基于人工神经网络的管壳式换热器静态预测与动态特性的分析研究/The Static Prediction and Dynamic Behavior of Shel

2018-06-30 09:48:07

heat 换热器 ANN tube exchangers



人工神经网络(Artificial Neural Networks 简称ANNs)广泛应用于模式识别、工业过程控制等领域。基于ANN在处理非线性问题方面的优势,ANN在热科学领域,如换热器的静态预测与动态特性及其控制方面的研究也越来越受到研究者关注,而换热器静态及动态特性研究对其设计优化具有十分重要的意义。通常对于工程中复杂流程及结构的换热器研究一般采用实验方法,理论上较难推导,而ANN方法能够对理论上难以推导的数学模型从实验角度来获取,应用于各种换热器静态及动态性能预测简单可行,能够在短时间内获得高精度要求,是一种有效的研究途径。
本文以管壳式换热器为研究背景,应用ANN方法分别研究了换热器静态以及动态特性。首先采用多层前向BP网络结构,在前人有限实验数据的基础上,实现对螺旋折流板换热器静态预测,应用BP标准算法、LM算法和贝叶斯三种不同算法进行分析比较,得出最优算法,同时考察了训练和预测样本数据的比例以及不同网络结构对预测性能的影响。在完成静态预测的基础上,对连续螺旋折流板换热器及弓型折流板换热器在以水、油为换热工质系统的动态特性进行实验研究,并将所得到的实验结果作为样本数据,应用ANN技术实现动态特性预测,与实验结果进行对比验证。最后应用数学分析法重点讨论了管壁热容及双侧扰动对管壳式换热器动态特性的影响。
静态研究结果表明:ANN预测与实验数据最大相对偏差为3%。在本文研究背景下,三种算法中贝叶斯算法效果最好。对于如何选取训练和预测样本数据的比例应该根据实际问题需要,在总样本一定的情况下,随着学习样本数据增多,预测精度变高,但泛化能力变弱。比较得出4-5-2-1 为最优ANN结构,适用于弓型及连续螺旋折流板两种换热器模型,预测误差均在11%以内。动态实验结果表明:在进口发生流量扰动时,出口温度响应随冷流体流量升高扰动而降低,随热流体流量升高扰动而升高。对于水-油换热器系统其换热工质比热容较大,故其热惯性也比较大,进口扰动所造成的温度变化比较小而且缓慢,响应时间比较长,且油侧温度稳定时间较水侧长。采用ANN方法进行动态预测,预测结果与实验数据对比,误差均在±1%以内,训练过程中CPU耗时均在10s左右,可见在短时间内可以获得高精度要求,足以满足一般的工程需要。数学分析法动态研究结果表明:管壁热容不变的情况下,同侧流体出口温度响应比隔侧流体出口温度响应幅度大,两侧流体出口温度响应受管壁热容的增大幅度变小,在双侧扰动下,响应幅度较单侧扰动大。可见,传统的管壳式换热器其简化模型比较简单,动态特性采用数学分析法能够得到计算结果,但这种方法是建立在一系列简化假设基础之上的,对于一些结构比较复杂的新型管壳式换热器,应用ANN方法能够在短时间内获得高精度结果,该方法对复杂的换热器提供一种有效的工程研究途径,研究结果为换热器的设计及控制提供一定的参考依据。

Artificial Neural Networks (ANNs) are widely used in the field of pattern recognition, industry control and so on. Because of the superiority of ANNs in solving non-linear problem, researchers pay much attention to the prediction and control of heat exchangers in thermal science by ANNs. The static prediction and dynamic behaviors are significant for design and optimization of heat exchangers. Usually, the experimental technique is used for the research of heat exchangers with complex structure. The mathematical model can be obtained by using ANNs from the experiments, rather than the theoretical analysis. Previous works show that ANNs are effective and powerful tools and can be feasibly used in the performance prediction of heat exchangers.
This dissertation predicts the performance of shell-and-tube heat exchangers with helical baffles by using Back Propagation (BP) standard algorithm, the Levenberg-Marquardt (LM) algorithm and the Bayesi algorithm and performance of shell-and-tube heat exchangers with continuous baffles and shell-and-tube heat exchangers with segmental baffles under different network architecture. After the static prediction of shell-and-tube heat exchanger, this dissertation analyzes the dynamic behaviors of shell-and-tube heat exchanger by ANN method and mathematics analytic method respectively. In order to obtain the sample data, the experimental system is built to investigate the dynamic behaviors of shell-and-tube heat exchangers with continuous baffles and with segmental baffles. With the experimental data, the dissertation analyzes the dynamic behavior predicted by ANN approach. The effects of the single-side and double-side temperature disturbances and the tube-wall heat capacity on dynamic behaviors of heat exchangers are investigated in detail by mathematics analytic method.
The results of static prediction show that the maximum relative error of predict result and the empirical data is 3%. The Bayesi algorithm is the best one in three algorithms. Selection of the proportion of the sample data for training and testing networks should be according to the requirement of the actual problem. Under certain total sample, higher the number of sample data, higher the prediction precision, however, weaker the ability of exudes. 4 - 5 - 2 - 1 ANN configuration is optional network to predict shell-and-tube heat exchangers with continuous baffles and with segmental baffles, and the predicted errors are less than 11%. The dynamic experimental results show that the response time is long for liquid-liquid heat exchangers. Under the different inlet flow rate disturbances, the temperature varies slowly and slightly changes in value, and the stable time of oil side is longer than that of the water side. Comparing the ANN prediction result with the empirical data, the error is in ±1%, CPU consumes about 10s, it shows that ANN can obtain high accuracy results in short time under engineering requirement. The dynamic behaviors by mathematics analytic method indicate that, when the tube-wall heat capacity is invariable, the response of the outlet temperature in the same side is higher than that of the different side, and the response of the outlet temperature changes lowly as the tube-wall heat capacity increases. Under the double-side disturbance, the temperature response becomes higher than the single-side temperature disturbance. Although the dynamic of shell-and-tube heat exchanger can be investigated based on the simplified model by mathematics analytic method, the mathematics analytic method is valid under a series of assumptions. For some newly heat exchanger with complex structure, ANN method can be feasibly provide a effective way in engineering research and provide the certain reference for design and control of the heat exchangers.