Intelligent Demand Forecasting Method And Application Of Fused Magnesium Furnace Group Combining Mechanism Model And Data-Driven Model

The fused magnesium furnace(FMF)is a high-energy-consuming device that converts magnesite into fused magnesia through complex physical and chemical processes.Fused mag-nesia is a raw material for advanced refractory materials in the manufacturing and aerospace industries.The electricity demand for the fused magnesium furnace group(FMFG)is calcu-lated by the moving average of the furnace group power within a given period,which is used to measure the electricity consumption of FMFG.In order to save energy,the demand for FM-FG can not exceed the peak demand set by the electric utility company.The fused magnesia producers set a limit value in order not to exceed the peak demand.When the actual demand for FMFG exceeds the limit value,the power supply of some FMFs must be cut off to ensure that the total demand is not exceeded.However,the balance between the exothermic and the endothermic effect in the furnace is affected by power cuts,resulting in a lower-quality and reduced production of magnesia product.The smelting current control system controls the cur-rent of FMF so that the error between the actual current and the setting value of smelting current is as small as possible,to ensure that the magnesite in the furnace is completely converted into fused magnesia and the consumed electric energy is as small as possible.When the content of impurity components of magnesite increases and the particle length of magnesite becomes larger,the impedance decreases,the current becomes larger,and the demand rises,exceeding the limit value.At this time,the current control system controls the current to track the set smelting current,so the current decreases,the demand decreases,and the demand is lower than the value limitation.Therefore,there is a demand spike phenomenon in which the demand first increases beyond the limit value and then falls below the limit value.When the demand spike exceeds the limit value,it will cause unnecessary power cutoff.Accurate demand forecasting is of great significance to prevent unnecessary cutoff of power supply caused by demand spikes,and it also demonstrates practical reference value for forecasting industrial process operation indicators.This paper relies on the National 973 Program "Comprehensive Production and Pro-cess Integrated Control System Overall Control Strategy and Operation Control Method(2009CB320601)",National Natural Science Foundation Project "Overall Optimization Con-trol Research Based on Data and Model for Complex Industrial Processes(61020106003)",taking the smelting process of FMFG of Dashiqiao City New Development Refractory Mate-rial Group Co.,Ltd.in Liaoning Province as the background,the research on the intelligent demand forecasting method of FMFG and its application is carried out.The main contributes are summarized as follows.1)Through the mechanism analysis of the smelting process of FMF and the dynamic characteristics analysis of the current control system,the changing rate of power model of a single FMF is established,and then the rate of change of power model of FMFG is obtained.On this basis,the demand forecasting model structure of FMFG is established,and the challenging problems of demand forecasting of FMFG by using the demand forecasting model structure are analyzed.2)Combining the smelting process mechanism modeling of FMF,system identification method and neural networks,the following three demand forecasting methods of FMFG are proposed:data-driven forecasting method based on RBF neural network,mechanism mod-el and data driven forecasting method based on offline identification,intelligent forecasting method using data-driven and mechanism model based on online alternating identification.The simulations experiments are carried out on the actual data of the smelting process of the FMFG for the three demand forecasting methods.The experimental results show that the mean abso-lute error(MAE),root mean square error(RMSE),the true positive rate(TPR)of the upward trend and the true negative rate(TNR)of the downward trend using the proposed intelligent forecasting method are significantly better than the first two methods.The above three demand forecasting methods are as follows.①It is analyzed that the key to the demand forecasting of FMFG is to obtain the rate of change of the power of FMFG.Considering that the rate of change of power of FMFG is an un-known nonlinear function with unknown order,it is related to the power of FMFG in the past.An order identification method based on the partial autocorrelation function(PACF)for this unknown nonlinear function is proposed.Based on this method,a data-driven demand fore-casting method consisting of PACF-based nonlinear function order identification module,RBF neural network based forecasting model of the rate of change of the power of FMFG and the demand forecasting model is proposed.The method is verified by the simulation experiments on the actual data of the smelting process of FMFG.The simulation results show that the MAE is 19.0950,the RMSE is 39.1921,the TPR of the upward trend is 80.34%,and the TNR of the downward trend is 80.48%.