Prediction model of converter gas production based on deep-learning
The prediction of the occurrence of converter gas provides important support for the micro differential pressure control at the converter inlet,the improvement of converter gas recovery quality,and the overall gas bal-ance scheduling of the plant.Based on the actual occurrence data of converter gas during the blowing process of a certain steel plant,a deep learning method was used to establish three prediction models for converter gas occur-rence:BP neural network,LSTM long short memory neural network,and RBFNN radial basis function neural net-work.The effects of three parameters,namely prediction steps,input sample size,and hidden unit number,on the accuracy and computational efficiency of the prediction model were compared and analyzed.The research results in-dicate that the prediction accuracy of the model decreases with the increase of prediction steps,and choosing 30 step prediction is more in line with the actual needs of steel mills.As the sample input increases,there is no significant change in the accuracy of LSTM,while the accuracy of BP shows a decreasing trend.The accuracy of RBF first in-creases significantly and then slowly decreases.The prediction efficiency of LSTM showed no significant change,BP significantly decreased,and RBF remained unchanged.When the three models are under the optimal sample input and 30 step prediction conditions,the accuracy of LSTM remains basically unchanged as the number of hidden units increases.BP first slightly increases and then slowly decreases,while RBF first increases significantly and then re-mains stable,and then decreases significantly.The prediction efficiency of LSTM has slightly decreased,BP has significantly decreased,and RBF remains unchanged.Finally,under the condition of 30 step prediction,the optimal parameter conditions for LSTM,BP,and RBF models are as follows:LSTM sample input quantity is 125,hidden unit number is 135,ERMS minimum is 13.38,and training duration is 4.7 min;The input amount of BP samples is 50,the number of hidden units is 60,and the minimum ERMS is 31.46,with a training duration of 16.8 min;The in-put amount of RBF samples is 210,and the number of hidden units is 210.At this time,the minimum ERMS is 2.07,and the training duration is 1.2 min.Compared with actual data,RBF has the best prediction effect.By using the prediction results of converter gas generation to regulate the speed of the fan,the micro differential pressure at the furnace mouth can be maintained in a more stable state,reducing the suction air volume,and improving the heat val-ue of the recovered gas.
blowing processconverter gasprediction of occurrence volumedeep-learningcomparative analysis of models
NETL
NSTL
万方数据
