Data-driven intelligent greenhouse temperature control method based on sampling data
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- INST OF WATER RESOURCES FOR PASTERAL AREA MINIST OF WATER RESOURCES P R C
- Filing Date
- 2026-03-19
- Publication Date
- 2026-08-07
AI Technical Summary
[0005]综上所述，本发明旨在克服现有半导体激光器温度控制方法存在的严重依赖精确数学模型、自适应能力差及抗干扰能力弱三方面不足，提供一种高精度、强自适应性的控制方案；具体而言，由于传统先进控制方法需要已知激光器及其温控系统的精确热力学模型，而实际模型参数难以准确获取且易漂移，导致模型失配与控制性能下降；同时，广泛应用的PID控制器参数固定，无法自适应激光器工作过程中因驱动电流、环境温度波动及自身老化引起的热动态特性变化，致使变工况下控制精度难以保证；此外，现有方案对温度传感器测量噪声等扰动缺乏有效在线处理机制，噪声直接影响控制器输入，引发控制指令振荡或偏差；因此，本发明提供一种不依赖被控对象精确数学模型、能够在线自适应的数据驱动控制方法，以实现在各种工况及存在测量噪声情况下对半导体激光器工作温度的快速、精确、稳定控制，从而确保其输出波长、功率关键特性的长期稳定性
1、本发明采用数据驱动的无模型自适应控制（MFAC）方法，其核心在于：基于系统运行中实时采集的输入输出数据，动态利用紧格式动态线性化（CFDL）模型，并在线估计该模型中的伪偏导数（PPD）。这一特性使得控制器能够主动适应被控对象（激光器热学系统）的动态变化，彻底摆脱了对复杂、难以获取的精确物理数学模型的依赖；从而克服了传统PID控制因模型失配或参数固定而导致的在变工况下控制性能下降问题，实现了在全工作范围内的持续高精度温度稳定；
Smart Images
Figure CN122526328A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of smart greenhouse temperature control technology, specifically a data-driven smart greenhouse temperature control method based on sampled data. Background Technology
[0002] As a core component of modern optoelectronic systems, the output characteristics of semiconductor lasers are highly dependent on the stability of their junction temperature. Any slight temperature fluctuation can cause laser wavelength drift, output power fluctuation, or even mode jump, which severely restricts their performance in high-end applications such as fiber optic communication, precision spectroscopy, sensing and detection, and quantum technology. Therefore, achieving high-precision and high-stability temperature control is a key prerequisite for ensuring the high-performance operation of semiconductor lasers and extending their service life.
[0003] Traditional temperature control solutions for semiconductor lasers mainly rely on analog temperature control circuits or microprocessor-based digital PID controllers. Analog temperature control circuits use analog electronic components to build control loops, which have fast response speeds but fixed control laws, making complex compensation difficult. They are also susceptible to component aging and temperature drift, making it difficult to guarantee long-term accuracy and consistency. While digital PID solutions have improved parameter tuning and anti-drift capabilities, their control performance largely depends on the precise mathematical model of the controlled object—the "thermal system" composed of the laser thermoelectric cooler (TEC) module and the laser chip. However, this thermal system is a complex system with nonlinear and time-varying characteristics. Its parameters, such as heat capacity and thermal resistance, change with the laser drive current, ambient temperature, and its own aging state. This makes it difficult for PID controllers based on fixed parameters to maintain optimal performance under various operating conditions, resulting in slow response, overshoot, or steady-state errors.
[0004] In recent years, with the development of high-precision data acquisition technology and advanced control theory, data-driven control methods have provided a new solution to overcome the limitations of traditional temperature control technology. Such methods do not require the prior establishment of an accurate physical model of the laser thermal system. Instead, they directly utilize the input and output data collected during system operation and dynamically adjust the control strategy through online or offline learning. This enables the system to have adaptive and self-learning capabilities, opening up a new technical path for achieving higher precision and stronger robustness in semiconductor laser temperature control.
