A disturbance rejection learning control method based on a dynamic combination model

Abstract

The application discloses an anti-interference learning control method based on a dynamic combination model, which is applied to a control system without model or weak model dependence. The method is characterized in that a weight system is added to the output of a controller, different order process local models of a controlled system are used as sub-models, a weight evaluation and updating mechanism is set, the connection weights of the sub-models are dynamically determined, the sub-models are combined into a dynamic combination model of the controlled system with variable order according to the connection weights, the comprehensive model parameters of the controlled system are calculated according to the dynamic combination model, and the anti-interference learning control of the controlled system is realized. The model structure of the dynamic combination model, the system order updating mode and the weight adjustment method are given. The dynamic combination model overcomes the problem of dynamic change of the system model in operation, realizes the modeling of the fractional order controlled system, and expands the application range of the dynamic combination model.

Description

Technical Field

[0001] This invention relates to the technical field of disturbance rejection learning control, and in particular to a disturbance rejection learning control method based on a dynamic combination model. Background Technology

[0002] Disturbance rejection control (DRC) and learning control are two very important control methods. These methods obtain real-time estimates of disturbances by learning them and then compensate at the system input through control, thus achieving disturbance rejection. Existing typical control methods include Active Disturbance Rejection Control (ADRC), Model-Free Control (MFC), and Model-Free Adaptive Control (MFAC), all of which have been widely studied and applied. These methods are based on local models or ultra-local models. The structure of these models can be obtained through mechanistic analysis, dynamic linearization, or system identification. However, once the model structure is determined, the model order remains unchanged, making it unsuitable for controlling systems with varying orders. Furthermore, the estimation of uncertainties in the model often relies on Extended State Observer (ESO), differential algebra, and system identification methods, failing to fully utilize historical data. This results in low historical data utilization; measurement data is often discarded after only one use in the ESO, leading to slow convergence of uncertainty estimation and frequent initial shocks.

[0003] In particular, for temperature control of chemical reactors, due to the complexity of their chemical reaction processes and the occurrence of phase changes, traditional control methods cannot achieve temperature control throughout the entire working state. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of existing technologies and propose an anti-disturbance learning control method based on a dynamic combinatorial model. The invention provides the model structure, system order update method, and weight adjustment method of the dynamic combinatorial model. It overcomes the problem of dynamic order changes in the system model during operation by using the dynamic combinatorial model. The iterative learning identification method in the learning time domain can improve the efficiency of model parameter estimation. The evaluation mechanism in the evaluation time domain estimates the connection weights of the dynamic combinatorial model, realizing the modeling function of fractional-order systems. This expands the applicability of the dynamic combinatorial model and enables anti-disturbance learning control of the controlled object.

[0005] The objective of this invention is achieved through the following technical solution: a disturbance rejection learning control method based on a dynamic combined model, applied to model-free or weakly model-dependent control systems. This method adds weighted systems to the controller output, uses process local models of different orders of the controlled system as sub-models, sets a weight evaluation and update mechanism, dynamically determines the connection weights of each sub-model, and combines the sub-models according to the connection weights to form a dynamic combined model of varying order for the controlled system. The comprehensive model parameters of the controlled system are calculated based on the dynamic combined model to achieve disturbance rejection learning control of the controlled system. This method includes the following steps:

[0006] S1. The process local models of different orders of the controlled system are used as sub-models, and the sub-models are combined into a dynamic combination model of the controlled system according to the connection weights.

[0007] S2. Establish the learning time domain and evaluation time domain in the time-series historical data of the controlled system;

[0008] S3. Iterative learning of dynamic combination models of different orders and process gains is performed in the learning time domain to estimate the parameters of the controlled system and the estimation error.

[0009] S4. Compare the estimation errors in the evaluation time domain, select the sub-model with the smaller estimation error, obtain the gain of the sub-model through the estimation error, set up a weight evaluation and update mechanism, determine the connection weights between each sub-model, and update the dynamic combination model according to the connection weights.

[0010] S5. Calculate the control components of the sub-models in the dynamic combination model, determine the final control quantity through the weight system, and output the final control quantity to the controlled system.

[0011] Furthermore, step S1 includes:

[0012] The process local model is adopted as shown in formula (1). Indicates the controlled system.

[0013] (1)

[0014] In formula (1), It is the order of the controlled system. It is process gain. It is the output of the controlled system. It is the total disturbance resulting from the superposition of internal and external disturbances within the controlled system. It is the input of the controlled system;

[0015] The local model of the process of different orders is represented as arrive ,common Using process local models of different orders as sub-models, a dynamic combination model of the controlled system is established, as shown in formula (2):

[0016] (2)

[0017] In formula (2), These are the connection weights of the sub-model, which are determined by a weight evaluation and update mechanism.

