基于改进回声状态网络的高炉煤气产耗预测  

Study on Prediction Method for Generation and Consumption of Blast Furnace Gas Based on Improved Echo State Network

作　　者：刘颖[1] 时飞飞[1] 赵珺[1] 王伟[1] 丛力群[2] 冯为民[2] 

机构地区：[1]大连理工大学信息与控制研究中心,大连116024 [2]上海宝信软件股份有限公司,上海201203

出　　处：《系统仿真学报》2011年第10期2184-2189,共6页Journal of System Simulation

基　　金：国家自然科学基金重点项目(61034003)

摘　　要：以钢铁企业高炉煤气系统为背景,针对其产生量和消耗量的预测问题,提出一种改进回声状态网络时间序列方法进行系统仿真预测,并根据最小均方差准则,以最小化网络训练误差为目标,采用随机梯度下降法对网络参数进行优化。该方法对于不同预测对象,可计算出合适的网络连接权值、储备池谱半径等参数,避免了传统回声状态网络方法中单凭经验选择网络参数的现状,提高了预测精度。采用该方法对高炉煤气系统现场实际产耗数据进行了仿真预测,仿真结果表明所提出方法的有效性。To the background of blast furnace gas （BFG） system in steel industry, an improved echo state network time series forecasting method with parameters was proposed to simulate gas generation and consumption. Based on least mean square error criterion, the network parameters were optimized with stochastic gradient descent method to achieve a minimal training error. This method can be used to obtain suitable network parameters of different objects such as connection weights, and related reservoir spectral radius. With this new method, the traditional echo state network method that selects network parameters only depending on human experience will be improved, and the prediction accuracy will be also promoted. The simulation results of practical gas generation and consumption have shown the validity of the proposed approach.

关 键 词：高炉煤气系统仿真预测 回声状态网络 梯度下降法 参数优化 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程]