TWI809592B - Model prediction control system and method thereof - Google Patents

Abstract

A model prediction control (MPC) system and a method thereof are disclosed. The model prediction control system includes a model simulation unit, an observation unit and a proxy model prediction control unit. The model simulation unit is configured to simulate dynamic behaviors of a target object according to a behavior model of the target object and a plurality of control decision parameters, and correspondingly generating a plurality of real state parameters. The observation unit is configured to observe these control decision parameters and these real state parameters, and generate a plurality of estimated state parameters. The proxy model prediction control unit is configured to perform approximate model prediction control operations based on the estimated state parameters and a plurality of target parameters to replace model prediction control operations with higher computation complexity.

Description

Model Predictive Control System and Its Method

The present disclosure relates to a control system and a control method, in particular to a model predictive control system and method for predicting and controlling the behavior of an object.

The model predictive control mechanism is to establish the dynamic model of the target object according to the dynamic behavior of the target object, and predict the dynamic change of the target object according to the dynamic model, and then use the optimization algorithm to calculate the optimal solution of the control decision input value, so that The behavior of the object satisfies the intended purpose.

However, the optimization algorithm of the model predictive control mechanism involves a large-dimensional optimization solution process. When the dynamic change of the target object is fast or the dynamic model of the target object is complex, the optimization algorithm will consume a lot of calculations As a result, the optimal solution cannot be completed within the limited time decision-making cycle, which will cause the delay of the control decision of the target object, affect the performance of the target object and even cause accidents.

Therefore, technicians in related industries in this technical field are committed to improving the model predictive control mechanism, which can quickly calculate the suitable control target in various application scenarios (such as financial transactions, self-driving cars, drones, factory production lines, etc.) The input value of the control decision can avoid the delay of the control decision; moreover, the improved model predictive control mechanism can also greatly reduce the amount of calculation to reduce the hardware specification requirements.

The disclosure provides a model predictive control system, including a model simulation unit, a monitoring unit and an agent model predictive control unit. The model simulation unit is used for simulating the dynamic behavior of the target according to the behavior model of the target and a plurality of control decision parameters of the previous state, and correspondingly generates a plurality of actual state parameters. The monitoring unit is configured to monitor the control decision parameters of the previous state and the actual state parameters, and generate a plurality of estimated state parameters according to the control decision parameters of the previous state and the actual state parameters. The agent model predictive control unit is used to perform approximate model predictive control operations according to the estimated state parameters and multiple target parameters, so as to obtain multiple control decision parameters of the current state, and return the control decision parameters of the current state So far the model simulates the unit. Wherein, the proxy model predictive control unit executes the approximate model predictive control operation to replace the model predictive control operation with high computational complexity.

The disclosure also provides a model predictive control method, including the following steps. Create a behavioral model of the target. The dynamic behavior of the target is simulated according to the behavior model of the target and multiple control decision parameters of the previous state. Correspondingly generate a plurality of actual state parameters. Such previous state control decision parameters and such actual state parameters are monitored. A plurality of estimated state parameters are generated according to the control decision parameters of the previous state and the actual state parameters. Set up the surrogate model predictive control unit. Based on the estimated state parameters and multiple target parameters, the proxy model predictive control unit performs approximate model predictive control operations to replace model predictive control operations with high computational complexity. And, return the control decision parameters of these current states to the model simulation unit.

By reading the following drawings, detailed descriptions and scope of patent application, Other aspects and advantages of the present disclosure can be seen.

1000: Model Predictive Control Systems

100:Model simulation unit

200: Monitoring unit

300: Optimizing Model Predictive Control Units

400: Proxy Model Predictive Control Unit

500: update unit

700: target

2000: Factory production line

u1, u2, u3: control decision parameters

y1, y2: actual state parameters

x1, x2: Estimated state parameters

S: set of target parameters

s1~sN: target parameters

track1: target movement track

track2: actual moving track

u1_mean, u2_mean: average value

u1_var, u2_var: variation value

2100: Feed amount

2200: steam volume

2300: Reboiling temperature or pressure

S110~S180: steps

FIG. 1 is a block diagram of a model predictive control system according to an embodiment of the present disclosure.