[0018] Furthermore, step S2 includes:

[0019] Time-series historical data of the controlled system In this process, a time series historical data of a preset time period is selected for parameter estimation of each sub-model, where k is the sampling time and the preset time period is the learning time domain; then, a time period close to the current time is selected from the learning time domain as the evaluation time domain; the sampling interval of the learning time domain is represented as shown in formula (3), and the sampling interval of the evaluation time domain is represented as shown in formula (4):

[0020] [ k L , k C ] , k L < k E < k C , (3)

[0021] [ k E , k C ] , k L < k E < k C , (4)

[0022] For the current moment, As the starting point for learning the time domain, the sub-model completes parameter learning and estimation within the interval of formula (3); To evaluate the starting point of the time domain, the sub-model completes the evaluation of parameter performance within the sampling interval of formula (4).

[0023] Furthermore, step S3 includes:

[0024] In the learning time domain, the total disturbance of the controlled system is obtained through multiple iterations of learning by the iteratively expanded state observer IL-ESO. The estimation is performed by estimating the process gain of the controlled system. The learning and updating are completed by up and down perturbation;

[0025] When IL-ESO is run for the first time, if there is no prior knowledge of the state and extended state, the original values ​​of the state and extended state are 0. After IL-ESO is run once, the final values ​​of the previous state and extended state are used as the initial values ​​of the next state and extended state. That is, after the learning time domain is moved forward by one sampling period, the initial values ​​of the state and extended state at the new time are taken as the values ​​of the state and extended state at the previous time.

[0026] Through offline reinforcement learning mechanisms, the original process gain in IL-ESO is improved. Based on this, up and down perturbations are performed to obtain the process gain at the next time step. Three candidate values , and ,in The perturbation coefficient is determined based on the actual rate of change of process gain. The three candidate values ​​are substituted into the IL-ESO to obtain error sequences for three different process gain values.

[0027] Furthermore, step S4 includes:

[0028] In the evaluation time domain, the error sequences of three different process gain values ​​are compared, and the evaluation index function is used to calculate the evaluation index of each error sequence. The evaluation index function is the sum of the absolute values ​​of the error sequence in the evaluation time domain. The evaluation index is selected accordingly. The minimum value corresponds to the process gain value. As shown in formula (5):

[0029] (5)

[0030] in, For error sequences, This represents the maximum number of iterations.

[0031] Apply offline reinforcement learning and IL-ESO to all sub-models to obtain the current time step of each sub-model. Evaluation indicators in the evaluation time domain , ;

[0032] For each sub-model Sort in ascending order, represented as , ,Right now Minimum, Maximum; if If the corresponding error meets the preset modeling accuracy and control accuracy requirements, then only this sub-model is selected as the dynamic combination model, i.e., the corresponding... The order is the order of the controlled system, and the weights are... =1, and the weights of other sub-models are 0; if If the preset modeling accuracy and control accuracy requirements are not met, multiple [items] will be added. Then, a combination test was conducted, and the top three sub-models that met the preset modeling accuracy and control accuracy were selected.

[0033] Furthermore, step S4 includes:

[0034] The weight evaluation and update mechanism is based on the fitting error of each sub-model. Let the top three sub-models that meet the preset modeling accuracy and control accuracy be SM1, SM2, and SM3, respectively. Then there are the following cases from Case 01 to Case 07:

[0035] Case 01: Select SM1 alone; Case 02: Select SM2 alone; Case 03: Select SM3 alone; Case 04: Select the combination of SM1 and SM2; Case 05: Select the combination of SM1 and SM3; Case 06: Select the combination of SM2 and SM3; Case 07: Select the combination of SM1, SM2, and SM3. The individual sub-model methods for Case 01, Case 02, and Case 03 have already been completed in the sorting process, i.e., Case 01 is selected; while for the combinations of Case 04 to Case 07, the evaluation indicators and weight calculation formulas described in formulas (6) to (9) are as follows:

[0036] Case 04: (6)

[0037] Case 05: (7)

[0038] Case 06: (8)

[0039] Case 07: (9)

[0040] In formula (6) For the connection weight of SM1, The connection weights of SM2; in formula (7) For the connection weight of SM1, The connection weights of SM3; in formula (8) For the connection weights of SM2, The connection weights of SM3; in formula (9) For the connection weight of SM1, For the connection weights of SM2, For SM3 connection weights;

[0041] Compare Case 01 Evaluation metrics for Case 04 Evaluation metrics for Case 05 Evaluation metrics for Case 06 And the evaluation metrics for Case07 The sub-model or combination of sub-models with the smallest value is selected as the dynamic combination model of the controlled system, and the corresponding weights are substituted into it. The weights of other unselected sub-models are set to 0.

