基于遗传算法优化的BP神经网络气化用煤灰流动温度预测模型  

Prediction model of fusion temperature of coal ash for gasification based on GA-BP neural network

作　　者：邱钱粮 白向飞 QIU Qianliang;BAI Xiangfei(CCTEG Chinese Institute of Coal Science,100013 Beijing,China;CCTEG China Coal Research Institute,100013 Beijing,China;National Energy Technology and Equipment Laboratory of Coal Utilization and Emission Control,100013 Beijing,China)

机构地区：[1]煤炭科学研究总院,北京100013 [2]煤炭科学技术研究院有限公司,北京100013 [3]国家能源煤炭高效利用与节能减排技术装备重点实验室,北京100013

出　　处：《煤炭转化》2023年第2期109-118,共10页Coal Conversion

基　　金：国家自然科学基金项目(21875095和42030807).

摘　　要：采用遗传算法优化的BP神经网络建立煤灰流动温度预测模型,模型以灰成分及酸碱质量比、硅铝质量比等组合参数作为输入变量,以煤灰流动温度作为输出量,对126组来自中国北部地区的煤灰样数据进行训练与测试,并建立常规BP神经网络模型,研究了各输入变量对网络模型预测精度的影响并对比与常规BP神经网络模型的预测能力。结果表明:不同输入层变量的GA-BP神经网络模型对训练集和测试集样本数据都具有较好的学习和泛化能力,所有预测结果相对平均预测误差均不超过4%。酸碱质量比和硅铝质量比参数作为神经网络输入层的添加,虽略微提高模型对训练样本的拟合程度,但也导致验证时过拟合现象的发生,模型对新样本的拟合优度下降。采用SiO_(2),Al_(2)O_(3),Fe_(2)O_(3),CaO,MgO和Na_(2)O+K_(2)O的质量分数6个参数作为输入变量的GA-BP模型最为适合,其对测试集数据的预测相对平均误差为3.45%,低于常规BP神经网络模型3.79%的误差。Genetic algorithm was introduced to optimize the BP neural network to establish a coal ash fusion temperature prediction model.Combined parameters such as ash composition,acid-base mass ratio and silicon-aluminum mass ratio are used as input variables,and the coal ash flow temperature is set as output.Then,the model trains and tests 126 groups of data of coal ash samples from northern China,and a conventional BP neural network model is established.The influence of each input variable on the prediction accuracy of the network model was studied and its prediction ability was compared with that of the conventional BP neural network model.The results show that the GA-BP neural network model with different input layers has good learning and generalization ability for training and testing sample data,and the relative average prediction error of all prediction results does not exceed 4%.In addition,the use of acid-base mass ratio and silicon-aluminum mass ratio as input layer of neural network slightly improves the fitting accuracy of the model to the training samples,but also results in the over-fitting during verification,and the goodness of fit of the model to the new samples decreases.The GA-BP model using six parameters of the mass fraction of SiO_(2),Al_(2)O_(3),Fe_(2)O_(3),CaO,MgO and Na_(2)O+K_(2)O as input variables is the most suitable.The relative average error of GA-BP model is 3.45%,which is lower than the 3.79%error of a conventional BP neural network model.

关 键 词：煤炭气化 煤灰流动温度 预测模型 遗传算法 BP神经网络 

分 类 号：TQ534[化学工程—煤化学工程]