METHOD OF GENERATING OPTIMAL CONTROL PROBLEMS FOR INDUSTRIAL PROCESSES

FIELD OF THE INVENTION

The invention relates to the field of industrial process modelling, optimal control and advanced scheduling. It departs from a method of merging two arbitrarily interconnected Mixed Logical Dynamic (MLD) subsystems into a combined MLD system and a method of designing a model based controller or scheduler for real-time control or scheduling of an industrial process.

BACKGROUND OF THE INVENTION

Mixed logical dynamical (MLD) systems, as introduced e.g. in the article "Control of Systems Integrating Logic, Dynamics, and Constraints", A. Bemporad and M. Morari, Automatica 35(3), 1999, pp 407-427, represent a mathematical framework for modeling systems described by interacting physical laws, logical rules, and operating constraints, generally called "hybrid systems". MLD systems are determined or described by linear dynamic equations subject to linear mixed-integer inequalities involving both continuous, i.e. real-valued, and binary, i.e. boolean-valued, variables. The variables include continuous and binary states x, inputs u and outputs y, as well as continuous auxiliary variables z and binary auxiliary variables δ, as described in the following equations: ■ x(t+l) = A x(l) + B1 u(t) + B2 δ(t) + B3 z(t) (Eq.1 a) y(t) = C x(t) + D1 u(ϊ) + D2 φ) + D3 z(t) (Eq. Ib) E2 %t) + E3 z(t) <= E] u(t) + E4 x(t) + E5 (Eq. Ic) In general, the variables mentioned are vectors and A, Bu C, D1-, E,- are matrices of suitable dimensions. In order to be well posed, the MLD system (Εqs.l) must be such that for any given x(t) and ιι(t) the values of δ(t) and z(t) are defined uniquely. Formulations or relationships in the form of Eqs.l appear naturally when logical statements are written as prepositional calculus expressions, or when bounds on the states are set explicitly. Among the advantages of the MLD framework are the possibility to automatically generate the matrices of MLD systems from a high-level description. MLD systems generalize a wide set of models, among which there are linear hybrid systems and even nonlinear systems whose nonlinearities can be expressed or at least suitably approximated by piecewise linear functions. The drawback of the MLD systems approach is the relative complexity of the theory, which in turn makes the modelling and maintenance of a complex industrial system difficult for people without a background in mixed integer optimization. Model predictive control (MPC) is a procedure of solving an optimal-control problem, which includes system dynamics and constraints on the system output, state and input variables. The main idea of model predictive control is to use a model of the plant or process, valid at least around a certain operating point, to predict the future evolution of the system. Based on this prediction, at each time step t the controller selects a sequence of future command inputs or control signals through an on-line optimization procedure, which aims at optimizing a performance, cost or objective function, and enforces fulfillment of the constraints. Only the first sample of the optimal sequence of future command inputs is actually applied to the system at time t. At time MT, a new sequence is evaluated to replace the previous one. This on-line re-planning provides the desired feedback control feature. Model Predictive Control can be applied for stabilizing a MLD system to an equilibrium state or to track a desired reference trajectory via feedback control. Standard routines can then be used to recast the previous control problem into a Mixed Integer Linear Programming (MBLP) of the following form:

min / (v) = min g 7 v s.t. A v ≤ b (Eq.2), V V ~

where the optimization vector v includes the aforementioned MLD input and auxiliary variables u, δ, z via vr = [MΓ Jf z. 3, the vector g denotes a cost vector, the matrix A is a constraint matrix and the vector & is termed a boundary vector. Standard procedures can be adapted to handle time varying parameters such as cost and price coefficients. The solution of the MILP problem can be obtained by reverting to the commercial optimisation problem solver for solving linear, mixed-integer and quadratic programming problems called CPLEX (http://www.ilog.com/products/cplex/). Model predictive control (MPC) in combination with Mixed Logical Dynamical (MLD) systems descriptions has been used for modeling and control of processes in the utility automation and process industry. By way of example, a method of scheduling a cement production is described in the article "Using Model Predictive Control and Hybrid Systems for Optimal Scheduling of Industrial Processes", by E. Gallestey et al., AT Automatisierungstechnik, Vol. 51, no. 6, 2003, pp. 285-293.

