Study On Prediction Of Blast Furnace Gas Production And Emission Based On Big Data

Aiming at the problems of low utilization rate of the related data about blast furnace gas,difficulty in establishing prediction model and low prediction accuracy,it is very important to establish an accurate and efficient prediction model of blast furnace gas production and emission by using a large amount of historical data reasonably.The establishment of prediction model for blast furnace gas production and emission can not only provide decision information for dispatchers,but also reduce energy consumption of steel enterprises and realize energy-saving and pollutant emission reduction.Based on the field investigation of a steel enterprise in Tangshan in the early stage,a total of 105 measuring points and more than 10 G of historical data were collected.The experimental data was preprocessed by Box diagram analysis and linear normalization,and seven main influencing factors of gas production were extracted by gray relation analyses.Based on the post-treatment data,two prediction models were used to predict the production and emission of blast furnace gas.Firstly,a combination prediction model based on LSTM and ARIMA is proposed,which is based on the comparison between the real values of experimental data,the prediction results of LSTM model and ARIMA model.The comparison results show that the LSTM prediction results are generally lower than the true values,and the ARIMA prediction results were generally higher than the true values.Based on the above phenomena,a combination model based on LSTM and ARIMA is proposed,which combines the prediction results of the two models by using CRITIC method.Compared with other models,mean square root error of the combination model reduces 2.325,obviously improves the drawbacks of the ARIMA and LSTM in predictive characteristics,reduces the prediction error,and is suitable for forecasting blast furnace gas emission.Secondly,the MR-BP was built.With rising data sample of blast furnace gas,the time spent on model training is also greatly increased.In order to solve this problem,a method of combining MapReduce architecture with BP parallelization was used to predict the production and emission.The model assigns BP neural network training to multiple map tasks,the network weight correction values obtained by each map task are summarized by reduce function,and using the final correction weights adjust the network parameters.By comparing BP model,combination model based on LSTM and ARIMA and MR-BP,it is concluded that training time of three models are 148.20 s,112.33 s and 57.524 s respectively,and the mean square root error of the prediction results is 5.193,2.335 and 3.436,respectively.Experimental results show that the MR-BP greatly shortens the time of model training,ensures the real-time prediction.The prediction is accurate,but it is slightly worse than the LSTM and ARIMA combination prediction model.Figure 29;Table 11;Reference 54.