基于改进ESN的高炉煤气系统预测方法的研究
|
摘要
钢铁企业中能源介质的传输与转换的过程复杂,确保连续、安全和经济的能源供应是企业能源管理的重要任务。从钢铁企业能源的使用情况来看,副产煤气是能源优化调度的重点。由于高炉煤气(BFG)热值最低,生产过程复杂,产出波动大,一般情况下企业在对煤气调度和平衡调整不善时首先选择将BFG作放散处理,这势必造成大量有害气体对环境造成污染。因此对BFG产生量和消耗量进行科学预测,充分回收和利用BFG,可为制定合理的煤气使用计划提供依据,提高钢铁企业节能降耗的水平。
本文以上海宝钢股份公司为研究背景,针对BFG系统的预测问题,将其系统中的产消用户分为非调节用户和调节用户,对不同类型用户采用不同的方法进行预测。对于非调节用户,提出一种带参数的改进回声状态网络(ESN)时间序列方法进行预测,同时根据最小均方差准则,以训练误差最小为目标,利用随机梯度下降法优化网络参数,并与其它几种常见神经网络的预测效果进行了比较,说明了该方法的优势。对于调节用户,则采用平均值法进行预测。对于BFG的受入量,由于其含噪水平较高,直接采用改进ESN方法进行预测,其效果不理想,为此提出一种基于经验模态分解(EMD)的除噪算法,在预测前首先对数据进行除噪。此外,分析了影响高炉煤气系统柜位升降的因素,采用ESN网络建立了高炉煤气系统柜位预测因素模型。
在高炉煤气系统预测模型的基础上,本文研究开发了高炉煤气系统预测软件。该软件分为服务端和客户端,服务端每隔一定时间对高炉煤气系统(包括各个用户和柜位)进行预测,客户端则将实时数据和预测数据通过用户界面进行显示。仿真实验和高炉煤气系统软件运行均取得了较好的效果,表明了本文提出算法的有效性。
The transportation and conversion process of energy in steel industry is relatively complicated. It is an important task for enterprise energy management to guarantee a continuous, safe and economical energy supply. At present, the byproduct gas is the key part of energy optimal scheduling problem. Blast furnace gas (BFG) is with the characteristics of least caloricity, complex generation process and output with big fluctuation, so it will usually be diffused firstly when poor energy scheduling and imbalance occurs, which will create much more harmful gas and degrade the environment. Therefore, the study on reliable prediction of the BFG generation and consumption and its recovery and reuse fully can provide scientific guidance for efficient utilization of the byproduct energy and improve the level to save energy in steel enterprise.
In this paper, on basis of the background of Shanghai Baosteel, the generation and consumption users of the BFG system are divided into two types for the prediction problem, which are adjustable users and non-adjustable users, respectively. For non-adjustable users, an improved echo state network (ESN) prediction method with parameters is proposed firstly. Based on Least Mean Square (LMS) error criterion, the network parameters are optimized by stochastic gradient descent in order to achieve a minimized training error. Compared to other methods, such method brings about a better prediction performance. Secondly, an average method is used to predict adjustable users. Since the data of the BFG generation includes a series of noises in practice, the performance is unsatisfactory if using the improved ESN method directly. For resolving this problem, an approach based on empirical mode decomposition (EMD) is employed to implement the data de-noise before prediction. Based on analyzing the influence factors of gas holder level, prediction factor model of the BFG gas holder is built by the ESN method.
An application software for the BFG system prediction is developed on the basis of the prediction method. This software consists of server and client terminal. The server predicts the BFG system which includes all users and gas holder at regular intervals and the client terminal presents the real-time data and predict data through the user interfaces. The prediction software has been applied in the Energy Center of Baosteel and the running results demonstrate the validity of the proposed method.
