A data-driven optimization control method for complex industrial processes with unknown operation processes

Abstract

The present invention discloses a data-driven optimization control method for complex industrial process operation with unknown operation process, including the following steps: modeling the dynamic model of the complex industrial process equipment layer and operation layer; based on the singular perturbation theory, decomposing into a pair of optimal regulation problems of fast subsystems and optimal set point tracking problems of slow subsystems; designing controllers for the fast and slow subsystems respectively, and designing a data / model hybrid driven composite tracking controller based on the incremental system to control the steady-state output tracking error to converge to zero. The present invention solves the problem that the slow dynamics of the operation process are difficult to model, realizes the optimal output regulation control of the complex industrial process equipment layer and the optimal fixed-point output tracking control of the operation layer, and the decomposition and combination controller design method avoids the potential dimensionality disaster and ill-conditioned numerical problems in the control design process.

Description

Technical Field

[0001] The present invention relates to the field of process industry operation optimization control, and in particular to a data-driven operation optimization control method for a complex industrial process with unknown operation process. Background Art

[0002] In the actual process industry, unit equipment (such as grinders, ovens, and concentrators) run fast on a fast time scale of seconds, while operating indicators (such as energy efficiency, product quality, and resource utilization) change slowly on a slow time scale of hours, with multi-time scale characteristics. Due to the high complexity and uncertainty of variables, for some flotation processes with unclear dynamic mechanisms, it becomes increasingly difficult or even impossible to establish an accurate system mathematical model and obtain complete system dynamics information. At this time, the use of model-based control methods will lose its feasibility to a certain extent. In addition, many industrial processes generate and store a large amount of process data at every moment of every day, which contains all the valuable status information of process operations and equipment. It will be very important to use this data to directly design controllers online / offline, predict and evaluate system status, evaluate performance, make decisions, and even diagnose faults. Summary of the invention

[0003] Purpose of the invention: The purpose of the present invention is to provide a data-driven optimization control method for complex industrial processes with unknown operation processes, which solves the problem of slow dynamics being difficult to model in the operation process, realizes the optimal output regulation control at the equipment level of complex industrial processes and the optimal fixed-point output tracking control at the operation level, and avoids the potential dimensionality disaster and ill-conditioned numerical problems in the control design process.

[0004] Technical solution: A data-driven optimization control method for complex industrial processes with unknown operation processes, comprising the following steps:

[0005] S1, based on the equipment layer and the operation layer, establish a complex industrial process dynamics model; then use Euler discretization to establish a model of complex industrial process dynamics with respect to discrete time variables n;

[0006] S2, based on singular perturbation theory, decomposes the operation optimization problem of dual-time-scale industrial systems into a pair of optimal regulation problems of fast subsystems and optimal set-point tracking problems of slow subsystems;

[0007] S3, redefine the standard linear quadratic tracking solution using convex optimization to obtain an equivalent convex optimization problem;

[0008] S4, design controllers for the fast subsystem and the slow subsystem respectively, and design a data / model hybrid driven composite tracking controller based on the incremental system to control the steady-state output tracking error to converge to zero.

[0009] Further, the establishment of the kinetic model includes the following steps:

[0010] S21, the dynamic model of the equipment unit is as follows:

[0011]

[0012] in, represents the state vector of the device, represents the control input, represents Euclidean space, n z 、n u Represents the dimension of the vector;

[0013] S22, the dynamic model of the motion process is as follows:

[0014]

[0015] r(t)=Cx(t)

[0016] Among them, A x , B x and C is the system parameter matrix, Indicates the operation process status. Indicates the operation index, n x 、n r Represents the dimension of the vector;

[0017] S23, introduce a parameter 0<ε<<1 to describe the time scale difference between the fast variable z(t) and the slow variable x(t), define z(t)=εy(t), εA z =A y , εB z =B x , build complex industrial process dynamics models:

[0018]

[0019] r(t) = Cx(t);

[0020] S24, select the sampling period εh, where h is a given time interval, and use Euler discretization to establish a model of complex industrial process dynamics with respect to discrete time variable n:

[0021]

[0022] r(n)=Cx(n)

[0023] in, and is the identity matrix of appropriate dimension.

