A MODEL PREDICTIVE APPROACH TO BLAST FURNACE OPERATIONAL MANAGEMENT AUTOMATION

2016, vol. 16, no. 4, pp. 36–49 36 Introduction A promising work direction to improving the efficiency of blast-furnace processes control is application of methods, based on modeling and predictive solutions. In general, the use of blast furnace models has a great history and a large number of sources on this topic is available. It is necessary here to note the works of national authors I. Tovarovskiy, A. Gotliba, G. Efimenko, A. Gimmelfarb, A. Pokhvisnev, O. Onorine, N. Spirin, A. Ramma, A. Dmitriev [1–26]. The works of V. Parshakov [12–15], devoted to study of influence of the cohesion zone parameters on the blast furnace process efficiency, deserve special attention. It is necessary to note among foreign authors the works of J. Kule, M. Sasaki, K. Ono, A. Suzuki, J.M. Burgess, D.R. Jenkins, K.L. Hockings, S.A. Kumar, N. Suresh, C.P. Jeffreson, M. Gobetto [27–41]. However, as far as the blast furnace process is quite sophisticated and its parameters are not fully observable the specified problem is not completely solved now and studies on the topic are still conducted. The main features of the proposed approach are: – usage of the operational data mining software to identify effective regions of the blast furnace technical parameters values, providing productivity increase and coke consumption reduction; – real-time software for identification of the furnace cohesion zone current parameters for the operational management correction; – forecasting of the blast furnace thermal state indicators dynamics when the blast parameters or charge load change. DOI: 10.14529/ctcr160405


The blast furnace control model general structure
The main difficulties preventing the achievement of high technical and economic efficiency levels are: 1) partial observability and controllability of processes; 2) the need for the processes stabilization in extreme boundary conditions; 3) incomplete knowledge about the current process state due to its complexity. To overcome the above difficulties an advanced methodology of model predictive management is now developed. The peculiarity of this methodology comes from usage of controlled object modelling software with permanent on-line updating based on constant parameters identification by the real operational data for the observability and controllability of processes.
Therewith, each managing step solves the problem of control action on technical and economic indicators optimization.
The general structure of the model-predicative management is shown on Fig. 1. Here in u k -the controlled parameters of the blast furnace process (BFp); z k -measured uncontrolled parameters of the BFp; w k -unmeasured disturbance factors of the BFp; q k -the blast furnace process efficiency indicator: 1, if BPf is satisfied with determined performance efficiency; 0, otherwise, k q     y k -the output measured parameters of the BFp; x k -the blast furnace process state vector, for a satisfactory prediction of BFp characteristics; р k -the measured parameters of the blast furnace process used to estimate its state vector; {(u, x, z, y, р, q): k = 1 … k -1} -previous BF melting parameters statistics; Мod i -the i-th model representation of the BFp, which provides state vector estimation according to technological instructions; МРС -the program of the model-predictive management calculation; k -the current melting index.
One BFp model representation for the state vector evaluation is the "AKOMM" Ltd. "Cohesion" system, which provides a quantitative assessment of the melting zone parameters.

Effective regions of the operating parameters clustering
The effective operating parameters regions are determined by the blast furnace target indicators settings, such as productivity, coke consumption, theoretical combustion temperature, furnace thermal state indicators (the cast iron silicon content, titanium module, blast-furnace gas utilization, etc.) For instance, Fig. 2 represents effective region detection, based on the target function: 1 , , 0, 1 k n ir c c n c n c e n b             , where п ir -relative cast iron perfomance; b c -the relative specific coke consumption; α п , α c -weights of partial indices n and c respectively as a part of a generalized target, reflecting the importance of performance and coke saving in the overall target structure.