本文以上海宝钢股份公司为研究背景,针对BFG系统的预测问题,将其系统中的产消用户分为非调节用户和调节用户,对不同类型用户采用不同的方法进行预测。对于非调节用户,提出一种带参数的改进回声状态网络(ESN)时间序列方法进行预测,同时根据最小均方差准则,以训练误差最小为目标,利用随机梯度下降法优化网络参数,并与其它几种常见神经网络的预测效果进行了比较,说明了该方法的优势。对于调节用户,则采用平均值法进行预测。对于BFG的受入量,由于其含噪水平较高,直接采用改进ESN方法进行预测,其效果不理想,为此提出一种基于经验模态分解(EMD)的除噪算法,在预测前首先对数据进行除噪。此外,分析了影响高炉煤气系统柜位升降的因素,采用ESN网络建立了高炉煤气系统柜位预测因素模型。
在高炉煤气系统预测模型的基础上,本文研究开发了高炉煤气系统预测软件。该软件分为服务端和客户端,服务端每隔一定时间对高炉煤气系统(包括各个用户和柜位)进行预测,客户端则将实时数据和预测数据通过用户界面进行显示。仿真实验和高炉煤气系统软件运行均取得了较好的效果,表明了本文提出算法的有效性。
The transportation and conversion process of energy in steel industry is relatively complicated. It is an important task for enterprise energy management to guarantee a continuous, safe and economical energy supply. At present, the byproduct gas is the key part of energy optimal scheduling problem. Blast furnace gas (BFG) is with the characteristics of least caloricity, complex generation process and output with big fluctuation, so it will usually be diffused firstly when poor energy scheduling and imbalance occurs, which will create much more harmful gas and degrade the environment. Therefore, the study on reliable prediction of the BFG generation and consumption and its recovery and reuse fully can provide scientific guidance for efficient utilization of the byproduct energy and improve the level to save energy in steel enterprise.
In this paper, on basis of the background of Shanghai Baosteel, the generation and consumption users of the BFG system are divided into two types for the prediction problem, which are adjustable users and non-adjustable users, respectively. For non-adjustable users, an improved echo state network (ESN) prediction method with parameters is proposed firstly. Based on Least Mean Square (LMS) error criterion, the network parameters are optimized by stochastic gradient descent in order to achieve a minimized training error. Compared to other methods, such method brings about a better prediction performance. Secondly, an average method is used to predict adjustable users. Since the data of the BFG generation includes a series of noises in practice, the performance is unsatisfactory if using the improved ESN method directly. For resolving this problem, an approach based on empirical mode decomposition (EMD) is employed to implement the data de-noise before prediction. Based on analyzing the influence factors of gas holder level, prediction factor model of the BFG gas holder is built by the ESN method.
An application software for the BFG system prediction is developed on the basis of the prediction method. This software consists of server and client terminal. The server predicts the BFG system which includes all users and gas holder at regular intervals and the client terminal presents the real-time data and predict data through the user interfaces. The prediction software has been applied in the Energy Center of Baosteel and the running results demonstrate the validity of the proposed method.
引文
[1]中国钢铁工业协会.“十一五”中国钢铁企业信息化发展建议[J].冶金自动化,2007,31(2):6-10.
[2]张应良.能源介质的综合调整法[J].能源技术,2000,6(2):115-117.
[3]Mohsen M S,Akash B A.Energy analysis of the steel making industry[J].International Journal of Energy Research,1998,22(12):1049-1054.
[4]Naoki H,Yoshitsugu W,Takeyoshi K,et al.Minimizing energy consumption in Industries by cascade use of waste energy[J].IEEE Transactions on Energy Conversion,1999,14(3):795-801.
[5]Valsalam S R,Muralidharan V,Krishnan N.Implementation of energy management system for an integrated steel plant[C].Proceedings of Energy Management and Power Delivery.1998,2:661-666.
[6]唐万梅.几个预测方法及模型研究[D].内蒙古:内蒙古大学,2006.
[7]John F.Applied Regression Analysis,Linear Models,and Related Methods[M].SAGE,1997.
[8]George E P B,Gwilym J.Time series analysis,forecasting and control[M].Holden-Day,Incorprated,1990.
[9]刘思峰,郭天榜,党耀国.灰色系统理论及其应用[M].北京:北京科学出版社,1999.
[10]Guo-Dong L,Daisuke Y,Masatake N,et al.Research on grey model by grey interval analysis[J].Journal of Grey System,2006,9(2):103-113.