[0005] In summary, this invention aims to overcome the three shortcomings of existing semiconductor laser temperature control methods: heavy reliance on precise mathematical models, poor adaptability, and weak anti-interference capabilities. It provides a high-precision, highly adaptive control scheme. Specifically, traditional advanced control methods require a known precise thermodynamic model of the laser and its temperature control system. However, actual model parameters are difficult to obtain accurately and are prone to drift, leading to model mismatch and decreased control performance. Furthermore, widely used PID controllers have fixed parameters and cannot adapt to changes in the laser's thermal dynamic characteristics caused by driving current, ambient temperature fluctuations, and aging, making it difficult to guarantee control accuracy under varying operating conditions. In addition, existing solutions lack effective online processing mechanisms for disturbances such as temperature sensor measurement noise, which directly affects the controller input, causing control command oscillations or deviations. Therefore, this invention provides a data-driven control method that does not rely on a precise mathematical model of the controlled object and can adapt online, enabling rapid, accurate, and stable control of the semiconductor laser's operating temperature under various operating conditions and in the presence of measurement noise, thereby ensuring the long-term stability of its key characteristics such as output wavelength and power. Summary of the Invention
[0006] In view of the problems existing in the prior art, the present invention discloses a data-driven smart greenhouse temperature control method based on sampled data. The technical solution adopted includes the following steps: Step 1: System Modeling and Discretization: The continuous-time temperature model of the smart greenhouse is sampled and discretized to obtain the input and output sample data of the system, and a discrete-time temperature model of the smart greenhouse is established. The discrete-time dynamic temperature model of the smart greenhouse satisfies the preset assumptions and lemmas. Step 2: Set the desired temperature: Based on the equipment's operating status, set the desired temperature value that the smart greenhouse needs to achieve and maintain; Step 3: Sensor noise suppression: Based on the noise model of sensor measurement data, the weighted width learning system (WBLS) algorithm is used to process the noisy measurement data. Reliable samples and abnormal samples are distinguished by weighted penalty factors. Combined with the weighted incremental learning algorithm, the model is quickly reconstructed to suppress the influence of noise and outliers, and output the temperature prediction value after noise suppression. Step 4: Construct a data-driven linearized model: Based on the discretized system input and output data obtained in Step 1, construct a model-free compact form dynamic linearized data model with an observer. This model dynamically approximates the complex nonlinear greenhouse temperature system as a linear system near each operating point. The model includes slowly time-varying pseudo-partial derivative parameters. Step 5: Output Feedback: In the controller design process, the noise-suppressed temperature prediction output value obtained in Step 3 is used to replace the direct measurement output value of the greenhouse temperature model; Step Six: Controller Design: Model-free adaptive control design for the smart greenhouse temperature system, including: online estimation of pseudo-partial derivatives in the compact form dynamic linearized data model, design of a pseudo-partial derivative resetting algorithm to prevent the estimated value from drifting or becoming too large, and comprehensive construction of the actual controller output of the smart greenhouse temperature system; Step 7: Temperature control execution: Using the actual controller obtained in Step 6, control commands are generated and applied to the actuators of the smart greenhouse, so that the actual temperature of the greenhouse can be quickly tracked and stabilized near the desired temperature, and the control error is limited to the preset allowable range.
[0007] As a preferred embodiment of the present invention, in step one, the smart greenhouse temperature model is as follows: Define external disturbance variables, where, Solar radiation, Outdoor temperature; , , All are system thermodynamic parameters; the standardized representation of the control input is: ,in, This indicates the input to the greenhouse heating system, controlled by the percentage of the heating equipment's power opening. This indicates the input to the greenhouse's natural ventilation system, controlled by the percentage of window opening. It is about A linear function, Indicates about The row vector.
[0008] As a preferred embodiment of the present invention, in step one, after sampling and discretizing the smart greenhouse temperature model to obtain sampled data, the smart greenhouse temperature model under discrete time is obtained as follows: ; in, Indicates the greenhouse The actual temperature at that moment; It is about A linear function; Indicates the first The actual temperature inside the greenhouse at all times; Indicates about The row vector; the sampling time of the smart greenhouse temperature controller is The intelligent greenhouse temperature model is only used to generate the system's input and output data and does not participate in the controller design.
[0009] As a preferred embodiment of the present invention, in step one, a dynamic linearization method is used to transform the original nonlinear system into a linear data model, facilitating subsequent controller design. The preset assumptions and lemmas are as follows: make , , For smart greenhouses in ( The actual change in temperature at any given moment; Assumption 1: Except for a finite number of time points, the function right , and It has continuous partial derivatives; Assumption 2: Function Satisfying the generalized Lipschitz condition, that is, for any and ,have ,in and It is the Lipschitz constant; Lemma 1: For a smart greenhouse temperature system that satisfies Assumption 1 and Assumption 2, when At that time, there exists a time-varying parameter vector that transforms the system into a sampled data input / output model, and uses replace The model is as follows: ; in, , For slow time-varying parameters, The symbol for any moment All remain unchanged. Always greater than 0 yes Constantly control the amount of change in the input. ; It is a nonlinear uncertainty term.
[0010] As a preferred embodiment of the present invention, in step two, the desired temperature is set by initially setting it based on the rated temperature and wavelength-temperature tuning coefficient provided in the laser datasheet.
