基于Bayesian-SVR的城市用水量软测量建模  

Soft Sensor Modeling of Urban Water Consumption Based on Bayesian-SVR

作　　者：王鲜芳[1,2] 杜志勇[3] 潘丰[2] 

机构地区：[1]河南师范大学计算机与信息技术学院,河南新乡453007 [2]江南大学通信与控制工程学院,江苏无锡214122 [3]河南机电高等专科学校计算机科学与技术系,河南新乡453002

出　　处：《江南大学学报（自然科学版）》2010年第4期409-413,共5页Joural of Jiangnan University (Natural Science Edition)　

基　　金：国家863计划重点项目(2006AA020301);河南省教育厅自然科学研究计划项目(2009A480001)

摘　　要：针对传统交叉验证方法确定SVR模型参数存在耗时过长的问题,利用贝叶斯置信框架推断SVR的模型参数。通过第一级推断确定支持向量机的权矢量ω和偏置项b,通过第二级推断确定模型确定不敏感系数ε、惩罚因子C,通过第三级推断得到核参数。然后利用该方法对某城市用水系统中用水量进行建模,预测结果表明该方法不仅能加快建模速度,而且提高了预测精度。Aiming at the problem that determining parameters in traditional Support Vector Regression（SVR） modeling by cross-validation costs a long time,the Bayesian evidence framework is proposed to infer the parameters of SVR model.The weights ω and bias term b of the SVR are inferred on the first level of inference.The model insensitive factor ε,penalty factor C,are inferred on the second level of inference.The model comparison is performed on the third level of inference to automatically tune the parameter of kernel function.Besides,an hourly water demand forecast model for a city water supply is built based on this method.The forecast result shows that this prediction model not only has higher modeling speed,but also has better prediction performance.

关 键 词：交叉验证 支持向量回归 贝叶斯置信框架 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程]