②In order to further improve the accuracy of demand forecasting,in response to the problem that the above forecasting method does not fully use the information of the mechanism model of the smelting process,the dynamic characteristics of the current control closed-loop system of FMF are used to establish a model of the rate of change of the power of FMFG.The model consists of a linear model and an unknown nonlinear term.The offline identification method is used to obtain the linear model parameters.The RBF neural network based on PACF order identification is used to estimate the unknown nonlinear term.Then a mechanism-and-data driven forecasting method based on offline identification is proposed,which consists of the demand at the current moment,the powers of FMFG at current and past times,and the estimate of the rate of change of the power of FMFG at the current moment.Simulation experiments are carried out through the actual data of smelting process of FMFG.The simulation results show that the MAE of this method is higher than that of the former forecasting method,but the RMSE is reduced by 8.88%,the TPR of the upward trend is increased by 0.44%,and the TNR of the downward trend is increased by 0.86%.③The linear model parameters and unknown nonlinear term of the power changing rate of FMFG are changed with the smelting and feeding process and the raw materials,while the linear model and the nonlinear term interact with each other.To solve this problem,the maximal information coefficient(MIC)method and the rule reasoning are combined,then an order identification algorithm for unknown nonlinear term is proposed.The saturation function is used to improve the traditional alternating identification algorithm,so that the linear model and the nonlinear term are switched in the direction thus the modeling error is reduced.On this basis,an intelligent forecasting method consisting of nonlinear term order identification,model alternation identification and demand forecasting model is proposed.Through the simulation experiments of the actual data of smelting process of FMFG,the simulation results show that the forecasting accuracy of this method is significantly improved compared with the previous two forecasting methods.Compared with the first method,the MAE is reduced by 4.66%;compared with the second method,the RMSE is reduced by 4.93%,and the TPR of the upward trend is improved by 7.61%,and the TNR of the downward trend is increased by 4.52%.3)The demand forecasting simulation experiment system is designed and developed.The experimental system consists of hardware platform and software system.The hardware plat-form consists of a cloud forecasting model training server,a data acquisition and processing computer,a demand forecasting computer,and a mobile Android terminal.The software sys-tem consists of forecasting model training software,data acquisition and processing software,demand forecasting software and mobile monitoring software.Among them,the forecasting model training software includes forecasting accuracy setting module,forecasting algorithm parameters and weight correction and downlink module,forecasting algorithm evaluation mod-ule and remote web monitoring module.The data acquisition and processing software includes modules such as furnace group power data acquisition module,data preprocessing module,and data communication uploading module.The demand forecasting software includes intelligent forecasting algorithm module combining mechanism model and data driven model,parameter correction module,real-time demand forecasting module,display-and-upload module.The mo-bile monitoring software includes modules such as demand monitoring module and forecasting accuracy display module.The forecasting model training software runs on the server in the cloud,and the initial value of the parameter is selected by the intelligent forecasting algorithm combined with the mechanism model and the data driven model.The simulation experiment is used to determine whether to download the corrected weigh parameters.The data acquisi-tion and processing software runs on the data acquisition and processing computer to collect and process the power data of FMFG.The demand forecasting software runs on the demand forecasting computer,and realizes the intelligent forecasting algorithm.The mobile monitor-ing software runs on the Android mobile phone terminal to realize mobile monitoring of the demand.Using the actual data of the smelting process of FMFG,the simulation experimental system is used to simulate the proposed forecasting method.The effectiveness of the proposed method and the demand forecasting simulation experimental system are verified.4)For the smelting process of FMFG consisting of 4 FMFs supported by No.1 power transformer of New Development Refractory Group Co.,Ltd.,the intelligent forecasting algo-rithm combined with mechanism model and data driven model is used to develop the demand forecasting application software and to complete the industrial experiments.The experimental results show that the MAE is 14.2372,the RMSE is 22.9244,the TPR of the upward trend is 84.61%,and the TNR of the downward trend is 85.20%.