[0042] A disturbance rejection learning control method for temperature of a chemical reactor, implemented according to the above-mentioned disturbance rejection learning control method based on a dynamic combinatorial model, includes the following steps:

[0043] a. Read the operating status of the chemical reactor, including temperature control, flow control, and stirring control;

[0044] b. Determine whether temperature control is required for the chemical reactor. If temperature control is not required, stop heating the reactor through the reactor heating system and allow the reactor temperature to change freely with the ambient temperature. In this case, no temperature control is performed. Estimate the total perturbation of the sub-model of the dynamic combination model of the reactor heating system in the learning time domain, but the control quantity of the heating power output of the reactor heating system is 0. If temperature control is required, proceed to step c.

[0045] c. Estimate the total perturbation and process gain of the sub-models of the dynamic combined model of the reactor heating system in the learning time domain;

[0046] d. In the evaluation time domain, based on the estimation error of the sub-models, determine the dynamic combination model representing the reactor heating system, and determine the connection weights of each sub-model;

[0047] e. Calculate the control components of the sub-models in the dynamic combination model, determine the final heating power control quantity by connecting the weights, and output the heating power control quantity to the reactor heating system to achieve temperature control of the reactor.

[0048] A non-transitory computer-readable medium storing instructions that, when executed by a processor, perform the steps of the disturbance rejection learning control method based on the dynamic combinatorial model described above.

[0049] A computing device includes a processor and a memory for storing a processor-executable program, wherein when the processor executes the program stored in the memory, it implements the above-described disturbance rejection learning control method based on a dynamic combination model.

[0050] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0051] (1) Compared with traditional control methods based on local and extremely local models, the present invention can handle the disturbance rejection learning control of complex controlled objects, including not only the case of uncertain model parameters, but also the case of uncertain model order.

[0052] (2) Compared with existing neural network control, the present invention reduces the need for process data and a large amount of computing power, and can use ordinary MCU to control complex systems.

[0053] (3) The method of iterative learning and identification improves the utilization rate of historical data of the controlled system. Through the dynamic combination model, the disturbance rejection learning control method based on the dynamic combination model has the characteristics of fractional-order control method. It can smoothly transition between fractional and integer order according to the characteristics of time series data, which greatly improves the scope of application of the present invention.

[0054] (4) Regarding the temperature control of chemical reactors, the chemical reaction process is complex and phase change occurs. The present invention can achieve temperature control in all working states. Attached Figure Description

[0055] Figure 1 This is a structural diagram of the dynamic combination model.

[0056] Figure 2 This is a structural diagram of the weight evaluation and update mechanism for the dynamic combination model.

[0057] Figure 3 This is a structural diagram of a disturbance rejection learning control system based on a dynamic combinatorial model.

[0058] Figure 4 This is a structural diagram of the temperature control system for the reactor.

[0059] Figure 5 This is a schematic diagram of the circuit for measuring the temperature of the reactor.

[0060] Figure 6 This is a data flow diagram for reactor temperature disturbance rejection learning control based on a dynamic combination model.

[0061] Figure 7 This is a flowchart of the reactor temperature disturbance rejection learning control based on a dynamic combination model. Detailed Implementation

[0062] The present invention will be further described below with reference to specific embodiments.

[0063] Example 1

[0064] The disturbance rejection learning control method based on the dynamic combination model provided in this embodiment first adopts the basic local model as shown in formula (1). To represent the controlled system:

[0065] (1)

[0066] In formula (1) It is the system order. It is process gain. It is system output. It is the total disturbance, including both internal and external disturbances of the controlled system. It is the system's input. In traditional MFC and ADRC, the system order is assumed to be... Known, then obtain the information from historical process data. and The estimated value and Then the controller is redesigned, but the order of a real system may vary, or even be fractional, which traditional MFC and ADRC cannot handle.

[0067] Using formula (1) as a sub-model, a dynamic combination model is established. See [link / reference]. Figure 1 As shown. Figure 1 Zhong Cong arrive There are ( ) ) sub-models, through ( The dynamic combination model is constructed from the sub-models, as shown in formula (2):

[0068] (2)

[0069] In formula (2), the sub-model connection weights Determined by the weight evaluation and update mechanism of the sub-model, see [link / reference]. Figure 2 As shown. In the time-series historical data of the actual controlled system. We take historical data for a preset time period to estimate the parameters of each sub-model, where k is the sampling time and the preset time period is the learning time domain; then we take a time period close to the current time from the learning time domain as the evaluation time domain; the sampling interval of the learning time domain is represented as shown in formula (3), and the sampling interval of the evaluation time domain is represented as shown in formula (4):

[0070] [ k L , k C ] , k L < k E < k C , (3)

[0071] [ k E , k C ] , k L < k E < k C , (4)

[0072] For the current moment, As the starting point for learning the time domain, the sub-model completes parameter learning and estimation within the interval of formula (3); To determine the starting point for evaluation in the time domain, the sub-model completes the evaluation of parameter performance within the sampling interval of formula (4). The length of the learning time domain is selected based on the transition time of the controlled system, and is usually set to 4 times the transition time, while the evaluation time domain is the same as the transition time.