[0024] Furthermore, the decomposition of the operation optimization problem of the dual-time-scale industrial system comprises the following steps:

[0025] S31, according to the singular perturbation theory, define y(n) = y f (n)+y s (n), u(n)=u f (n)+u s (n), where y f (n),u f (n) is the fast component of the system variable, y s (n),u s (n) is the slow component of the system variable; when the fast component reaches a steady state, y f (n),u f (n) is ignored, and the slow component y s (n) Remain constant over fast time intervals;

[0026] S32, build the slow subsystem model:

[0027]

[0028] Introduce a slow time scale k satisfying n = k[1 / ε], k = 0, 1, 2, ..., [1 / ε] represents the largest integer less than 1 / ε, define 1 / ε and rewrite the slow subsystem model as:

[0029]

[0030] Let ε→0, and we get the continuous-time form of the slow subsystem:

[0031]

[0032] The slow component u of the control input s (n) remains constant within the sampling interval [k[1 / ε],(k+1)[1 / ε]) by discretizing the continuous-time slow subsystem on the slow time scale:

[0033] x s (k+1)=A s x s (k)+B s u s (k)

[0034] r s (k) = Cx s (k)

[0035] in Note x s (k) = x s (k[1 / ε])u s (k) = us (k[1 / ε]);

[0036] S33, build fast subsystem model:

[0037] y f (n+1)=A f y f (n)+B f u f (n)

[0038] Among them, the fast subsystem parameter matrix A f =I nz +hA y , B f =hB y ;

[0039] S34, define the increments between adjacent states Δx(n)=x(n+1)-x(n), Δy(n)=y(n+1)-y(n), and the tracking error is expressed as e(n+1)=e(n)+CΔx(n), thereby defining new variables: in, Represents the transpose of a matrix;

[0040] Construct an operation optimization problem to design the optimal controller u for a complex industrial process. The standard linear quadratic regulation problem is as follows:

[0041]

[0042]

[0043] Among them, min means minimization, J represents the cost function, st represents obedience, and the weight matrix

[0044] S35, construct the optimal regulation problem of the fast subsystem and design the optimal controller u for the fast subsystem f (n), the standard linear quadratic regulation problem is as follows:

[0045]

[0046] stΔy f (n+1)=A f Δy f (n)+B f Δu f (n)

[0047] Among them, J f represents the cost function of the fast subsystem, the weight matrix

[0048] S36, let the tracking error es (k) = r s (k)-r * (k), where r * (k) is the target setting value, state increment Δx s (k) = x s (k+1)-x s (k), Δe s (k) = e s (k+1)-e s (k) and Get e s (k+1)=e s (k)+CΔx s (k);

[0049] Define the augmentation vector Construct an augmented system:

[0050]

[0051] Among them, the system parameter matrix

[0052] Construct the optimal fixed-point tracking control problem of the augmented slow subsystem and design the optimal controller u for the slow subsystem s (k), the standard linear quadratic regulation problem is as follows:

[0053]

[0054] Among them, J s Represents the cost function of the slow subsystem, the weight matrix

[0055] Further, the calculation of the equivalent convex optimization problem includes the following steps:

[0056] S41, construct an equivalent convex optimization problem for the optimal regulation problem of a fast subsystem;

[0057] S42, construct an equivalent convex optimization problem for the optimal fixed-point tracking problem of the slow subsystem.

[0058] Furthermore, the construction of an equivalent convex optimization problem of the optimal regulation problem of the fast subsystem includes the following steps:

[0059] S51, suppose that for all discrete time variables n, The following formula is satisfied:

[0060]

[0061] The cumulative sum of both sides of the above formula from 0 to n is:

[0062]

[0063] The lower bound of the cost index is

[0064] S52, construct the inequality form of the optimal regulation problem of the fast subsystem:

[0065]

[0066] in,

[0067] S53, we get the equivalent optimization problem of the optimal regulation problem of the fast subsystem:

[0068]

[0069] Among them, max means maximization.