DESCRIPTION OF THE INVENTION

It is therefore an objective of the invention to facilitate the establishment of Mixed Logical Dynamic (MLD) systems in order to model and control complex industrial processes, based on MLD subsystems representing sub-processes of said industrial processes and in view of an automated control, scheduling or simulation of said industrial processes. These objectives are achieved by a method of and a computer program for merging two arbitrarily interconnected MLD subsystems into one combined MLD system according to the claims 1 and 8, as well as a method of designing a model based controller or scheduler for real-time control or scheduling of an industrial process according to claim 6. Further preferred embodiments are evident from the dependent patent claims. According to the invention, the original MLD subsystems are related or connected via pairs of dependent input and output variables, which pairs are transformed into additional auxiliary variables of the combined MLD system. The equations describing the original MLD subsystems and involving said dependent input and output variables in turn are converted into constraints of the combined MLD system involving the additional auxiliary variables. Hence the core technical aspect of the invention concerns the automatic merging of two arbitrarily connected MLD subsystems into one composite MLD system. In a preferred variant of the invention, a repeated use of the merging procedure allows to generate any arbitrarily complex MLD system containing the complete description of an industrial process. Provided that the industrial process can be decomposed in several sub- processes, each MLD subsystem representing a sub-process is established independently from the others in the MLD framework, and the industrial process is reproduced as a sequence of interconnected MLD blocks. In a preferred variant of the invention, an optimal control problem is generated via adding an ultimate MLD block whose unique output becomes the objective function of the problem. In combination with Model Predictive Control (MPC) techniques, both on-line optimal control problems and scheduling applications can be generated and solved. The invention simplifies the modelling of complex industrial processes in the Mixed Logical Dynamic (MLD) framework and allows to automatically generate and solve optima] control problems and scheduling applications for said industrial processes. Preferably, this is done by graphically representing any MLD subsystem by a block whose input / output ports match the input / output variables of the MLD formulation. In a graphical environment, any specific industrial process may then be reproduced by instantiating blocks from a library of basic MLD elements or atomic MLD blocks and by properly connecting them in order to obtain a graphical representation of the process. In case an appropriate library is available, this procedure will not require any expert knowledge from the end user apart from the ability to build the graphical interconnections mentioned.

BRIEF DESCRIPTION OF THE DRAWINGS

The subject matter of the invention will be explained in more detail in the following text with reference to preferred exemplary embodiments which are illustrated in the attached drawings, in which: Fig.l shows two MLD subsystems being merged into a combined MLD system, and Fig.2 depicts the MLD subsystems representing the sub-processes of the grinding section of a cement plant.