[0073] Figure 3 This is a system architecture diagram for disturbance rejection learning control. Figure 3 Medium TD iFor tracking differentiators, the transition process of the input is arranged, and the input is obtained. Approximate differentials of each order v ( t ) = [ v 1 , v 2 , ⋯ , v n ] T The i-th disturbance rejection learning controller within the dashed box is an improvement on the Active Disturbance Rejection Control (ADRC) system structure, adding a weighted system (which can be zero) and employing a novel total disturbance response mechanism. estimation methods and process gain The learning method described here is different from traditional ESO learning, which involves learning from sampled data only once and then discarding the data. In this embodiment, however, the ESO estimates the total perturbation through multiple iterations in the learning time domain, which is called iterative ESO, or IL-ESO (Iterative learning extended state observer). In traditional ESO, the process gain is generally a nominal value that remains constant during the control process. In this embodiment, the learning and updating are accomplished by perturbing the process gain, thus unlike traditional adaptive learning which requires continuous excitation conditions, improving the online learning capability of the parameters.

[0074] This embodiment uses a linear IL-ESO, but this is not intended to limit the scope of the invention; therefore, it can be extended to nonlinear IL-ESOs. In the i-th disturbance rejection learning controller in the learning time domain, the state equation of the controlled system model with system order n is as follows:

[0075] (5)

[0076] The corresponding IL-ESO is:

[0077] { ξ ^ i 1 ( j , k ) = x i 1 ( k ) − x ^ i 1 ( j , k ) ， x i 1 ( k ) = y ( k ) ， j ∈ [ 1 , j m a x ] ， k ∈ [ k L , k C ] x ^ i 1 ( j , k + 1 ) = x ^ i 1 ( j , k ) + h ⋅ x ^ i 2 ( j , k ) + h ⋅ β i 1 ξ ^ i 1 ( j , k ) , x ^ i 2 ( j , k + 1 ) = x ^ i 2 ( j , k ) + h ⋅ x ^ i 3 ( j , k ) + h ⋅ β i 2 ξ ^ i 1 ( j , k ) , ⋯ x ^ i n ( j , k + 1 ) = x ^ i n ( j , k ) + h ⋅ x ^ i ( n + 1 ) ( j , k ) + h ⋅ β i n ξ ^ 1 ( j , k ) + b ^ i ( k ) ⋅ u ( k ) , x ^ i ( n + 1 ) ( j , k + 1 ) = ρ i L ⋅ x ^ i ( n + 1 ) ( j − 1 , k ) + h ⋅ β i ( n + 1 ) ( ρ i f b ⋅ ξ ^ i 1 ( j , k ) + ρ i L C ⋅ ξ ^ i 1 ( j − 1 , k ) ) (6)

[0078] in, , , ..., and To ensure a stable positive value for ESO, the bandwidth method is used for tuning. , , ..., and These are estimates of the state and expansion state of IL-ESO, respectively; These are robust iterative learning coefficients, representing the proportion learned from the previous estimate. ,when When the value is 1, it represents full learning; generally, 0.99 is used. It is the feedforward correction coefficient for learning control. ; It is the feedback correction coefficient for learning control. ; The process gain is denoted by j; j represents the iteration number, where j=1 indicates only one iteration. =1 and This is the traditional ESO. When running IL-ESO for the first time, if there is no prior knowledge of the state and extended state, the original values ​​of the state and extended state can be 0. After one run, the final values ​​of the previous state and extended state can be used as the initial values ​​for the next run. This is because, after shifting the learning time domain forward by one sampling period, the initial values ​​of the state and extended state at the new time step can be taken as the values ​​of the state and extended state at the previous time step. Since the learning time domain continuously shifts with the number of samplings, and the previous sampling period is stored in the database, the estimation of the total perturbation has a long-term accumulation effect over time. Therefore, the number of iterations does not need to be very high; typically, 5 iterations are taken, which is the maximum number of iterations. =5.

[0079] In traditional ADRC and MFC, the process gain is generally a constant nominal value. Through offline reinforcement learning mechanisms, the original process gain in formula (6) is... After applying up and down perturbations to the base, the process gain at the next time step is obtained. Three possible candidate values , and ,in The perturbation coefficient is determined based on the rate of change of gain in the actual process. =0.01, that is, perturbed by a 1% change. Substituting the three candidate values ​​into formula (6) respectively, we obtain the error sequences of three different process gain values. In the evaluation time domain, the three error sequences are compared, and the process gain with the minimum evaluation index is selected. The evaluation index function is the sum of the absolute values ​​of the errors in the evaluation interval, as shown in formula (7):

[0080] (7)

[0081] The offline reinforcement learning in the evaluation time domain is obtained. By learning IL-ESO in the time domain, and .right Figure 1 All sub-models are applied offline reinforcement learning and IL-ESO to obtain the evaluation metrics of each sub-model in the evaluation time domain at the current time step. , .right Sort in ascending order, represented as , ,in Minimum, Maximum. If If the error already meets the requirements for modeling accuracy and control accuracy, then only this sub-model can be selected as the controlled system model, i.e., the dynamic combination model, whose corresponding... The order is the order of the controlled system, and its weight is... =1, and the weights of other sub-models are 0; if If the modeling accuracy requirements are not met, then add multiple [modeling methods]. Then, conduct combination tests and select the top three sub-model combinations.