Fig. 2 is a schematic diagram of an embodiment of the application of the model predictive control system disclosed in the present disclosure to a self-driving car.

FIG. 3 is a schematic diagram of an agent model predictive control unit performing model predictive control calculations according to the present disclosure.

FIG. 4 is a schematic diagram of an embodiment of applying the model predictive control system of the present disclosure to a factory production line.

FIG. 5 is a flowchart of a model predictive control method according to an embodiment of the present disclosure.

The technical terms in this specification refer to the customary terms in this technical field. If some terms are explained or defined in this specification, the explanations or definitions of these terms shall prevail. Each embodiment of the disclosure has one or more technical features. On the premise of possible implementation, those skilled in the art may selectively implement some or all of the technical features in any embodiment, or selectively combine some or all of the technical features in these embodiments.

FIG. 1 is a block diagram of a model prediction control (MPC) system 1000 according to an embodiment of the present disclosure. Please refer to Fig. 1, the model predictive control system 1000 includes a model simulation (model simulation) unit 100, a monitoring (observation) unit 200, an optimal model predictive control (optimal model prediction control) unit 300, proxy model prediction control (proxy model prediction control) unit 400, and update unit 500.

The model simulation unit 100 can establish a behavior model of the object 700 to simulate the dynamic behavior of the object 700 . In one example, the target object 700 (or may be referred to as "target object", "operating target", generally refers to the control target for which the model predictive control system 1000 performs predictive control) is, for example, a self-driving car, and the model simulation unit 100 can establish a self-driving car The behavioral model of the self-driving car simulates the dynamic behavior of the "vehicle speed", "traveling direction" and so on. Among them, the control object associated with the behavior "vehicle speed" of the self-driving car is the "accelerator" of the self-driving car, and the control object associated with the "direction of travel" is the "steering wheel" of the self-driving car. In operation, the model simulation unit 100 can receive a plurality of control decision parameters, one of the control decision parameters u1 is, for example, "throttle size", and its associated control object is "throttle". Simulate controlling the "accelerator" of the self-driving car to adjust the behavior of the self-driving car "vehicle speed". Moreover, the model simulation unit 100 can simulate and calculate the actual measurement of the "vehicle speed", that is, the actual state parameter y1 of the "vehicle speed"; in other words, the actual state parameter y1 represents the "actual vehicle speed" of the self-driving car.

On the other hand, another control decision parameter u2 of the model simulation unit 100 is, for example, the "steering wheel rotation amount", and its associated control object is the "steering wheel"; the model simulation unit 100 can simulate and control the "steering wheel" of the self-driving car according to the control decision parameter u2 ” to adjust the self-driving car’s behavior “direction of travel”. Moreover, the model simulation unit 100 can simulate and calculate the actual measurement value of the "traveling direction", that is, the actual state parameter y2; in other words, the actual state parameter y2 represents the "actual traveling direction" of the self-driving car.

To sum up, the model simulation unit 100 can simulate and control the control object "throttle" of the self-driving car according to the control decision parameter u1 "throttle size" to adjust the behavior of the self-driving car "vehicle speed", and the model simulation unit 100 correspondingly outputs the actual state parameter y1 "Actual Speed". On the other hand, the model simulation unit 100 can simulate and control the control object "steering wheel" of the self-driving car according to the control decision parameter u2 "steering wheel rotation amount" to adjust the behavior "traveling direction" of the self-driving car, and the model simulation unit 100 outputs the actual state parameters correspondingly y2 "actual direction of travel".

The monitoring unit 200 is connected to the model simulation unit 100. The monitoring unit 200 can receive actual state parameters y1, y2 from the model simulation unit 100, and generate estimated state parameters x1, x2 according to the control decision parameters u1, u2. Among them, the estimated state parameter x1 represents the "estimated speed" of the self-driving car, and the estimated state parameter x2 represents the "estimated direction of travel" of the self-driving car. More specifically, the monitoring unit 200 can receive the actual state parameters y1(k) and y2(k) of the current state (or called: the current iteration (iteration), ie, the kth iteration), and receive the previous The control decision parameters u1(k-1), u2(k-1) of the state (or called: the previous iteration, ie, the (k-1)th iteration). The monitoring unit 200 generates the estimated state parameters of the current iteration according to the actual state parameters y1(k) and y2(k) of the current iteration and the control decision parameters u1(k-1) and u2(k-1) of the previous iteration x1(k), x2(k).

