计算机与应用化学
計算機與應用化學
계산궤여응용화학
COMPUTERS AND APPLIED CHEMISTRY
2013年
10期
1113-1116
,共4页
硅含量预测%模糊神经网络%RBF神经网络%集成预测模型
硅含量預測%模糊神經網絡%RBF神經網絡%集成預測模型
규함량예측%모호신경망락%RBF신경망락%집성예측모형
prediction of silicon content%fuzzy neural network%RBF neural network%integrated prediction model
在炼铁生产过程中，有效的降低铁水硅含量并且使其保持在合理的低水平，有利于提高生铁的质量和产量。但是在实际生产过程中，生铁的硅含量只有在生铁样本送到实验室经过化验后方可得知，检测存在严重的滞后性，这样采取的调整铁水硅含量的措施也会滞后。为了消除检测的滞后性，及时的对铁水硅含量的调整采取及时有效的措施，有必要对铁水硅含量进行预测。基于模块化和信息融合思想，本文提出了集成模糊神经网络铁水硅含量预测方法，选取了2个预测子模块单独学习并训练，然后经过一个决策融合模块得到最终的铁水硅含量预测结果。建立了铁水硅含量预测模型，并在模型训练完成后进行了MATLAB实验仿真。仿真数据采集自凌源钢厂2号高炉，样本数据均在高炉炉况基本稳定的运行条件下获得，用这些样本数据训练预测模型。预测模型训练结束后，又选取50个样本对模型进行测试预测。仿真结果验证了该方法的有效性，集成模糊神经网络预测模型预测精度很高，相对误差较小，能够给予高炉生产过程给予有效的指导。
在煉鐵生產過程中，有效的降低鐵水硅含量併且使其保持在閤理的低水平，有利于提高生鐵的質量和產量。但是在實際生產過程中，生鐵的硅含量隻有在生鐵樣本送到實驗室經過化驗後方可得知，檢測存在嚴重的滯後性，這樣採取的調整鐵水硅含量的措施也會滯後。為瞭消除檢測的滯後性，及時的對鐵水硅含量的調整採取及時有效的措施，有必要對鐵水硅含量進行預測。基于模塊化和信息融閤思想，本文提齣瞭集成模糊神經網絡鐵水硅含量預測方法，選取瞭2箇預測子模塊單獨學習併訓練，然後經過一箇決策融閤模塊得到最終的鐵水硅含量預測結果。建立瞭鐵水硅含量預測模型，併在模型訓練完成後進行瞭MATLAB實驗倣真。倣真數據採集自凌源鋼廠2號高爐，樣本數據均在高爐爐況基本穩定的運行條件下穫得，用這些樣本數據訓練預測模型。預測模型訓練結束後，又選取50箇樣本對模型進行測試預測。倣真結果驗證瞭該方法的有效性，集成模糊神經網絡預測模型預測精度很高，相對誤差較小，能夠給予高爐生產過程給予有效的指導。
재련철생산과정중，유효적강저철수규함량병차사기보지재합리적저수평，유리우제고생철적질량화산량。단시재실제생산과정중，생철적규함량지유재생철양본송도실험실경과화험후방가득지，검측존재엄중적체후성，저양채취적조정철수규함량적조시야회체후。위료소제검측적체후성，급시적대철수규함량적조정채취급시유효적조시，유필요대철수규함량진행예측。기우모괴화화신식융합사상，본문제출료집성모호신경망락철수규함량예측방법，선취료2개예측자모괴단독학습병훈련，연후경과일개결책융합모괴득도최종적철수규함량예측결과。건립료철수규함량예측모형，병재모형훈련완성후진행료MATLAB실험방진。방진수거채집자릉원강엄2호고로，양본수거균재고로로황기본은정적운행조건하획득，용저사양본수거훈련예측모형。예측모형훈련결속후，우선취50개양본대모형진행측시예측。방진결과험증료해방법적유효성，집성모호신경망락예측모형예측정도흔고，상대오차교소，능구급여고로생산과정급여유효적지도。
Reducing silicon content in hot metal effectively and keeping it in a reasonable low level have significant impact. However, the silicon content in hot metal is not known until the steel samples are sent to laboratory to be tested. So, it is necessary to predict silicon content advanced. A method of modular integrated fuzzy neural network prediction of silicon content in hot metal is proposed based on the theory of modularization. To improve forecasting accuracy and decrease training time, integration method is also leaded into this paper. The whole forecasting network is divided into two sub-networks. These two sub-networks are trained respectively. Through this way, the training time is decreased by a large margin. Forecasting accuracy and system stability are all improved effectively at the same time. Date samples are acquired from a domestic steel blast furnace which runs in a stable operating condition. These samples are used to train the model. After training, 50 samples are selected to test the prediction model. The prediction results show that the method has good prediction effect.