[0082] Figure 2 The weighted evaluation and update mechanism is based on the fitting error of each sub-model. Let the first three sorted sub-models be SM1, SM2, and SM3. Then there are the cases described in Cases 01 to 07 as follows:

[0083] Case 01: Select SM1 alone; Case 02: Select SM2 alone; Case 03: Select SM3 alone; Case 04: Select the combination of SM1 and SM2; Case 05: Select the combination of SM1 and SM3; Case 06: Select the combination of SM2 and SM3; Case 07: Select the combination of SM1, SM2, and SM3. The individual sub-model methods for Case 01, Case 02, and Case 03 have already been completed in the sorting process, i.e., Case 01 is selected; while for the combinations of Case 04 to Case 07, the evaluation indicators and weight calculation formulas of the following formulas (8) to (11) are as follows:

[0084] Case 04: (8)

[0085] Case 05: (9)

[0086] Case 06: (10)

[0087] Case 07: (11)

[0088] In formula (8) For the connection weight of SM1, The connection weights of SM2; in formula (9) For the connection weight of SM1, The connection weights of SM3; in formula (10) For the connection weights of SM2, The connection weights of SM3; in formula (11) For the connection weight of SM1, For the connection weights of SM2, For SM3 connection weights;

[0089] Compare Case 01 Evaluation metrics for Case 04 Evaluation metrics for Case 05 Evaluation metrics for Case 06 And the evaluation metrics for Case07 The size of the values ​​is used to select the sub-model or combination of sub-models with the smallest value as the dynamic combination model of the controlled system. For example... If the minimum value is found, then Case 07: the combination of SM1+SM2+SM3 forms the dynamic combination model of the controlled system, with the corresponding weights being... , and The weights of other sub-models are 0. Unlike neural network models, the sub-models in dynamic combinatorial models have physical meaning and interpretability. If their contribution to fitting accuracy is too small, their weights can be 0. The learning algorithm of dynamic combinatorial models is offline reinforcement learning and iterative learning methods, rather than error backpropagation algorithms. Compared to conventional integer-order differential equation models or difference equation models, dynamic combinatorial models have the properties of fractional-order models and can better reflect the actual process model.

[0090] See Figure 3 As shown, in the i-th disturbance rejection learning controller, TDi and SEFi adopt the same form as ADRC, and the linear SEF (State error feedback) control is as follows:

[0091] (12)

[0092] in , , ..., To ensure the stability of the closed-loop system, the bandwidth method is used for tuning. , , ..., , e i ( t ) = [ e i 1 , e i 2 , ⋯ , e i n ] T ; , , ..., To track the output of the differentiator TDi, the final output of the disturbance rejection learning control based on the dynamic combination model is... for

[0093] (13).

[0094] Example 2

[0095] See Figure 4As shown, a temperature control system for the reactor is provided. In addition to temperature control, the industrial reactor also includes inlet flow rate (A), inlet flow rate (B), outlet flow rate (C), and stirring control. These controls can be implemented using conventional PID control, specifically the on-chip ADC and timer PWM operation of the MCU shown in the diagram. However, many factors influence temperature control, and the temperature process model is complex. Therefore, the disturbance rejection learning control method based on a dynamic combined model, as described in Example 1, is adopted. Figure 4 The device employs a three-wire platinum resistance temperature sensor and an off-chip analog-to-digital converter (ADC), specifically a 24-bit ΔΣ ADC (ADS1220) with an SPI interface. For details on the interface between the temperature sensor and the off-chip ADC, please refer to [link to documentation]. Figure 5 As shown.