[0070] Furthermore, the construction of an equivalent convex optimization problem for the optimal fixed-point tracking problem of the slow subsystem includes the following steps:

[0071] S61, let for all slow time scales k, Satisfies the following formula:

[0072]

[0073] The cumulative sum of both sides of the above formula from 0 to k is:

[0074]

[0075] The lower bound of the cost index is

[0076] S62, construct the inequality form of the optimal fixed-point tracking problem of the slow subsystem:

[0077]

[0078] in,

[0079] S63, we get the equivalent convex optimization problem of the optimal fixed-point tracking problem of the slow subsystem:

[0080]

[0081] S64, using Schur’s lemma to consider and Equivalently, satisfy K s is the controller gain of the slow subsystem, and the closed-loop form of the equivalent convex optimization problem of the optimal fixed-point tracking problem of the slow subsystem is obtained:

[0082]

[0083] Further, the step S4 comprises the following steps:

[0084] S71, Design of fast subsystem controller, Δu f (n) = K f Δy f (n), where K f represents the controller gain of the fast subsystem;

[0085] S72, Design of Slow Subsystem Controller, Among them, K s represents the slow subsystem controller gain;

[0086] Let K s =[K s1 K s2 ],in but It can be written as: Cumulatively sum the equation from 0 to k-1, let u s (0) = K s1 x s (0), and the optimal control strategy is obtained:

[0087]

[0088] S73, data expression of the slow subsystem controller, runs the original system and collects a set of system I / O data groups with a sequence length of N:

[0089] U 0,N :=[u(0)u(1)…u(N-1)]

[0090] X 0,N :=[x(0)x(1)…x(N-1)]

[0091] X 1,N :=[x(1)x(2)…x(N)]

[0092] Among them, U 0,N Indicates that N control input data are continuously collected from time 0, X 0,N Indicates that the data of N state variables are continuously collected from time 0, X 1,N Indicates that the data of N state variables are continuously collected from time 1, and the definition get:

[0093]

[0094] Using the original system data, the optimal controller gain of the slow subsystem is calculated as matrix R s , is the solution to the following convex optimization problem:

[0095]

[0096] S74, build a composite controller:

[0097] u comp (n) = u f (n)+u s (n)

[0098] Among them, u s (n) remains unchanged in the interval [k[1 / ε],(k+1)[1 / ε]) and is s (k) Update at n=k[1 / ε].

[0099] Compared with the prior art, the present invention has the following significant effects: compared with the controller design method without singular perturbation decomposition, the controller design of the reduced-order subsystem avoids potential numerical stiffness problems; compared with the popular discounted tracking control method, a cost function without discount factors is constructed by using system increments, and the obtained controller can force the steady-state tracking error to converge to zero, thereby improving the tracking accuracy; compared with the popular reinforcement learning method, through convex duality, the method for solving the optimal solution of the optimal fixed-point tracking problem with infinite time domain is equivalently converted into a convex optimization problem that requires less data and is easy to solve; the present invention realizes the design of data-driven controllers when the operating process is unknown in complex industrial processes, and avoids potential numerical stiffness problems. BRIEF DESCRIPTION OF THE DRAWINGS

[0100] Figure 1 It is a design flow chart of the present invention;

[0101] Figure 2 It is a structural schematic diagram of the simulated hematite mixed separation and thickening process of the present invention;

[0102] FIG. 3( a ) is a diagram showing the operating status data of a virtual slow subsystem and a real original system obtained by simulation according to an embodiment of the present invention.

[0103] Figure 3(b) is a diagram showing the difference in operating status data between the real original system and the virtual slow subsystem;

[0104] FIG4(a) is a fixed-point output tracking trajectory diagram obtained by simulation according to an embodiment of the present invention.

[0105] FIG4( b ) is a diagram showing the optimal control input and its fast and slow component trajectories obtained by simulation according to an embodiment of the present invention.

[0106] FIG4( c ) is a diagram of the optimal controlled output and its fast and slow component trajectories obtained by simulation according to an embodiment of the present invention; DETAILED DESCRIPTION

[0107] The present invention is further described in detail below in conjunction with the accompanying drawings and specific implementation methods.

[0108] A data-driven optimization control method for complex industrial processes with unknown operation processes, such as Figure 1 As shown, including:

[0109] Step 1: Establish the dynamic model of the equipment layer and operation layer of the complex industrial process;

[0110] The establishment of the kinetic model includes the following steps:

[0111] Step 21, Modeling Equipment Unit Dynamics:

[0112]

[0113] Among them, z(t) represents the state vector of the device, u(t) represents the control input, and A y , B y is the system parameter matrix;

[0114] Step 22, modeling motion process dynamics:

[0115]

[0116] r(t)=Cx(t) (3)

[0117] Among them, x(t) represents the operation process state, r(t) represents the operation index, and A x , B x and C is the system parameter matrix.

