摘要
为了提高水泥生产过程的能耗建模和预测的精度,提出了一种基于神经网络与马尔科夫修正的水泥生产集成能耗预测模型:在数据预处理阶段,为了减小处理数据的规模,采用平均影响值法进行数据降维,筛选敏感变量,从12个变量中选出对能耗输出影响较大的6个,构建一个6输入单输出的神经网络,使能耗建模阶段选用的神经网络模型结构更为简单,可以有效减少训练神经网络所需的时间。在能耗建模阶段,为了建立性能更佳的能耗模型,在以神经网络作为能耗建模元学习器的基础上采用集成学习思想,组合数个元学习器成一个性能更佳的强学习器,即将多个神经网络的预测输出值求平均作为集成模型的预测结果,采用水泥烧成系统的依赖变量和对应的炉窑能耗值作为试验数据进行模型的训练、验证和预测,结果表明,集成模型预测结果的决定系数R 2值较单个神经网络提高了0.019,预测值与真实值的相对残差的均值较单个神经网络也减少了0.027,模型性能有所提高。在能耗预测阶段,为了进一步提高模型的预测精度,引入马尔科夫残差修正法,即依据历史预测能耗值与实际能耗值的残差修正当前预测值,提升集成能耗模型的预测精度。结果表明,经马尔科夫修正法修正的预测值相对残差从-0.6%降至-0.25%,能耗预测值更加接近实际值,预测精度提升,可更好地挖掘水泥炉窑烧成系统电能耗变化与依赖变量的规律,实现能耗精确预测,为水泥生产过程的能耗监管提供了更精确的参考依据。根据水泥生产能耗建模3个阶段的描述,提出一种基于神经网络与马尔科夫修正的水泥生产集成能耗预测模型,在水泥生产能耗预测上有更佳的预测效果和更高的预测精度。
In order to improve the accuracy of energy consumption modeling and prediction in the cement production process,an integrated energy consumption prediction model for cement production based on neural network and Markov correction was proposed in this paper.In the data preprocessing stage,in order to reduce the scale of processing data,the average influence value method was used to reduce the data dimension,and the sensitive variables were filtered,six of the 12 variables that have a greater impact on the energy consumption output were selected to construct a 6-input single-output neural network,which made the energy consumption modeling stage selection.The structure of the neural network model is simpler,and could effectively reduce the time required to train the neural network.In the energy consumption modeling stage,in order to establish a better performance energy consumption model,the integrated learning idea was adopted on the basis of the neural network as the energy consumption modeling meta-learner,and several meta-learners were combined into a stronger performance.The average of the predicted output values of multiple neural networks was used as the prediction result of the integrated model.The dependent variables of the cement firing system and the corresponding furnace energy consumption value were used as experimental data for model training,verification and prediction.The experimental results show that the determination coefficient of the prediction results of the integrated model is improved by 0.019 compared with a single neural network.The mean value of the relative residual size between the predicted value and the true value is also reduced by 0.027 compared with the single neural network.The performance of the model is improved.In the energy consumption prediction stage,in order to further improve the prediction accuracy of the model,the Markov residual correction method is introduced,that is,the current predicted value is corrected based on the residual of the historical predicted energy consum
作者
黄堃
杨文
丁孝华
HUANG Kun;YANG Wen;DING Xiaohua(NARI Technology Co.,Ltd.,Nanjing 211106,China)
出处
《洁净煤技术》
CAS
2020年第5期103-110,共8页
Clean Coal Technology
基金
国家重点研发计划资助项目(2016YFB0601501)。
关键词
神经网络
能耗建模
水泥生产
集成算法
马尔科夫修正
neural network
energy consumption modeling
cement production
integrated algorithm
Markov correction