[0011] As a preferred embodiment of the present invention, in step three, the noise model of the sensor measurement data is: ,in The noise is unbiased white Gaussian noise with a small fluctuation amplitude.
[0012] As a preferred embodiment of the present invention, the specific implementation process of the weighted width learning system (WBLS) algorithm in step three is as follows: The control time of the intelligent greenhouse temperature control system is divided into time windows, let the first... The width of the training data in each time window is The training input data includes control input. and , will the The training input data for each window is denoted as . , ; Group feature nodes; each group contains One node; Grouped enhancement nodes; each group contains One node; No. The output of the group feature nodes is: , No. The output of the group enhancement node is: , The output of the feature node is represented as ; and From the distribution range Randomly generated, and It is the activation function, and the output of all boosted nodes is represented as ; A generalized learning network is represented as: ,in, This is the output label vector; the WBLS used is built based on the standard BLS, and the connection weights of WBLS can be calculated using the weighted ridge regression algorithm as follows: ; ; in, It is a weighted penalty factor. This represents the extended input matrix. It is a regularization coefficient close to zero. It is the identity matrix; The weighted penalty factor is calculated using the Huber weighting function: ; in, Indicates the first A weighted penalty factor, Indicates an adjustable positive parameter; It is the first in the window data The standardized residuals of each sample are calculated using the following formula: ; in, Indicates the first The residuals between the predicted and measured values of each sample; the residuals of all samples form a residual matrix. ; It is a robust sizing estimator, calculated using the following formula: MAR Where MAR represents the median absolute residual, which can be calculated using the following formula: ; in, The median function is represented by the Huber weight function; WBLS based on the Huber weight function is called Huber-WBLS. The connection weights and weighted penalty factors can be calculated iteratively to obtain higher accuracy. The maximum absolute value of the difference between consecutive solutions is represented as follows: ; When adding new samples to the model, a weighted incremental learning algorithm is used, assuming the first... The output matrices of the feature layer and enhancement layer for training data within each time window are respectively represented as follows: , ; in, Is it towards the first The weights are calculated after incremental updates are added to the training data within each time window; therefore, the output layer weights are updated as follows: ; ; ; ; in, It is being updated gradually. The system's predictions are calculated using training data from multiple time windows; these predictions can be obtained in the following ways: ; in, It is the expanded version Input matrix; by combining the generated connection matrix with Multiplication allows us to calculate the predicted data for each time period. .
[0013] As a preferred embodiment of the present invention, in step four, the sampled data with the observer is model-free compact-format dynamic linearized data model, which will... Defined as The estimated value is obtained as follows: ; in, This represents the total uncertainty of the system. This indicates the parameter estimation error; Based on the model-free compact-form dynamic linearized data model of the sampled data, a proposed... estimation algorithm ; in, and It is the step size factor; The sampling data control law is: ; ; in, , These are the constraint values that control the input; It is the step size factor. It is a bounded expected temperature value; Since the data is unknown, a sampled data extended state observer is proposed.
[0014] As a preferred embodiment of the present invention, the observer is a sampled data extended state observer, designed as follows: definition The expression for the sampled data extended state observer is: ; The sampled data extended state observer is proposed as follows: ; in, and They represent respectively to and The estimate, and It is the observer gain.
[0015] The beneficial effects of this invention are: 1. This invention employs a data-driven model-free adaptive control (MFAC) method. Its core lies in dynamically utilizing a compact form dynamic linearization (CFDL) model based on real-time input / output data acquired during system operation, and estimating the pseudo-partial derivatives (PPD) in this model online. This characteristic enables the controller to proactively adapt to the dynamic changes of the controlled object (laser thermal system), completely eliminating the dependence on complex and difficult-to-obtain precise physical and mathematical models. This overcomes the performance degradation problem of traditional PID control under varying operating conditions caused by model mismatch or fixed parameters, achieving continuous high-precision temperature stability across the entire operating range. 2. By introducing a weighted width learning system (WBLS) to preprocess sensor measurement data, this invention can effectively suppress the interference of measurement noise and outliers on the control system. (WBLS) automatically distinguishes reliable samples from outliers through a weighted penalty factor, reducing the negative impact of noise on controller design. This technical feature directly brings stronger anti-interference capability, enabling the control system to maintain stable operation and precise control even when there are uncertainties such as sensor noise and external environmental fluctuations. 