Prediction Of Gas Production And Supply Based On LSTM/SARIMA Time Series Model

The steel industry has always been an industry with high energy consumption and high emissions.In recent years,various enterprises have put forward various countermeasures in response to the national policy of energy conservation and emission reduction and "dual carbon",efficient use of various energy resources and reduction of environmental pollution.By-product gas,mainly blast furnace gas,coke oven gas and converter gas,is an important secondary energy produced in the process of metallurgical production.The rational use of them can not only improve the secondary utilization efficiency of the overall energy,but also achieve energy-saving effect.On the one hand,since the components of by-product gas are mostly sources of environmental pollution,reducing the release of by-product gas can reduce the harm to the environment to a certain extent.The premise of efficient and reasonable utilization of these by-product gas is not only to have reasonable gas scheduling rules,but also to accurately predict the production and consumption of gas in the future.Regarding the amount of gas production,this paper discusses the deficiencies of domestic and foreign scholars’ prediction models for by-product gas,mainly including the following points:(1)The time granularity is too large,the sample is too small,and the number of prediction steps and times is too small,which cannot be comprehensively investigated.Model prediction ability;(2)The model lacks the comparison of prediction results under special working conditions(such as equipment failure,production reduction,production stoppage,etc.),or the model is difficult to adapt to special working conditions.In view of the above problems,this paper selects small-granularity data for multi-step prediction of large samples,and uses the long-short-term memory model(LSTM)in deep learning and the seasonal difference autoregressive moving average model(SARIMA)in linear regression for prediction,comparative analysis Different time series models predict the effect.Aiming at the problem of special working conditions,this paper proposes a gradient-driven time series prediction composite model.According to the gradient change of the input data,the most suitable model is selected for prediction,which greatly increases the prediction accuracy.By comparing the predicted value of gas flow with the actual collected gas flow,the average relative error(MAPE)is used as the judgment basis.The results show that:under normal conditions,the accuracy of LSTM is higher than SARIMA,and in the30-step prediction,LSTM The average relative error is 0.0428,while the SARIMA model is 0.0648;the LSTM model has the smallest error when the input sample is 100,the average relative error is 0.0428,the SARIMA model has the smallest error when the input sample is 200,and the average relative error is 0.0570;LSTM model in During self-update,it is most appropriate to select 700 min of data as each training sample,which not only ensures the accuracy of the model(the average relative error is 0.0448),but also saves computing resources;under variable working conditions,the SARIMA model is more accurate than the LSTM model.The average relative error of SARIMA is 0.0694,and the average relative error of LSTM is 0.094.Therefore,it is recommended to use the gradient-driven time series prediction composite model for prediction under field operating conditions.Higher prediction accuracy.Finally,the prediction model of each gas production amount is established,and the established prediction model is deployed to the energy management and control center of a steel plant through SOCKET communication technology.Scheduling provides a more efficient method.