction error. The artificial neural network with three hidden layers (5, 2, and 3 neurons in the first, second, and third hidden layer, respectively) yields the minimum prediction error and thus is determined as the preferred network. The square of the correlation coefficient of measured and predicted fiber diameters shows the good performance of the model. Using the established ANN model, computer simulations of the effects of the processing parameters on the fiber diameter are carried out. The results show great promise for this research in the field of computer assisted design of melt blowing technology. © 2005 Wiley Periodicals, Inc. J Appl Polym Sci 99: 424–429, 2006     

Fast property prediction in an industrial rubber mixing process with local ELM model

Online property prediction in industrial rubber mixing processes is not an easy task. An efficient data‐driven prediction model is developed in this work. The regularized extreme learning machine (RELM) is utilized as the fundamental soft sensor model. To better capture distinguished characteristics in multiple recipes and operating modes, a just‐in‐time RELM modeling method is developed. The number of hidden neurons and the value of regularization parameter of the just‐in‐time RELM model can be efficiently selected using a fast leave‐one‐out strategy. Consequently, without the time‐consuming laboratory analysis process, the Mooney viscosity can be online predicted once a mixing batch has been discharged. The industrial Mooney viscosity prediction results show its better prediction performance in comparison with traditional approaches. © 2017 Wiley Periodicals, Inc. J. Appl. Polym. Sci. 2017 , 134, 45391.