[11]Christopher M B.Neural networks for pattern recognition[M].Oxford University Press,1995.
[12]Clemen R T.Combining forecasts:A review and annotated bibliography[J].International Journal of Forecasting,1989,5:559-583.
[13]何书元.应用时间序列分析[M].北京:北京大学出版社,2003
[14]Muller K R,Smola A J,Ratsch J,et al.Predicting time series with support vector machines[C].Proc.7th Int.Conf.Artificial Neural Networks,1997,1327:999-1004.
[15]Vapnik V N.The nature of statistical learning theory[M].Springer-Verlag,New York,1995.
[16]Feldkamp L A,Prokhorov D,Eagen C F,et al.Enhanced multi-stream Kalman filter training for recurrent networks[R].Nonlinear modeling:advanced black-box techniques,1998,29-53.
[17]向小东.基于神经网络与混沌理论的非线性时间序列预测研究[D].成都:西南交通大学,2001.
[18]张玉瑞,陈建波.基于RBF神经网络的时间序列预测[J].计算机工程与应用,2005(11):74-76.
[19]Cowper M R,Mulgrew B,Unsworth C P.Nonlinear prediction of chaotic signals using a normalized radial basis function network[C].Signal Processing,2002,82(5):775-789.
[20]Vesanto J.Using the SOM and local models in time-series prediction[C].Proc.Workshop Self-Organizing Maps(WSOM),1997:209-214.
[21]Principe J C,Kuo J M.Dynamic modeling of chaotic time series with neural networks[C].Advances in Neural Information Processing Systems,1995,7:311-318.
[22]Jaeger H.Tutorial on training recurrent neural networks,covering BPTT,RTRL,EKF and "echo state network" approach[R].German National Research Center for Information Technology,2002.
[23]Jaeger H,Haas H.Harnessing nonlinearity:Predicting chaotic systems and saving energy in wireless communication[J].Science,2004,304(5667):78-80.
[24]Kountouriotis P A,Obradovic D,Mandic D P,et al.Multi-step forecasting using echo state networks[J].Computer as a tool,2005,2:1574-1577.J
[25]Peter J B,Richard A D.时间序列与预测[M].北京:人民邮电出版社,2009.
[26]杨行俊,郑君里.人工神经网络[M].北京:高等教育出版社,1992.
[27]Jaeger H.Adaptive nonlinear system identification with echo state networks[C].Advances in Neural Information Processing Systems.2003,15:593-600
[28]王瑟.基于改进的回声状态神经网络的非线性预测[D].南京:南京工业大学,2006.
[29]Salihoglu U.Chaos in small recurrent neural networks:theoretical and practical studies[M].Anne academic 2003-2004 of University Libre de Bruxelles.
[30]刘颖,时飞飞,赵珺等.基于改进冋声状态网络的高炉煤气产耗预测仿真研究[J].系统仿真学报(已录用,待发表).
[31]Cowper M R,Mulgrew B,Unsworth C P.Nonlinear prediction of chaotic signals using a normalized radial basis function network[C].Signal Processing,2002,82(5):775-789.
[32]Muller K R,Smola A J,Ratsch J,etal.Predicting time series with support vector machines[C].Proc.7th Int.Conf.Artificial Neural Networks,1997,1327:999-1004.
[33]刘红星,左洪福,姜澄宇.往复机械特征频段信号的解调分解及应用[J].振动工程学报,2000,13(2):283-289.
[34]孙晖.经验模态分解理论与应用研究[D].浙江:浙江大学,2005.
[2]张应良.能源介质的综合调整法[J].能源技术,2000,6(2):115-117.
[3]Mohsen M S,Akash B A.Energy analysis of the steel making industry[J].International Journal of Energy Research,1998,22(12):1049-1054.
[4]Naoki H,Yoshitsugu W,Takeyoshi K,et al.Minimizing energy consumption in Industries by cascade use of waste energy[J].IEEE Transactions on Energy Conversion,1999,14(3):795-801.
[5]Valsalam S R,Muralidharan V,Krishnan N.Implementation of energy management system for an integrated steel plant[C].Proceedings of Energy Management and Power Delivery.1998,2:661-666.