Fig. 2
Complexity of the considered effective region of operating parameters clustering task, in accordance with the division into target areas defined by specified levels of the target function ( Fig. 2) caused due to is its high dimensionality. The number of operating parameters can be more than 70. To simplify the problem solution we used the method of exact area decomposition on the two-dimensional crosssection, analytically described by second order elliptical regions [42][43]. Fig. 3 represents an example of an increased efficiency area in the coordinates of the "specific coke consumption -cast iron silicon content" including furnace thermal state constraints.   To solve this problem a quadratic solution residual of inequalities system is formulated: where f ij -discriminant function analytically describing the effective region of the BF parameters.
The valid values of the parameters are restricted by inequalities: min max , .
The residual of the solution (1) is based on the technological process monitoring statistics. This statistics may be incomplete. In this case, the minimizing of solution residual problem formulation will be incorrect. We introduce the additional regularizing constraint to streamline problem formulation.
Herein{ } Ri x -valid values of operating parameters used for regularization; for example, the base parameters values obtained on the basis of technological calculations. Regularization usage is the central point of the approach proposed. Regularization allows generating consistent solutions based on both operating data and technological calculations.
The total residual of the inequalities solution (3) including constraints (5) is formulated as a penalty function: If the recurrent process converges, the result is a generalized solution of problem (7) under the given constraints.
For testing the BFp model-predictive control algorithms reporting forms are generated daily. They display BFp technological parameters of the current day, including the targets and factors of adaptive control. Based on reporting forms, a comparison is held between the ac the BFp efficiency and an effective regime.
Developed algorithms based on the models produced in SCADA "PolyTER" are clarified to specify the area of high quality thermal state including the effective values of the curren In addition, the approach for constructing the set of Pareto implemented, when solving the blast furnace modes optimization problem.
The minimum coke consumption was used as the target function. Calculations wer the following dependencies P i (C operational mode, along with thermal state constraints considering different silicon content. Fig. 4 represents the dependence between the optimal cast iron performance and coke quality ( As we can see from the graph, cast iron performance increases while the M 10 productivity growth rate slows significantly. This is particularly evident when the values of are close to a 8.0 and 8.2. valid values of managed parameters by the criterion of minim sidual (1) considering that the uncontrollable parameters within defined limits are streaming to provide the maximum of specified residual. This problem is a minimax mathematical programming problem: In general, the solution of the problem will be implemented by the gradient descent method. In this case, the recurrence algorithm for solving the problem would be as follows: If the recurrent process converges, the result is a generalized solution of problem (7) under the given predictive control algorithms reporting forms are generated daily. They display BFp technological parameters of the current day, including the targets and factors of adaptive control. Based on reporting forms, a comparison is held between the actual values of factors influencing the BFp efficiency and an effective regime.
Developed algorithms based on the models produced in SCADA "PolyTER" are clarified to specify the area of high quality thermal state including the effective values of the current mode.
In addition, the approach for constructing the set of Pareto-optimal non-improvable solutions was implemented, when solving the blast furnace modes optimization problem.
The minimum coke consumption was used as the target function. Calculations wer C c , W vapor ), Si(W vapor ), М 10 (W vapor ), W vapor when operational mode, along with thermal state constraints considering different silicon content. Fig. 4 represents the dependence between the optimal cast iron performance and coke quality ( from the graph, cast iron performance increases while M 10 falls, therewith decreasing productivity growth rate slows significantly. This is particularly evident when the values of Pareto region dependency between the cast iron performance andcoke quality ( valid values of managed parameters by the criterion of minimum constraints residual (1) considering that the uncontrollable parameters within defined limits are streaming to provide the maximum of specified residual. This problem is a minimax mathematical programming problem: (7) the gradient descent method. In this If the recurrent process converges, the result is a generalized solution of problem (7) under the given predictive control algorithms reporting forms are generated daily. They display BFp technological parameters of the current day, including the targets and factors of adaptive tual values of factors influencing Developed algorithms based on the models produced in SCADA "PolyTER" are clarified to specify t mode. improvable solutions was The minimum coke consumption was used as the target function. Calculations were made for setting the desired BF operational mode, along with thermal state constraints considering different silicon content. Fig. 4 represents the dependence between the optimal cast iron performance and coke quality (M 10 ). falls, therewith decreasing productivity growth rate slows significantly. This is particularly evident when the values of

Operational BF real-time management
The solution of the BF operational management problem is implemented from point of view assu ing the regime parameters dynamic stabiliza tion to this problem is very difficult, since the blast plex properties: 1) BFp dynamics, taking long time intervals (up to 40 hours); 2) nonlinear nonstationary characteristics; 3) distributed parameters; 4) high level of disturbances; 5) low observability of many process

Methods of operational management
To illustrate the operational management techniques we will consider th con content regulation via the channel of specific coke consumption influence ( The transfer function between the cast iron silicon can be represented by two sequential dynami The transport delay estimation is based on calculating correlation function between the cast iron silicon content and B coke . Correlation function maximum defines the value of transport delay. A recu rence relation describes the inertial delay: where a, b -unknown coefficients, identified on current operational data; Identification of the a, b coefficients is carried out with synchronized real time operat and B coke data. For silicon level regulating using the recurrent formula (8) necessary correction of the coke flow is calculated to achieve the required level of the cast iron silicon content. The required amount content is set through the previously discussed optimization problem based on the usage of efficient BF mode clusters.
For instance, Fig. 6 represents an example of silicon level regulation in an operational environment of SCADA "PolyTER".
Currently implemented operational content, iron temperature,titanium module, furnace shaft temperature, furnacehearth temperature.
The framework of SCADA "PolyTER" is displayed in Fig. 7.