Moreover, the optimized model predictive control unit 300 is connected to the monitoring unit 200 , and the optimized model predictive control unit 300 can receive estimated state parameters x1(k), x2(k) from the monitoring unit 200 . In addition, the optimization model predictive control unit 300 can receive a target parameter set S, and the target parameter set S can include a plurality of target parameters s1, s2, . . . , sN can be expressed as S={s1, s2, . . . , sN}. Please refer to FIG. 2, which shows a schematic diagram of an embodiment of the disclosed model predictive control system 1000 applied to a self-driving car; the target parameter set S can represent the target behavior that the user expects the target object 700 (self-driving car) to achieve, The target trajectory track1 that the self-driving car should achieve as shown in Figure 2 is the target parameter set S. Moreover, the coordinates of each point of the target moving track track1 of the self-driving car are the target parameters s1, s2, . . . , sN of the target parameter set S. According to the target parameter set S and estimated state parameters x1(k), x2(k), the optimized model predictive control unit 300 can perform model predictive control to calculate corresponding control decision parameters u1(k), u2(k). In other words, in order to make the behavior of the self-driving car meet the target parameter set S (that is, to make the actual moving track track2 of the self-driving car satisfy the target moving track track1), the optimization model predictive control unit 300 estimates the state parameters x1(k), x2( k) (that is, the "estimated vehicle speed" and "estimated direction of travel" of the self-driving car) calculate the corresponding control decision parameters u1(k), u2(k) (that is, the "accelerator size" and "steering wheel rotation amount" of the self-driving car ") to control the "accelerator" and "steering wheel" of the self-driving car respectively, so as to adjust the behavior of the self-driving car "vehicle speed" and "direction of travel". The adjusted "vehicle speed" and "traveling direction" can be reflected in the actual state parameters y1, y2 (ie, the "actual speed" and "actual traveling direction" of the self-driving car) through the simulation of the model simulation unit 100 .

In this embodiment, the optimized model predictive control unit 300 executes the model predictive control according to the optimized algorithm, so as to calculate the optimal solution (optimal solution). However, if considering other more control objects of self-driving cars (for example, gear, engine temperature degree, etc.), then the optimization model predictive control unit 300 must perform optimization operations on a large number of control decision parameters u1, u2, u3, ..., uM, which will consume a lot of computing time and cannot be used in The optimal solution of the control decision parameters u1, u2, u3, ..., uM can be calculated in a short decision cycle (for example, the decision cycle of self-driving car driving can be as short as about 0.1 second), so real-time control cannot be achieved , which may cause performance impairment or safety accidents of self-driving cars.

In view of this, the pre-established agent model predictive control unit 400 can be used to replace the optimized model predictive control unit 300 . The model predictive control unit 400 is used as an agent to perform model predictive control operations with lower computational complexity (thus the decision cycle is shorter) to replace the computational complexity performed by the optimized model predictive control unit 300 (thus the decision cycle is shorter) Long) operation of model predictive control. Moreover, although the surrogate model predictive control unit 400 does not perform optimal calculations, the surrogate model predictive control unit 400 can still obtain a result close to the optimal solution, and thus can effectively adjust the behavior of the control target. In this embodiment, the model of the agent model prediction control unit 400 can be pre-established in an off-line manner (ie, offline modeling), and the manner of establishing the model can include imitation learning and other machine learning methods. Also, refer to the schematic diagram of the agent model predictive control unit 400 shown in FIG. 3 performing model predictive control operations; the agent model predictive control unit 400 can operate according to a random program to perform model predictive control, for example, according to a Gaussian process (Gaussian Process, GP ) to perform model predictive control. Still taking the two control decision parameters u1, u2 of the self-driving car (that is, the "accelerator size" and "steering wheel rotation" of the self-driving car) as an example, the agent model prediction control unit 400 estimates the state parameters x1, x2 And the target parameter set S executes the operation of the Gaussian program to calculate the respective mean values u1_mean, u2_mean and variation values u1_var, u2_var of the control decision parameters u1, u2. In one example, the control decision parameters u1 and u2 are output with mean values u1_mean and u2_mean (ie, control decision parameter u1=mean value u1_mean, control decision parameter u2=mean value u2_mean). On the other hand, the variation values u1_var and u2_var of the control decision parameters u1 and u2 are used as the model reliability index rindex to determine whether to update the proxy model predictive control unit 400 .