[6]唐万梅.几个预测方法及模型研究[D].内蒙古:内蒙古大学,2006.
[7]John F.Applied Regression Analysis,Linear Models,and Related Methods[M].SAGE,1997.
[8]George E P B,Gwilym J.Time series analysis,forecasting and control[M].Holden-Day,Incorprated,1990.
[9]刘思峰,郭天榜,党耀国.灰色系统理论及其应用[M].北京:北京科学出版社,1999.
[10]Guo-Dong L,Daisuke Y,Masatake N,et al.Research on grey model by grey interval analysis[J].Journal of Grey System,2006,9(2):103-113.
[11]Christopher M B.Neural networks for pattern recognition[M].Oxford University Press,1995.
[12]Clemen R T.Combining forecasts:A review and annotated bibliography[J].International Journal of Forecasting,1989,5:559-583.
[13]何书元.应用时间序列分析[M].北京:北京大学出版社,2003
[14]Muller K R,Smola A J,Ratsch J,et al.Predicting time series with support vector machines[C].Proc.7th Int.Conf.Artificial Neural Networks,1997,1327:999-1004.
[15]Vapnik V N.The nature of statistical learning theory[M].Springer-Verlag,New York,1995.
[16]Feldkamp L A,Prokhorov D,Eagen C F,et al.Enhanced multi-stream Kalman filter training for recurrent networks[R].Nonlinear modeling:advanced black-box techniques,1998,29-53.
[17]向小东.基于神经网络与混沌理论的非线性时间序列预测研究[D].成都:西南交通大学,2001.
[18]张玉瑞,陈建波.基于RBF神经网络的时间序列预测[J].计算机工程与应用,2005(11):74-76.
[19]Cowper M R,Mulgrew B,Unsworth C P.Nonlinear prediction of chaotic signals using a normalized radial basis function network[C].Signal Processing,2002,82(5):775-789.
[20]Vesanto J.Using the SOM and local models in time-series prediction[C].Proc.Workshop Self-Organizing Maps(WSOM),1997:209-214.
[21]Principe J C,Kuo J M.Dynamic modeling of chaotic time series with neural networks[C].Advances in Neural Information Processing Systems,1995,7:311-318.
[22]Jaeger H.Tutorial on training recurrent neural networks,covering BPTT,RTRL,EKF and "echo state network" approach[R].German National Research Center for Information Technology,2002.
[23]Jaeger H,Haas H.Harnessing nonlinearity:Predicting chaotic systems and saving energy in wireless communication[J].Science,2004,304(5667):78-80.
[24]Kountouriotis P A,Obradovic D,Mandic D P,et al.Multi-step forecasting using echo state networks[J].Computer as a tool,2005,2:1574-1577.J
[25]Peter J B,Richard A D.时间序列与预测[M].北京:人民邮电出版社,2009.
[26]杨行俊,郑君里.人工神经网络[M].北京:高等教育出版社,1992.
[27]Jaeger H.Adaptive nonlinear system identification with echo state networks[C].Advances in Neural Information Processing Systems.2003,15:593-600
[28]王瑟.基于改进的回声状态神经网络的非线性预测[D].南京:南京工业大学,2006.
[29]Salihoglu U.Chaos in small recurrent neural networks:theoretical and practical studies[M].Anne academic 2003-2004 of University Libre de Bruxelles.
[30]刘颖,时飞飞,赵珺等.基于改进冋声状态网络的高炉煤气产耗预测仿真研究[J].系统仿真学报(已录用,待发表).
[31]Cowper M R,Mulgrew B,Unsworth C P.Nonlinear prediction of chaotic signals using a normalized radial basis function network[C].Signal Processing,2002,82(5):775-789.
[32]Muller K R,Smola A J,Ratsch J,etal.Predicting time series with support vector machines[C].Proc.7th Int.Conf.Artificial Neural Networks,1997,1327:999-1004.
[33]刘红星,左洪福,姜澄宇.往复机械特征频段信号的解调分解及应用[J].振动工程学报,2000,13(2):283-289.
[34]孙晖.经验模态分解理论与应用研究[D].浙江:浙江大学,2005.