[0118] Step 23, introduce a small parameter 0<ε<<1 to describe the time scale difference between the fast variable z(t) and the slow variable x(t), define z(t)=εy(t), εA z =A y , εB z =B x , build complex industrial process dynamics models:

[0119]

[0120] r(t) = Cx(t);

[0121] Step 24, select a sampling period εh, where h is a given time interval, and use Euler discretization to establish a model of complex industrial process dynamics with respect to discrete time variable n:

[0122]

[0123] r(n)=Cx(n) (8)

[0124] in, and is the identity matrix of appropriate dimension.

[0125] Step 2, according to the established dynamic model and based on singular perturbation theory, the operation optimization problem of the dual-time-scale industrial system is decomposed into a pair of optimal regulation problems of the fast subsystem and the optimal set point tracking problem of the slow subsystem.

[0126] The decomposition of the operation optimization problem of the dual-time-scale industrial system includes the following steps:

[0127] Step 31, according to singular perturbation theory, define:

[0128] y(n)=y f (n)+y s (n) (9)

[0129] u(n)=u f (n)+u s (n) (10)

[0130] Among them, y f (n),u f (n) Fast component of system variable, y s (n),u s (n) is the slow component of the system variable; when the fast component reaches a steady state, y f (n),u f (n) can be ignored, and the slow component y s (n) Remain constant over fast time intervals;

[0131] Step 32, construct the slow subsystem model:

[0132]

[0133] Introduce a slow time scale k satisfying n = k[1 / ε], k = 0, 1, 2, ..., [1 / ε] represents the largest integer less than 1 / ε;

[0134] Define 1 / ε and rewrite the slow subsystem model as:

[0135]

[0136] Let ε→0, and we get the continuous-time form of the slow subsystem:

[0137]

[0138] The slow component u of the control input s (n) remains constant within the sampling interval [k[1 / ε],(k+1)[1 / ε]) by discretizing the continuous-time slow subsystem on the slow time scale:

[0139] x s (k+1)=A s x s (k)+B s u s (k) (14)

[0140] r s (k) = Cx s (k) (15)

[0141] in, Note x s (k) = x s (k[1 / ε]),u s (k) = u s (k[1 / ε]);

[0142] Step 33, build the fast subsystem model:

[0143] y f (n+1)=A f y f (n)+B f u f (n) (16)

[0144] in, B f =hB y ;

[0145] Step 34, define the increments between adjacent states Δx(n)=x(n+1)-x(n), Δy(n)=y(n+1)-y(n), and the tracking error can be expressed as e(n+1)=e(n)+CΔx(n);

[0146] Defining variables in, Represent the transpose of the matrix, construct the operation optimization problem, and design the optimal controller u for a complex industrial process:

[0147]

[0148] Among them, Δu(n) represents the control increment, min represents minimization, J represents the cost function, st represents compliance with, and the weight matrix

[0149] Step 35, construct the optimal adjustment problem of the fast subsystem:

[0150] Design the optimal controller u for the fast subsystem f (n), the standard linear quadratic regulation problem is as follows:

[0151]

[0152] stΔy f (n+1)=A f Δy f (n)+B f Δu f (n),

[0153] Among them, J f represents the cost function of the fast subsystem; weight matrix

[0154] Step 36, set the tracking error e s (k) = r s (k)-r * (k), state increment Δx s (k) = x s (k+1)-x s (k), Δe s (k) = e s (k+1)-e s (k) and Get s (k+1)=e s (k)+CΔx s (k);

[0155] definition Construct an augmented system:

[0156]

[0157] Among them, the system parameter matrix

[0158] Construct the optimal fixed-point tracking control problem of the augmented slow subsystem and design the optimal controller u for the slow subsystem s (k), the standard linear quadratic regulation problem is as follows:

[0159]

[0160] Among them, J s represents the cost function of the slow subsystem; weight matrix

[0161] Step 3: Redefine the standard linear quadratic tracking solution using convex optimization to obtain an equivalent convex optimization problem.