[0096] Figure 4In this design, the MCU uses an STM32G474CBT6. Its SPI3 (PA15, PB3, PB4, PB5) is connected to an external ADC. PA15, PB3, PB4, and PB5 are connected to nCS, SCLK, DOUT, and DIN of the ADS1220, respectively. PA12 is connected to nDRDY of the ADS1220 for reading the ADC data status. PB11 of the MCU drives a relay via a transistor, which in turn controls the motor of the mixer to achieve start and stop control of the mixer. PB6 of the MCU operates in the PWM state of timer TIM4 to adjust the power of the heating wire in the reactor. The PWM output of PB6 drives the bidirectional thyristor BTA41 via the opto-isolated driver FOD4108 to control the heating power of the heating wire, thereby controlling the temperature of the reactor. Three BC7277 LED display drivers and keyboard scanning chips are selected. Each BC7277 has a 9-digit LED display. The three BC7277 chips are used for temperature setpoint and measurement display, flow rate setpoint and measurement display, and operating status and time display, respectively. For temperature display, the first LED on the left displays a letter indicating the heating operating status; the following eight LEDs display the temperature setpoint (4 digits) and the temperature measurement (4 digits). For flow rate display, the first LED on the left indicates the flow rate channel, the second indicates the flow rate operating status, and the third is empty (this LED is optional and not intended to limit the scope of the invention). The following six LEDs alternately display the flow rate setpoint and measurement values ​​for inlet A, inlet B, and outlet C, each occupying three digits. For global status and time display, the first two LEDs on the left display the global operating status, the third is empty, and the following six LEDs display the time. The three chips share a single MCU's SPI2 (PB13, PB14, PB15) interface. PB13, PB14, and PB15 are connected to the BC7277's CLK (pin 24), MISO (pin 23), and MOSI (pin 2), respectively. The MCU's PB7, PB8, and PB9 are connected to the BC7277's chip select CS (pin 4) for temperature, flow rate, and global status / time, respectively. The BC7277 also features a keyboard function for global status and time display, connecting to a 16-key keypad. The MCU connects to the BC7277's KEY pin (pin 22) via PC13. When the BC7277 detects a key press, it notifies the MCU via the KEY pin. The MCU then reads the key status and value via the SPI interface. The MCU uses the keys to input setpoints and control the reactor's start / stop functions.

[0097] Figure 5A schematic diagram of temperature measurement using a platinum resistance thermometer is provided. The platinum resistance thermometer is connected in a three-terminal configuration. R511, R512, and R513 are the wire resistances of the platinum resistance thermometer, and RTD is the thermal resistance of the platinum resistance thermometer itself, using a PT100 resistor. Figure 5 The center-connection method cancels out lead resistance, improving temperature measurement accuracy. ADS1220's AIN2 and AIN3 are used as reference current outputs, set to 1 mA. The voltage across the RTD is connected to the ADC differential input via ANI0 and ANI1, and then to the ADC's internal 24-bit analog-to-digital converter via the ADC's internal PGA (multiplier 8). After conversion, the ADC informs the MCU via nRDY, and the MCU reads the temperature sample value via the SPI interface. The capacitors in the diagram serve as high-frequency filters: C522 is 1 μF, and the others are 0.1 μF; R521 is 1.65 kΩ, R522 and R520 are also 1.65 kΩ, and R510 and R511 are 1 kΩ.

[0098] Figure 6 This is a data flow diagram for disturbance rejection learning control of reactor temperature based on a dynamic combinatorial model. The RTC clock in the diagram is generated by a 32768Hz low-frequency crystal oscillator connected to the MCU, used to generate hours, minutes, and seconds, which are displayed on the digital tube. The system clock is generated by the MCU's SysTick, with a frequency source provided by an external 16MHz main crystal oscillator. After frequency downsampling, it is timed in 1-millisecond increments, used to calculate the sampling period and control period. The MCU samples each process variable according to its sampling period: temperature sampling period is 100 milliseconds, and flow rate sampling period is 50 milliseconds. The MCU contains a single-precision floating-point actuator, generating a 160MHz clock signal from an external 16MHz clock via a PLL as the MCU core frequency. Since the sampling periods are all above 50 milliseconds, the selected MCU's computing power is sufficient to run disturbance rejection learning control based on a dynamic combinatorial model. Figure 6 The key detection and processing program is used to set the setpoints of process variables and the system operating status, and can also calibrate the real-time clock. The digital tube display program reads data from the real-time clock register, process measurement database, process setpoint and operating status database, and displays it on the digital tube. The control program updates according to the system clock and compares it with the control cycle. When the control time is reached, it reads the process setpoint and process measurement value, and completes the control according to the control algorithm. The control cycle is set to be equal to the sampling period. Flow control uses the traditional PID control algorithm, while temperature control uses a disturbance rejection learning control method based on a dynamic combined model. The control program stores the intermediate results required for the learning time domain and evaluation time domain algorithms, such as the previous total disturbance and process gain.

[0099] Figure 7A flowchart of the reactor temperature disturbance rejection learning control based on a dynamic combined model is presented, called once per control cycle. Starting from Step 700, in Step 701, the temperature control status of the system is read. The system operating status includes temperature control, flow control, and stirring control, each controlled separately. In Step 702, it is determined whether temperature control is needed. If not, Step 711 is entered, the PWM output of PB6 is turned off, the pin outputs a low level, the bidirectional thyristor output is turned off, and the heating of the reactor is stopped. The reactor temperature then freely changes with the ambient temperature. However, even if the MCU does not perform temperature power control, it still estimates the total disturbance of each sub-model of the dynamic combined model in the learning time domain. The result is 0, which completes the task in Step 712. Then, proceed to Step 703, returning to the previous level of the program that called the temperature control subroutine. If temperature control is required in the decision made in Step 702, proceed to Step 721. In Step 721, the total perturbation and process gain of the sub-models in the dynamic combination model are estimated in the learning time domain, including obtaining the process gain through offline reinforcement learning in the evaluation time domain. The total perturbation is obtained by learning the IL-ESO in the time domain. Evaluation indicators of error Then proceed to Step 722. In Step 722, in the evaluation time domain, based on the estimation error of the sub-models, determine the dynamic combination model representing the controlled object and determine the weights. When there are three sub-models, calculate according to formulas (8) to (11) described in Example 1, and then compare. , , , , The model with the smallest error is selected as the dynamic combination model of the controlled system, and then the process proceeds to Step 723. In Step 723, the control components of the sub-models in the dynamic combination model are calculated, the final control quantity is determined by the weights, and the control is output. Using the linear SEF algorithm, we have formulas (12) and (13). Finally, the process proceeds to Step 703 and returns to the previous level program that calls the temperature control subroutine.