Автоматизация оперативного управления доменным процессом с использованием модельно-упреждающего подхода
Вестник ЮУрГУ. Серия «Компьютерные технологии, управление, радиоэлектроника». Fig. 5 shows the obtained dependency between the optimal values of coke consumption and coke leads to coke consumption decrease, however the significant deceler of coke consumption rate is not observed, when reducing M 10 . time management The solution of the BF operational management problem is implemented from point of view assu ing the regime parameters dynamic stabilization in the defined increased effectiveness regions. The sol tion to this problem is very difficult, since the blast-furnace process as a controlled object has very co 1) BFp dynamics, taking long time intervals (up to 40 hours); ear nonstationary characteristics; 5) low observability of many process characteristics.

Methods of operational management
To illustrate the operational management techniques we will consider the example of cast iron sil con content regulation via the channel of specific coke consumption influence (B coke The transfer function between the cast iron silicon content (Si) and the specific coke consumption can be represented by two sequential dynamic delays: transport and inertial.
The transport delay estimation is based on calculating correlation function between the cast iron . Correlation function maximum defines the value of transport delay. A recu es the inertial delay: unknown coefficients, identified on current operational data; k -current time.
coefficients is carried out with synchronized real time operat For silicon level regulating using the recurrent formula (8) necessary correction of the coke flow is calculated to achieve the required level of the cast iron silicon content. The required amount ough the previously discussed optimization problem based on the usage of efficient BF ig. 6 represents an example of silicon level regulation in an operational environment Currently implemented operational management uses the following parameters:the cast ironsilicon content, iron temperature,titanium module, furnace shaft temperature, furnacehearth temperature.
The framework of SCADA "PolyTER" is displayed in Fig. 7.
Автоматизация оперативного управления доменным процессом упреждающего подхода 41 Fig. 5 shows the obtained dependency between the optimal values of coke consumption and coke leads to coke consumption decrease, however the significant decelerabetween the coke consumption and coke quality (M10) 5, were built when processing statistical information for the pe-10 OJSC "MMK" excluding the downtime periods.
The solution of the BF operational management problem is implemented from point of view assumtion in the defined increased effectiveness regions. The solufurnace process as a controlled object has very come example of cast iron silicoke ). and the specific coke consumption The transport delay estimation is based on calculating correlation function between the cast iron . Correlation function maximum defines the value of transport delay. A recur- current time. coefficients is carried out with synchronized real time operational silicon For silicon level regulating using the recurrent formula (8) necessary correction of the coke flow is calculated to achieve the required level of the cast iron silicon content. The required amount of silicon ough the previously discussed optimization problem based on the usage of efficient BF ig. 6 represents an example of silicon level regulation in an operational environment management uses the following parameters:the cast ironsilicon content, iron temperature,titanium module, furnace shaft temperature, furnacehearth temperature.
The left part of the diagram shows the graphs of the BF thermal state indicators: cast iron silicon content, titanium module, furnace shaft temperature and furnace hearth temperature. Here each indicator matches its current value (gray-green color), filtered current values (black color), forecast values (yellow color), target values (red color) and its visual level (purple color).
The middle part of the graphs shows current values of the controlled parameters: steam consumption for humidification, natural gas consumption, oxygen content in the blast, theoretical combustion temperature, and control parameters values 2-hour averaged: coke consumption per feeding, specific coke consumption per ton of calculated iron, ore load. Moreover, system displays job graphs for the specific coke consumption: 1) defined according to the desired value of the cast iron silicon content (red line); 2) based on weighted decision (blue line). Weighted decision is using specific coke tasks defined by the required values of the BF thermal state indicators.
Coke consumption per feeding considering cast iron per feeding is counted through the framework charts of specific coke consumption.

1.
To improve the blast furnace process efficiency we propose a model predictive control approach, based on clustering of blast furnace process effective parameters values regions and operational parameters stabilization within an efficient cluster using the forecast based on their dynamics.
2. We introduce method of multidimensional blast furnace effective parameters values regions decomposition using two-dimensional cross-sections and their analytical representation based on second order elliptic regions.
3. We provide the algorithmic solution for the optimal choice of blast furnace process controlled parameters with uncontrollable parameters possible variations using the multidimensional effective parameters values regions.
4. It is shown that using Pareto regions in the blast furnace process parameters coordinates, derived from the optimization problem solution, allows to assess the potential boundaries of high performance efficiency depending on the operating parameters. The obtained dependences can be used for technological calculations of the mode parameters estimating potential attainability of the effective values.
5. Blast furnace process regime parameters stabilization should be carried out based on predicting of their values using models of the dynamics. The real time dynamic models parameters identification algorithms are developed.
6. The automated blast furnace process decision support system was developed based on the methodical and algorithmic maintenance, implemented in SCADA "PolyTER".