中国电机工程学报
中國電機工程學報
중국전궤공정학보
Proceedings of the CSEE
2015年
23期
6083-6088
,共6页
刘吉臻%秦天牧%杨婷婷%吕游
劉吉臻%秦天牧%楊婷婷%呂遊
류길진%진천목%양정정%려유
选择性催化还原(SCR)脱硝%偏最小二乘%多尺度核%自适应%数据建模
選擇性催化還原(SCR)脫硝%偏最小二乘%多呎度覈%自適應%數據建模
선택성최화환원(SCR)탈초%편최소이승%다척도핵%자괄응%수거건모
selective catalytic reduction (SCR) denitration%partial least squares (PLS)%multi-scale kernel%self-adaptive%data modeling
针对选择性催化还原(selective catalytic reduction,SCR)脱硝系统中输入变量的多尺度特性及时变特性问题,将核偏最小二乘与多核学习相结合,同时引入自适应模型更新策略,提出了自适应多尺度核偏最小二乘(self-adaptive multi-scale KPLS,SMKPLS)方法。通过优化算法确定每个自变量对应的核函数宽度,然后利用多尺度核偏最小二乘方法建立非线性模型,采用自适应模型更新方法对模型进行更新。将该方法应用于SCR脱硝系统建模,并与其他建模方法进行对比,结果表明,SMKPLS 预测精度明显高于其他模型,计算时间远小于其他模型,具有更好的泛化能力及鲁棒性。
針對選擇性催化還原(selective catalytic reduction,SCR)脫硝繫統中輸入變量的多呎度特性及時變特性問題,將覈偏最小二乘與多覈學習相結閤,同時引入自適應模型更新策略,提齣瞭自適應多呎度覈偏最小二乘(self-adaptive multi-scale KPLS,SMKPLS)方法。通過優化算法確定每箇自變量對應的覈函數寬度,然後利用多呎度覈偏最小二乘方法建立非線性模型,採用自適應模型更新方法對模型進行更新。將該方法應用于SCR脫硝繫統建模,併與其他建模方法進行對比,結果錶明,SMKPLS 預測精度明顯高于其他模型,計算時間遠小于其他模型,具有更好的汎化能力及魯棒性。
침대선택성최화환원(selective catalytic reduction,SCR)탈초계통중수입변량적다척도특성급시변특성문제,장핵편최소이승여다핵학습상결합,동시인입자괄응모형경신책략,제출료자괄응다척도핵편최소이승(self-adaptive multi-scale KPLS,SMKPLS)방법。통과우화산법학정매개자변량대응적핵함수관도,연후이용다척도핵편최소이승방법건립비선성모형,채용자괄응모형경신방법대모형진행경신。장해방법응용우SCR탈초계통건모,병여기타건모방법진행대비,결과표명,SMKPLS 예측정도명현고우기타모형,계산시간원소우기타모형,구유경호적범화능력급로봉성。
Based on the multi-scale characteristics and time-varying characteristics of selective catalytic reduction system, combining kernel partial least squares and multiple kernel learning, introducing the self-adaptive model updating method at the same time, self-adaptive multi-scale kernel partial least squares regression(SMKPLS) was proposed. Optimization algorithm was used to determine the kernel function width of each variable. Then multi-scale kernel partial least squares method was used to establish the nonlinear model. The model was updated with adaptive model updating method. By applying the method to SCR system modeling and comparing with other modeling methods, results show that the prediction accuracy of SMKPLS is significantly higher, calculation time of SMKPLS is far less than that of others, generalization ability and robustness of SMKPLS are both better.