DETAILED DESCRIPTION OF PREFERRED EMBODIMENTS

Fig.l shows two MLD blocks or subsystems Sl and S2, defined respectively by the matrices A1, B1U, C1, D}},..ι3, E1 h.._5 and A2, B2j.j, C2, D2lt..,3, E2]^, as well as by the vectors of states xj, &, the vectors of inputs uj, u∑, the vectors of outputs yj, yj, the vectors of real auxiliary variables zj, &, and the vectors of Boolean auxiliary variables δj, δ?. The MLD subsystem Sl, S2 are interconnected as indicated by the arrows joining the two blocks. The goal is to merge Sl and 52 in order to obtain one single or combined MLD system 6" described by the matrices A, B1..3, C, Eirψ<s and the corresponding vectors of states x, inputs u, outputs y_, real auxiliary variables g and Boolean auxiliary variables δ. The matrices of the combined system 5 are obtained by creating matrices comprising the corresponding matrices of the subsystems Sl, S2 along their diagonal as a starting ~Al 0 point, e.g. by setting A = and by subsequently eliminating or exchanging rows _ 0 A2 and/or columns as explained below. Likewise, the vectors of the combined system S are created by augmenting the vectors of Sl with the corresponding vectors of S2 as a starting point, e.g. by setting x7 = [xj1 x/], and by subsequently eliminating or rearranging their elements if necessary. It is to be noted that the matrices and vectors of the combined system S are uniquely designated and maintain their designation throughout the subsequent operations performed on their rows, columns or elements. The vectors of states, inputs and outputs can contain both real (denoted with the index r) and Boolean (index b) variables. If this is the case, all the real variables will have to be grouped and placed in the corresponding vector before the Boolean ones. Among the inputs and outputs certain variables are independent of the other MLD block (denoted with the prefix z) and some are dependent (prefix d). Pairs of the dependent variables are represented in Fig.l by a common arrow pointing from one MLD block (Sl, S2) to the other MLD block (S2, Sl). The treatment of the embodiment according to Fig.l will be detailed in the following, yet less general configurations with e.g. only one, two or three pairs of dependent variables are also conceivable and can be treated in the framework of the present invention. The resulting state vector x of the combined MLD block S consists of a composition of all the states of Sl and S2 and is defined as x = xr2, Xbi, Due to the real variables of both systems (xri, xa) being placed before the Boolean variables [xj,j, Xti), the matrix A of the combined MLD system S, based on the matrices A1, A2 of the subsystems Sl, S2 as shown above, has to have the rows and columns corresponding to xj,i swapped with the rows and columns of xa. Likewise, the other matrices Bj j involved in the state dynamic of the system S, based on the corresponding matrices (B1J j, B2J.J) of the subsystems Sl, S2, undergo, in a first step, a corresponding swap of rows. Equally, in a first step a swap of columns is executed on the matrices C, E4 of the system S multiplying the resulting state vectors and being based on the corresponding matrices (C1, E14\ C2, E24) of the subsystems Sl, S2. The resulting output vector v. of the combined MLD block S consists of a composition of the independent outputs of Sl and S2 and, taking into account the "real" before "Boolean" principle, is defined by 3; = [iy,], ijr2, ijbi, iyjbil- Again, the matrices involved in the output dynamic (C, Dj .^) of the system S are based on the corresponding matrices (C1, D i,..,3\ Cr> D i,..,s) of the subsystems Sl, S2. Along the same idea, the rows of these matrices corresponding to jvj,; will have to be swapped in a first step with the rows corresponding to iχr2. Additionally, as the dependent outputs are lost, the corresponding rows of the C, D 2 ,..,3 matrices have to be deleted. The resulting input vector u of the combined MLD block 6" consists of a composition of the independent inputs of Sl and 52 and is defined as u = [iurj, iur2, iubi, iubi\- Again, the real variables of both systems are being placed before the Boolean variables. The dependent inputs have been dropped. This implies that the input matrix Bj, in a second step, will undergo a swap between the columns corresponding to iubi and iu^ and, in a third step, will have to have the columns corresponding to the dependent inputs (durjt dubj and duj-2-, dubi) deleted. Likewise, the other input matrices Dj, Ej of the system S, being based on the matrices (D j, E j; D2J, E2f) of the subsystems Sl, S2, will undergo a swap of columns between iu±,ι and iur2 and, in a second step, will have the columns corresponding to the dependent inputs (du,j, deleted. The eliminated real dependent inputs durj, dur2 will then be substituted by additional real auxiliary variables zj2, Z21 added to the real auxiliary variables z of the combined system S. Thus the real auxiliary variables of the MLD system S, z = [zj, ∑2, Zi 2, 22;], include the original auxiliary variables (gj, z£) of Sl and S2 plus the ones introduced by the S1-S2 and S2-S1 real interconnections. Similarly, the Boolean dependent inputs dubi, dub2 will be replaced by additional Boolean auxiliary variables δ.12, hi added to the Boolean auxiliary variables δ of the combined system 5. This transformation is obtained or executed by appending the columns deleted from the matrices Bj, Dj and Ej (observing a change of sign for the latter) to the matrices B3, D3, and E3 if real or to the matrices B2, D2, and E2 if Boolean. The equations dy_j = du^ and Jv^ = duj between the dependent inputs and outputs of the two subsystems S2 and Sl represent the connections between the subsystems. The information contained therein has to be retained and included in the formulation of the combined MLD system 5. However, as the dependent variables (dj, du) do not appear in the final formulation, this is not a straightforward task. In the case of the S1-S2 connection (short straight arrows in Fig.l), the original expression for the dependent outputs dy_j of subsystem Sl is comprised in the set of equations yj = C! xj + D'J U1 + D^ S] + DJ3 ZJ- (Eq. Ib) As a result of the row-elimination procedure above, these expressions for dyj are absent from the resulting equations for the output variable v. of the combined MLD block S. Eq. Ib comprises terms of the type D j dui wherein the dependent inputs duj have to be replaced by the additional auxiliary real and/or Boolean variables {zj2, δ_n) introduced above. Likewise, the dependent inputs dua have to be replaced by the additional auxiliary variables (zai, <hi) introduced above. Thus, the equations du2 = dyj translate into