[0162] Step 41, constructing an equivalent convex optimization problem of the optimal regulation problem of the fast subsystem, including the following steps:

[0163] Step 411, suppose for all n, The formula satisfied is as follows:

[0164]

[0165] The sum of both sides of the equation from 0 to n is:

[0166]

[0167] The lower bound of the cost index is

[0168] Step 412, construct the inequality form of the optimal regulation problem of the fast subsystem:

[0169]

[0170] in,

[0171] Step 413, obtaining an equivalent optimization problem of the optimal regulation problem of the fast subsystem,

[0172]

[0173] Among them, max means maximization.

[0174] Step 42, constructing an equivalent convex optimization problem for the optimal fixed-point tracking problem of the slow subsystem, including the following steps:

[0175] Step 421, suppose for all k, The formula satisfied is

[0176]

[0177] The sum of both sides of the equation from 0 to k is:

[0178]

[0179] The lower bound of the cost index is

[0180] Step 422, construct the inequality form of the optimal fixed-point tracking problem of the slow subsystem:

[0181]

[0182] in,

[0183] Step 423, obtain the equivalent convex optimization problem of the optimal fixed-point tracking problem of the slow subsystem:

[0184]

[0185] Step 424, use Schur's lemma to simultaneously consider and Equivalently, satisfy K s The controller gain of the slow subsystem is given as the closed-loop form of the equivalent convex optimization problem of the optimal fixed-point tracking problem of the slow subsystem:

[0186]

[0187] Step 4, design controllers for the fast and slow subsystems respectively, and design a data / model hybrid driven composite tracking controller based on the incremental system to control the steady-state output tracking error to converge to zero, including the following steps:

[0188] Step 71, design the fast subsystem controller:

[0189] Δu f (n) = K f Δy f (n)(30)

[0190] Among them, K f represents the controller gain of the fast subsystem;

[0191] Step 72, design the slow subsystem controller:

[0192]

[0193] Among them, K s represents the slow subsystem controller gain;

[0194] Let K s =[K s1 K s2 ],in but It can be written as:

[0195]

[0196] For equation (32), the sum from 0 to k-1 is: s (0) = K s1 x s (0), and the optimal control strategy is obtained:

[0197]

[0198] Among them, k≥1.

[0199] Step 73, data expression of the slow subsystem controller, running the original system, collecting a set of system I / O data groups with a sequence length of N:

[0200] U 0,N :=[u(0) u(1) … u(N-1)] (34)

[0201] X 0,N :=[x(0) x(1) … x(N-1)] (35)

[0202] X 1,N :=[x(1) x(2) … x(N)] (36)

[0203] definition We can get:

[0204]

[0205] Using the original system data, the optimal controller gain of the slow subsystem can be calculated as matrix R s , is the solution to the following convex optimization problem:

[0206]

[0207]

[0208] Step 74, build the composite controller:

[0209] u comp (n) = u f (n)+u s (n) (41)

[0210] Among them, u s (n) remains unchanged in the interval [k[1 / ε],(k+1)[1 / ε]) and is s (k) Update at n=k[1 / ε].

[0211] The hematite mixed selection and thickening process is combined to verify the effectiveness of the data-driven operation optimization control method for complex industrial processes with unknown operation processes.

[0212] After the grinding process, the low-concentration slurry enters the thickener and is stirred with a rake to increase the concentration. By adjusting the frequency u of the underflow slurry pump, the underflow slurry flow y and the underflow concentration r are controlled to provide the flotation process with a qualified underflow slurry flow and underflow slurry concentration. Its structure is as follows: Figure 2 shown.

[0213] The linearized dynamics of the system:

[0214]

[0215] r(t)=x(t).

[0216] The actual production requires that the underflow concentration r be controlled within the required range by adjusting the underflow slurry pump frequency u, that is, Select the median value 33 as the expected underflow concentration value r * . Choose ε = 0.2, h = 1. Discretize the linearized dynamics:

[0217] x(n+1)=0.9886x(n)+0.0022y(n)

[0218] y(n+1)=0.864y(n)+2.6u(n)

[0219] r(n)=x(n).

[0220] Set the initial values ​​x(0) = 31.5 and y(0) = 35. Considering that the decomposed slow subsystem is a virtual system, it is impossible to measure x s (k). The measured underflow concentration x(n) can be collected at a fast time scale n.