[0100] Example 3

[0101] This embodiment discloses a non-transitory computer-readable medium storing instructions that, when executed by a processor, perform the steps of the disturbance rejection learning control method based on a dynamic combinatorial model as described in Embodiment 1.

[0102] In this embodiment, the non-transitory computer-readable medium can be a disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), USB flash drive, portable hard drive, etc.

[0103] Example 4

[0104] This embodiment discloses a computing device, including a processor and a memory for storing processor-executable programs. When the processor executes the program stored in the memory, it implements the disturbance rejection learning control method based on a dynamic combinatorial model as described in Embodiment 1.

[0105] The computing device described in this embodiment may be a desktop computer, laptop computer, smartphone, PDA handheld terminal, tablet computer, programmable logic controller (PLC), or other terminal device with processor function.

[0106] The above-described embodiments are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Therefore, any changes made in accordance with the shape and principle of the present invention should be covered within the protection scope of the present invention.

Claims

1. A disturbance rejection learning control method based on a dynamic combinatorial model, applied to model-free or weakly model-dependent control systems, characterized in that: This method adds a weighted system to the output of the controller, uses the process local models of different orders of the controlled system as sub-models, sets a weight evaluation and update mechanism, dynamically determines the connection weights of each sub-model, and combines the sub-models into a dynamic combination model of the controlled system with different orders based on the connection weights. The method involves calculating the comprehensive model parameters of the controlled system based on the dynamic combination model to achieve disturbance rejection learning control of the controlled system. This method includes the following steps: S1. The process local models of different orders of the controlled system are used as sub-models, and the sub-models are combined into a dynamic combination model of the controlled system according to the connection weights. S2. Establish a learning time domain and an evaluation time domain from the time-series historical data of the controlled system, including: Time-series historical data of the controlled system In this process, a time series historical data of a preset time period is selected for parameter estimation of each sub-model, where k is the sampling time and the preset time period is the learning time domain; then, a time period close to the current time is selected from the learning time domain as the evaluation time domain; the sampling interval of the learning time domain is represented as shown in formula (3), and the sampling interval of the evaluation time domain is represented as shown in formula (4): (3)； (4)； For the current moment, As the starting point for learning the time domain, the sub-model completes parameter learning and estimation within the interval of formula (3); To evaluate the starting point of the time domain, the sub-model completes the evaluation of parameter performance within the sampling interval of formula (4); S3. Iterative learning of dynamic combination models of different orders and process gains in the learning time domain to estimate the parameters of the controlled system and the estimation error, including: In the learning time domain, the total disturbance of the controlled system is obtained through multiple iterations of learning by the iteratively expanded state observer IL-ESO. The estimation is performed by estimating the process gain of the controlled system. The learning and updating are completed by up and down perturbation; When IL-ESO is run for the first time, if there is no prior knowledge of the state and extended state, the original values ​​of the state and extended state are 0. After IL-ESO is run once, the final values ​​of the previous state and extended state are used as the initial values ​​of the next state and extended state. That is, after the learning time domain is moved forward by one sampling period, the initial values ​​of the state and extended state at the new time are taken as the values ​​of the state and extended state at the previous time. Through offline reinforcement learning mechanisms, the original process gain in IL-ESO is improved. Based on this, up and down perturbations are performed to obtain the process gain at the next time step. Three candidate values , and ,in The perturbation coefficient is determined based on the actual rate of change of process gain. The three candidate values ​​are substituted into the IL-ESO to obtain the error sequences of three different process gain values. S4. Compare the estimation errors in the evaluation time domain, select the sub-model with the smaller estimation error, obtain the gain of the sub-model through the estimation error, set up a weight evaluation and update mechanism, determine the connection weights between each sub-model, and update the dynamic combination model according to the connection weights. S5. Calculate the control components of the sub-models in the dynamic combination model, determine the final control quantity through the weight system, and output the final control quantity to the controlled system.