[zl2\ / ^2i]= dyrl ldyM *= C11 X1 + D1-1I «i + D^17S127 f + D1-1S ^17 Zn7J

(Eq.ld) where the tilde and the subscript matrix index i (C1- ,D, 1...3 ) signify that only the i-th rows and the relevant columns of the matrices C1 and D!j.,j are exploited. This equality constraints involving only state variables (xj), independent input variables (?«;) and auxiliary variables £gj zzi Zj2; δj δ_n h.Λ obtained with the previous manipulations or substitutions can therefore be added as additional rows to the matrices E1 of the combined MLD system S. Additional constraints in terms of MLD inequalities are formed by turning every single equality or line in the above vector equation Eq.ld into two inequalities or rows by noting the former both in form of <= and in the form of >=. The case of the S2-S1 connection (long returning arrows in Fig.l) is similar and results in exchanged indices. The combined or merged MLD system S resulting from the procedure outlined above in turn can be considered as a subsystem itself. Via iterated application of the procedure, a complex MLD system S!ot representing all relevant technical aspects of an arbitrarily complex industrial installation or process can be generated. Thereby order of merging of the "basic" MLD subsystems or "atomic" MLD blocks does not have any impact on the final result and may be chosen at will. In order to use the MLD system for a model based control procedure an objective function or cost functional /has to be defined. In the context of the present invention, this is most conveniently done via adding a functional MLD subsystem SCOst comprising exactly one independent scalar output variable ycos! representing the objective to be optimized. This functional MLD subsystem Scos, is ultimately merged to the complex MLD system S10, representing all relevant technical aspects of the physical process. Thereby, all the outputs of the complex MLD system Stot that are related to costs are regarded as dependent outputs with corresponding dependent inputs of the functional MLD subsystem SCOst- Obviously, there is no restriction on the number of independent inputs u serving as control variables or command inputs of the single-output MLD system S that results from the aforementioned ultimate merger.

The method according to the invention is completely general and can be applied to many industrial processes with a limited effort. In the following, a particular implementation of the invention in a cement plant is presented in more detail with reference to Fig.2. Cement mill scheduling implies deciding when to produce a certain cement grade on which mills. Due to the number of mills, grades, silos, conveyor belts and the various operating constraints the problem is quite complex. In a hybrid system framework, the mills are treated as finite state or on/off machines that associate different energy consumptions and throughputs depending on the cement grade produced. Generally, a library of basic MLD blocks or subsystems is provided for each basic element or sub- process of the complex industrial process being modeled. In a cement mills scheduling example, these blocks typically consist of the MLD formulation for a "mill", "silo" and "electricity". In particular, the representation of the grinding section of a cement plant according to Fig.2 comprises one MLD block associated to the Cement Mill (one input, three outputs), one MLD block associated to the Cement Silo (one input) and one associated to the Electricity section (one input, two outputs). The optimal control problem is generated via adding a functional or cost MLD block, of which the unique output variable ycost becomes the objective function of the problem. The next step is to iteratively merge two interconnected MLD blocks to obtain a combined block, until only one block is left. The final MLD block, resulting from the merging process, will have as inputs all the independent inputs of the various MLD subsystems, i.e. the input "production" in the exemplary embodiment of Fig.2, and as output the unique output of the cost block, representing the objective function of the optimal control problem. The arrows in Fig.2 represent interconnections between the individual blocks or subsystems just as in Fig.l.