Continuing from the above, the update unit 500 can set the threshold t1 of the model reliability index rindex, and compare the model reliability index rindex with the threshold t1. When the model reliability index rindex is greater than the threshold value t1 (that is, the variation values u1_var and u2_var of the control decision parameters u1 and u2 are greater than the threshold value t1), it means that the reliability of the currently used proxy model predictive control unit 400 is low and needs to be carried out. Adaptive update. In one example, various operation data (such as the operation data related to the "vehicle speed" and "traveling direction" of the self-driving car) can be collected online (on-line), and these operation data can be used to predict the control unit of the agent model 400 for adaptive updates.

From the above, the model of the agent model predictive control unit 400 can be pre-established offline (ie, offline modeling), and the agent model predictive control unit 400 can be adaptively updated online (ie, online update). Moreover, when the agent model predictive control unit 400 is performing adaptive updating, the agent model predictive control unit 400 cannot calculate the control decision parameters u1, u2 in real time. Optimized model predictions control (that is, model predictive control with high computational complexity (and therefore long decision-making period)) to calculate the control decision parameters u1, u2.

Figure 4 shows a schematic diagram of an embodiment of the model predictive control system 1000 disclosed herein applied to a factory production line 2000. As shown in Figure 4, the object to be controlled is the "distillation tower" in the factory production line 2000, and the model The control objects of the predictive control system 1000 include the control variables of each PID controller set point in the entire distillation column system, such as "feed (feed) amount" 2100, "steam (steam) amount" 2200, "reboiler (reboiler) ) temperature or pressure" 2300, etc. The control decision parameters u1, u2, u3 of the model predictive control system 1000 respectively correspond to the PID controller set points of the above-mentioned control objects, for example: the control decision parameter u1 is the PID controller set point of the “feed amount” 2100 . The control decision parameter u2 is the set point of the PID controller of the “steam amount” 2200 . Moreover, the control decision parameter u3 is the set point of the PID controller of the “reboiling temperature or pressure” 2300 .

In addition, the target parameter s1 of the target "distillation tower" is, for example, "the output per unit time of the distillate": in order to achieve the output per unit time desired by the user, the proxy model predictive control of the model predictive control system 1000 can be used The unit 400 performs approximate model predictive control to calculate the control decision parameters u1, u2, u3 that are close to the optimal solution, and import the control decision parameters u1, u2, u3 into the entire distillation column system in the actual factory production line 2000 The corresponding PID controller set point of . The decision-making cycle of the above-mentioned factory production line 2000 (taking the chemical factory of the distillation tower system as an example) is about 15 seconds to 30 seconds, and the agent model predictive control unit 400 can be used to execute relatively fast Approximate model predictive control can calculate the control decision parameters u1, u2, u3 that are close to the optimal solution within a decision period of 15 seconds to 30 seconds.

In addition, in the factory production line 2000, the proxy model predictive control system 1000 of this embodiment can also optionally cooperate with the control loop performance assessment (CLPA) system, the plant-wide disturbance source tracking (PDT) system, and controller parameter tuning ( PID tunner) system or system identification (modeling) system to control and adjust various variables in the factory production line 2000.

FIG. 5 is a flowchart of a model predictive control method according to an embodiment of the present disclosure. Please refer to FIG. 5 , the model predictive control method can be implemented in cooperation with the model predictive control system 1000 in FIG. 1 and the agent model predictive control unit 400 in FIG. 3 . First, in step S110 , a plurality of control decision parameters are received by the model simulation unit 100 , for example, two control decision parameters u1 "accelerator size" and control decision parameters u2 "steering wheel rotation amount" of the self-driving car are received. Moreover, the actual state parameter y1 "actual vehicle speed" and the actual state parameter y2 "actual traveling direction" are simulated and generated according to the control decision parameters u1 and u2 by the model simulation unit 100 .