[0221] Figure 3(a) , 3(b) describes the measured undercurrent concentration x(k) reconstructed on the slow time scale, x generated by the slow subsystem s (k) and the error between them x(k)-x s (k).

[0222] Let Q f =0.1, R f =1, we can get the optimal control gain K of the fast subsystem at the device layer f =-0.1710. Let Q s =1, R s =1, the optimal control gain K of the slow subsystem in the operating layer can be obtained s =[-2.8247 -0.7410].

[0223] Figure 4(a) , 4(b) 4(c) describes the bottom flow concentration r gradually converges to the set reference value r * ; Actual control input u and its decomposed input u f 、u s The trajectory of the underflow slurry velocity y, the transient value y f converges to zero, and y eventually reaches its quasi-steady-state value y s .

[0224] Simulation results show that the data-driven operation model of the complex industrial process optimization control method proposed in this invention is unknown. When the actual operation process is unknown, the designed data-driven controller can effectively realize the optimal control of the system.

Claims

1. A data-driven optimization control method for complex industrial processes with unknown operation processes, characterized in that: The following steps are involved: S1, building a complex industrial process dynamics model based on the equipment layer and operation layer; Then, using Euler discretization, a model of complex industrial process dynamics with respect to discrete time variable n is established; S2, based on singular perturbation theory, decomposes the operation optimization problem of dual-time-scale industrial systems into a pair of optimal regulation problems of fast subsystems and optimal set-point tracking problems of slow subsystems; S3, redefine the standard linear quadratic tracking solution using convex optimization to obtain an equivalent convex optimization problem; S4, design controllers for the fast subsystem and the slow subsystem respectively, and design a data / model hybrid driven composite tracking controller based on the incremental system to control the steady-state output tracking error to converge to zero; The establishment of the kinetic model comprises the following steps: S21, the dynamic model of the equipment unit is as follows: in, represents the state vector of the device, represents the control input, represents Euclidean space, n z 、n u They represent the vector dimensions respectively; S22, the dynamic model of the motion process is as follows: r(t)=Cx(t) Among them, A x , B x and C is the system parameter matrix, Indicates the operation process status. Indicates the operation index, n x 、n r Represents the dimension of the vector; S23, introduce a parameter 0<ε<<1 to describe the time scale difference between the fast variable z(t) and the slow variable x(t), define z(t)=εy(t), εA z =A y , εB z =B x , build complex industrial process dynamics models: r(t) = Cx(t); S24, select the sampling period εh, where h is a given time interval, and use Euler discretization to establish a model of complex industrial process dynamics with respect to discrete time variable n: r(n)=Cx(n) in, and is the identity matrix of appropriate dimension; The decomposition of the operation optimization problem of the dual-time-scale industrial system includes the following steps: S31, according to the singular perturbation theory, define y(n) = y f (n)+y s (n), u(n)=u f (n)+u s (n), where y f (n),u f (n) is the fast component of the system variable, y s (n),u s (n) is the slow component of the system variable; when the fast component reaches a steady state, y f (n),u f (n) is ignored, and the slow component y s (n) Remain constant over fast time intervals; S32, build the slow subsystem model: Introduce a slow time scale k satisfying n = k[1 / ε], k = 0, 1, 2, ..., [1 / ε] represents the largest integer less than 1 / ε, define 1 / ε and rewrite the slow subsystem model as: Let ε→0, and we get the continuous-time form of the slow subsystem: The slow component u of the control input s (n) remains constant within the sampling interval [k[1 / ε],(k+1)[1 / ε]) by discretizing the continuous-time slow subsystem on the slow time scale: x s (k+1)=A s x s (k)+B s u s (k) r s (k)=Cx s (k) Among them Let x s (k) = x s (k[1 / ε])u s (k) = u s (k[1 / ε]); S33, build fast subsystem model: y f (n+1)=A f y f (n)+B f u f (n) Among them, the fast subsystem parameter matrix B f =hB y ; S34, define the increments between adjacent states Δx(n)=x(n+1)-x(n), Δy(n)=y(n+1)-y(n), and the tracking error is expressed as e(n+1)=e(n)+CΔx(n), thereby defining new variables: in, Represents the transpose of a matrix; Construct an operation optimization problem to design the optimal controller u for a complex industrial process. The standard linear quadratic regulation problem is as follows: Among them, min means minimization, J represents the cost function, st represents obedience, and the weight matrix S35, construct the optimal regulation problem of the fast subsystem and design the optimal controller u for the fast subsystem f (n), the standard linear quadratic regulation problem is as follows: stΔy f (n+1)=A f Δy f (n)+B f Δu f (n) Among them, J f represents the cost function of the fast subsystem, the weight matrix S36, let the tracking error e s (k) = r s (k)-r * (k), where r * (k) is the target setting value, state increment Δx s (k) = x s (k+1)-x s (k), Δe s (k) = e s (k+1)-e s (k) and Get s (k+1)=e s (k)+CΔx s (k); Define the augmentation vector Construct an augmented system: Among them, the system parameter matrix Construct the optimal fixed-point tracking control problem of the augmented slow subsystem and design the optimal controller u for the slow subsystem s (k), the standard linear quadratic regulation problem is as follows: Among them, J s Represents the cost function of the slow subsystem, the weight matrix 2. The data-driven operation optimization control method for a complex industrial process with unknown operation process according to claim 1 is characterized in that: The calculation of the equivalent convex optimization problem includes the following steps: S41, construct an equivalent convex optimization problem for the optimal regulation problem of a fast subsystem; S42, construct an equivalent convex optimization problem for the optimal fixed-point tracking problem of the slow subsystem.