2. The disturbance rejection learning control method based on a dynamic combinatorial model according to claim 1, characterized in that, Step S1 includes: The process local model is adopted as shown in formula (1). Indicates the controlled system. (1)； In formula (1), It is the order of the controlled system. It is process gain. It is the output of the controlled system. It is the total disturbance resulting from the superposition of internal and external disturbances within the controlled system. It is the input of the controlled system; The local model of the process of different orders is represented as arrive ,common Using process local models of different orders as sub-models, a dynamic combination model of the controlled system is established, as shown in formula (2): (2)； In formula (2), These are the connection weights of the sub-model, which are determined by a weight evaluation and update mechanism.

3. The disturbance rejection learning control method based on a dynamic combinatorial model according to claim 2, characterized in that, Step S4 includes: In the evaluation time domain, the error sequences of three different process gain values ​​are compared, and the evaluation index function is used to calculate the evaluation index of each error sequence. The evaluation index function is the sum of the absolute values ​​of the error sequence in the evaluation time domain. The evaluation index is selected accordingly. The minimum value corresponds to the process gain value. As shown in formula (5): (5)； in, For error sequences, This represents the maximum number of iterations. Apply offline reinforcement learning and IL-ESO to all sub-models to obtain the current time step of each sub-model. Evaluation indicators in the evaluation time domain , ; For each sub-model Sort in ascending order, represented as , ,Right now Minimum, Maximum; if If the corresponding error meets the preset modeling accuracy and control accuracy requirements, then only this sub-model is selected as the dynamic combination model, i.e., the corresponding... The order is the order of the controlled system, and the weights are... =1, and the weights of other sub-models are 0; if If the preset modeling accuracy and control accuracy requirements are not met, multiple [items] will be added. Then, a combination test was conducted, and the top three sub-models that met the preset modeling accuracy and control accuracy were selected.

4. The disturbance rejection learning control method based on a dynamic combinatorial model according to claim 3, characterized in that, Step S4 includes: The weight evaluation and update mechanism is based on the fitting error of each sub-model. Let the top three sub-models that meet the preset modeling accuracy and control accuracy be SM1, SM2 and SM3, respectively. Then there are the following cases from Case01 to Case07: Case 01: Select SM1 alone; Case 02: Select SM2 alone; Case 03: Select SM3 alone; Case 04: Select the combination of SM1 and SM2; Case 05: Select the combination of SM1 and SM3; Case 06: Select the combination of SM2 and SM3; Case 07: Select the combination of SM1, SM2, and SM3. The individual sub-model methods for Case 01, Case 02, and Case 03 have already been completed in the sorting process, i.e., Case 01 is selected; while for the combinations of Case 04 to Case 07, the evaluation indicators and weight calculation formulas described in formulas (6) to (9) are as follows: Case04: (6); Case05: (7); Case06: (8); Case07: (9); In formula (6) For the connection weight of SM1, The connection weights of SM2; in formula (7) For the connection weight of SM1, The connection weights of SM3; in formula (8) For the connection weights of SM2, The connection weights of SM3; in formula (9) For the connection weight of SM1, For the connection weights of SM2, For SM3 connection weights; Compare Case 01 Evaluation metrics for Case 04 Evaluation metrics for Case 05 Evaluation metrics for Case 06 And the evaluation metrics for Case07 The sub-model or combination of sub-models with the smallest value is selected as the dynamic combination model of the controlled system, and the corresponding weights are substituted into it. The weights of other unselected sub-models are set to 0.

5. A method for temperature disturbance rejection learning control of a chemical reactor, characterized in that, The disturbance rejection learning control method based on a dynamic combinatorial model according to any one of claims 1-4 includes the following steps: a. Read the operating status of the chemical reactor, including temperature control, flow control, and stirring control; b. Determine whether temperature control is required for the chemical reactor. If temperature control is not required, stop heating the reactor through the reactor heating system and allow the reactor temperature to change freely with the ambient temperature. In this case, no temperature control is performed. Estimate the total perturbation of the sub-model of the dynamic combination model of the reactor heating system in the learning time domain, but the control quantity of the heating power output of the reactor heating system is 0. If temperature control is required, proceed to step c. c. Estimate the total perturbation and process gain of the sub-models of the dynamic combined model of the reactor heating system in the learning time domain; d. In the evaluation time domain, based on the estimation error of the sub-models, determine the dynamic combination model representing the reactor heating system, and determine the connection weights of each sub-model; e. Calculate the control components of the sub-models in the dynamic combination model, determine the final heating power control quantity by connecting the weights, and output the heating power control quantity to the reactor heating system to achieve temperature control of the reactor.

6. A non-transitory computer-readable medium storing instructions, characterized in that, When the instruction is executed by the processor, the steps of the disturbance rejection learning control method based on a dynamic combinatorial model according to any one of claims 1-4 are performed.

7. A computing device, comprising a processor and a memory for storing a processor-executable program, characterized in that, When the processor executes the program stored in the memory, it implements the disturbance rejection learning control method based on the dynamic combination model as described in any one of claims 1-4.

Images