Then, in step S120, the model simulation unit 100 transmits the actual state parameters y1(k), y2(k) of the current state to the monitoring unit 200, and the monitoring unit 200 further receives the control decision parameter u1(k- 1), u2(k-1). Accordingly, the monitoring unit 200 generates an estimated state parameter x1(k) "estimated vehicle speed" and an estimated The state parameter x2(k) "estimated direction of travel".

Then, in step S130, it is judged whether it is within the decision-making period of the model predictive control system 1000; taking the self-driving car as an example, the decision-making period is about 0.1 second. If it is still within the decision cycle, then execute step S135, and use the proxy model predictive control unit 400 to perform approximate model predictive control; wherein, the proxy model predictive control unit 400 receives the target parameter set S and the estimated state parameters x1(k), x2( k) Based on this, the average values u1_mean and u2_mean and the variation values u1_var and u2_var of the control decision parameters u1 and u2 are generated, and the variation values u1_var and u2_var are used as the model reliability index rindex. The generated control decision parameters u1 and u2 can be provided to the model simulation unit 100 in step S110 to perform simulation.

Then, in step S140, a threshold t1 of the model reliability index rindex is set; and it is judged whether the model reliability index rindex is greater than the threshold t1. If the model reliability index rindex is less than the threshold t1 (that is, the variation values u1_var and u2_var are less than the threshold t1), then execute step S150, and use the average values u1_mean and u2_mean as the final output control decision parameters u1 and u2 (that is, the control decision parameter u1=mean value u1_mean, control decision parameter u2=mean value u2_mean).

If the model reliability index rindex is greater than the threshold t1 (that is, the variation values u1_var and u2_var are greater than the threshold t1), it means that the proxy model predictive control unit 400 must be updated, and step S160 is performed to prepare for updating the proxy model predictive control unit 400. Then, step S170 is executed, and the optimal model predictive control unit 300 is used to calculate the control decision parameters u1 and u2 . Then, step S180 is executed to update the agent model predictive control unit 400; for example, an online adaptive update method can be used to collect various operational data related to the behavior model of the target object from the line, by The agent model predictive control unit 400 is updated from these data. After the update, step S150 can be executed to predict the control unit 400 to calculate the control decision parameters u1 and u2 through the updated agent model.

To sum up, the model predictive control system 1000 disclosed in this disclosure cooperates with the implementation of the model predictive control method. The agent model predictive control unit 400 is pre-established in the form of offline modeling imitation learning or machine learning; and the agent model predictive control unit is used 400 performs an approximate model predictive control operation to replace the optimized model predictive control unit 300 performing an optimal model predictive control operation with higher computational complexity (and therefore a longer decision period). Therefore, the agent model predictive control unit 400 can greatly reduce the calculation time, and can calculate the control decision parameters of the approximate optimal solution in a short decision cycle (especially for a short decision cycle of about 0.1 seconds for self-driving cars) . Moreover, the agent model predictive control unit 400 performs model predictive control operations according to the Gaussian program model to obtain the average value and variation value of the control decision parameters, and uses the variation value of the control decision parameters as the model reliability index. If the model reliability index is greater than the threshold value, online adaptive update is performed on the agent model predictive control unit 400 .

Although the present invention has been disclosed above in detail with preferred embodiments and examples, it should be understood that these examples are meant to be illustrative rather than limiting. It is expected that those skilled in the art can conceive various modifications and combinations, and the various modifications and combinations fall within the spirit of the present invention and the scope of the appended patent application.