3. The data-driven operation optimization control method for a complex industrial process with unknown operation process according to claim 2 is characterized in that: The construction of the equivalent convex optimization problem of the optimal regulation problem of the fast subsystem includes the following steps: S51, suppose that for all discrete time variables n, Satisfies the following formula: The cumulative sum of both sides of the above formula from 0 to n is: The lower bound of the cost index is S52, construct the inequality form of the optimal regulation problem of the fast subsystem: in, S53, we get the equivalent optimization problem of the optimal regulation problem of the fast subsystem: s.t.Φ≥0, Among them, max means maximization.

4. The data-driven operation optimization control method for a complex industrial process with unknown operation process according to claim 2 is characterized in that: The construction of the equivalent convex optimization problem of the optimal fixed-point tracking problem of the slow subsystem includes the following steps: S61, let for all slow time scales k, Satisfies the following formula: The cumulative sum of both sides of the above formula from 0 to k is: The lower bound of the cost index is S62, construct the inequality form of the optimal fixed-point tracking problem of the slow subsystem: in, S63, we get the equivalent convex optimization problem of the optimal fixed-point tracking problem of the slow subsystem: s.t.Ψ≥0, S64, using Schur’s lemma to consider and Equivalently, satisfy K s is the controller gain of the slow subsystem, and the closed-loop form of the equivalent convex optimization problem of the optimal fixed-point tracking problem of the slow subsystem is obtained:

5. According to claim 1, the data-driven operation process unknown complex industrial process operation optimization control method is characterized by: The step S4 comprises the following steps: S71, Design of fast subsystem controller, Δu f (n) = K f Δy f (n), where K f represents the controller gain of the fast subsystem; S72, Design of Slow Subsystem Controller, Among them, K s represents the slow subsystem controller gain; Let K s =[K s1 K s2 ],in represents Euclidean space, then It can be written as: Cumulatively sum the equation from 0 to k-1, let u s (0) = K s1 x s (0), and the optimal control strategy is obtained: S73, data expression of the slow subsystem controller, runs the original system and collects a set of system I / O data with a sequence length of N: U 0,N :=[u(0) u(1) … u(N-1)] X 0,N :=[x(0) x(1) … x(N-1)] X 1,N :=[x(1) x(2) … x(N)] Among them, U 0,N Indicates that N control input data are continuously collected from time 0, X 0,N Indicates that the data of N state variables are continuously collected from time 0, X 1,N Indicates that the data of N state variables are continuously collected from time 1, and the definition get: Using the original system data, the optimal controller gain of the slow subsystem is calculated as matrix is the solution to the following convex optimization problem: S74, build a composite controller: u comp (n)=u f (n)+u s (n) Among them, u s (n) remains unchanged in the interval [k[1 / ε],(k+1)[1 / ε]) and is s (k) Update at n=k[1 / ε].

Images