1000: Model Predictive Control Systems

100:Model simulation unit

200: Monitoring unit

300: Optimizing Model Predictive Control Units

400: Proxy Model Predictive Control Unit

500: update unit

700: target

u1, u2: control decision parameters

y1, y2: actual state parameters

x1, x2: Estimated state parameters

S: set of target parameters

s1~sN: target parameters

Claims (20)

A model predictive control system, comprising: a model simulation unit, used to simulate a dynamic behavior of a target object according to a behavior model of a target object and a plurality of control decision parameters of the previous state, and correspondingly generate a plurality of actual states Parameters; a monitoring unit for monitoring the control decision parameters of the previous state and the actual state parameters, and generating a plurality of estimated state parameters according to the control decision parameters of the previous state and the actual state parameters; and An agent model predictive control unit, which is used to perform an approximate model predictive control operation according to the estimated state parameters and a plurality of target parameters, so as to obtain a plurality of control decision parameters of the current state, and convert the control decision parameters of the current state and sent back to the model simulation unit; wherein, the proxy model predictive control unit executes the approximate model predictive control operation to replace a model predictive control operation with higher computational complexity. The model predictive control system as in Claim 1, wherein the target object has a plurality of control objects, and these control objects are simulated and controlled according to the control decision parameters, so that the actual state parameters of the target object satisfy the target parameter. The model predictive control system according to claim 1, wherein the proxy model predictive control unit executes a random program operation to perform the approximate model predictive control operation. The model predictive control system of claim 3, wherein the stochastic program operation performed by the agent model predictive control unit is a Gaussian program operation, and the agent model predictive control unit executes the Gaussian program operation to obtain the respective control decision parameters mean and variance. The model predictive control system as claimed in claim 4, wherein the average values are used as the final output control decision parameters, and the variation values are used as a model reliability index. For example, the model predictive control system of claim item 5 further includes: an updating unit for judging whether the model reliability index is greater than a threshold value, and if the model reliability index is greater than the threshold value, the proxy model predictive control unit performs renew. Such as the model predictive control system of claim 6, further comprising: an optimized model predictive control unit, when the proxy model predictive control unit is updated, the optimized model predictive control unit is used to perform the calculation with high complexity The model predictive control operation to obtain the control decision parameters satisfying the optimal solution. As the model predictive control system of claim 6, wherein the agent model The predictive control unit is updated in an online adaptive manner. The model predictive control system according to claim 1, wherein the agent model predictive control unit pre-establishes the behavior model of the target in an offline manner. The model predictive control system according to claim 9, wherein the agent model predictive control unit pre-establishes the behavior model by means of imitation learning or machine learning. A model predictive control method, comprising: establishing a behavior model of a target object; simulating a dynamic behavior of the target object according to the behavior model of the target object and a plurality of control decision parameters of a previous state; correspondingly generating a plurality of actual State parameters; monitoring the control decision parameters of the previous state and the actual state parameters; generating a plurality of estimated state parameters according to the control decision parameters of the previous state and the actual state parameters; setting an agent model predictive control unit ; according to the estimated state parameters and a plurality of target parameters, an approximate model predictive control operation is performed by the agent model predictive control unit to replace a model predictive control operation with higher computational complexity; and the current state of these The control decision parameters are sent back to a model simulation unit. The model predictive control method according to claim 11, wherein the target object has a plurality of control objects, and the model predictive control method includes: simulating and controlling the control objects according to the control decision parameters, so that the actual The state parameters satisfy these target parameters. The model predictive control method according to claim 11, wherein the proxy model predictive control unit executes a random program operation to execute the approximate model predictive control operation. The model predictive control method of claim 13, wherein the stochastic program operation performed by the agent model predictive control unit is a Gaussian program operation, and the model predictive control method includes: executing the Gaussian program operation by the model predictive control unit, In order to obtain the average value and variation value of these control decision parameters. The model predictive control method as claimed in claim 14 further includes: using the average values as the final output control decision parameters; and using the variation values as a model reliability index. For example, the model predictive control method of claim 15 further includes: judging whether the model reliability index is greater than a threshold; and If the model reliability index is greater than the threshold, the agent model predictive control unit updates. The model predictive control method as claimed in claim 16 further includes: setting an optimized model predictive control unit; and when the agent model predictive control unit is updated, the optimized model predictive control unit executes the computational complexity Higher model predictive control operations to obtain the control decision parameters satisfying the optimal solution. The model predictive control method according to claim 16, wherein the agent model predictive control unit is updated in an online adaptive manner. The model predictive control method according to claim 11, wherein the behavior model of the target is pre-established by the agent model predictive control unit in an off-line manner. The model predictive control method according to claim 19, wherein the behavior model is pre-established by the agent model predictive control unit by means of imitation learning or machine learning.

Images