3. By utilizing a compact form dynamic linearization (CFDL) data model, this invention establishes a dynamic linearization data model for the complex nonlinear, time-varying semiconductor laser temperature control system at each dynamic operating point. This technical feature transforms the complex nonlinear control problem into a series of linear control problems that are easy to solve online, greatly reducing the difficulty of controller design and implementation while ensuring good control performance. 4. The pseudo-partial derivative (PPD) online estimation algorithm and its reset mechanism designed in this invention have low computational cost and meet the requirements of real-time control. At the same time, the assumption and guarantee of constant PPD sign (always greater than zero) and the boundedness analysis of all signals in the scheme theoretically ensure the stability and reliability of the control system and avoid the risk of system runaway due to parameter drift or calculation divergence. 5. By implementing this invention, an extremely stable and precise operating temperature can be provided for semiconductor lasers; this directly translates into long-term stability of laser output power, effective locking of the center wavelength, suppression of spectral linewidth, and elimination of mode jumps, thereby ensuring their performance and reliability in high-end applications and helping to extend their service life. Attached Figure Description
[0016] Figure 1 This is an overall flowchart of the present invention;
[0017] Figure 2 This is a diagram of the width learning structure of the present invention;
[0018] Figure 3 This is a simulation diagram of the intelligent greenhouse temperature system of the present invention;
[0019] Figure 4 This is a temperature tracking error diagram for the smart greenhouse of this invention. Detailed Implementation
[0020] Example 1
[0021] like Figures 1 to 4 As shown, this invention discloses a data-driven smart greenhouse temperature control method based on sampled data. The technical solution adopted includes the following steps: Step 1: System Modeling and Discretization: The continuous-time temperature model of the smart greenhouse is sampled and discretized to obtain the input and output sample data of the system, and a discrete-time temperature model of the smart greenhouse is established. The discrete-time dynamic temperature model of the smart greenhouse satisfies the preset assumptions and lemmas. The temperature model for the smart greenhouse is as follows: Define external disturbance variables, where, Solar radiation, Outdoor temperature; , , All are system thermodynamic parameters; the standardized representation of the control input is: ,in, This indicates the input to the greenhouse heating system, controlled by the percentage of the heating equipment's power opening. This indicates the input to the greenhouse's natural ventilation system, controlled by the percentage of window opening. It is a linear function of Tin. Indicates about The row vector; the control objective is to bring the greenhouse temperature to the optimal reference temperature while minimizing energy consumption; After sampling and discretizing the temperature model of the smart greenhouse to obtain the sampled data, the discrete-time temperature model of the smart greenhouse is obtained as follows: ; in, Indicates the greenhouse The actual temperature at that moment; It is about A linear function; Indicates the first The actual temperature inside the greenhouse at all times; Indicates about The row vector; the sampling time of the smart greenhouse temperature controller is The intelligent greenhouse temperature model is only used to generate the system's input and output data and does not participate in the controller design. The pre-defined assumptions and lemmas are as follows: set up , , For smart greenhouses in ( The actual change in temperature at any given moment; Assumption 1: Except for a finite number of time points, the function right , and It has continuous partial derivatives; Assumption 2: Function Satisfying the generalized Lipschitz condition, that is, for any and ,have ,in and It is the Lipschitz constant; Lemma 1: For a smart greenhouse temperature system that satisfies Assumption 1 and Assumption 2, when At that time, there exists a time-varying parameter vector that transforms the system into a sampled data input / output model, and uses replace The model is as follows: ; in, , These are slow, time-varying parameters, representing pseudo-partial derivatives of the compact-scheme dynamic linearization data model. They approximate complex nonlinear systems as linear systems around each operating point, reducing the control difficulty of nonlinear systems. The symbol for any moment All remain unchanged. Always greater than 0 yes Constantly control the amount of change in the input. ; It is a nonlinear uncertainty term; Step 2: Set the desired temperature: Based on the equipment's operating status, set the desired temperature value that needs to be reached and maintained inside the smart greenhouse; Setting the reference temperature of the semiconductor laser is a systematic process that balances its performance, lifespan, and reliability. The desired temperature is set as follows: Initially set according to the rated temperature and wavelength-temperature tuning coefficient provided in the laser datasheet; Strictly adhere to the maximum permissible case temperature to prevent damage. Based on this, adjustments should be made according to the specific application scenario. Step 3: Sensor Noise Suppression: Based on the noise model of the sensor measurement data, the Weighted Width Learning System (WBLS) algorithm is used to process the noisy measurement data. A weighted penalty factor distinguishes reliable samples from outliers, and a weighted incremental learning algorithm is combined to achieve rapid model reconstruction, suppressing the influence of noise and outliers, and outputting the noise-suppressed temperature prediction value. The noise model of the sensor measurement data is as follows: ,in The noise is unbiased white Gaussian noise with small fluctuation amplitude; it is used to simulate the random error of sensor measurements in reality. To eliminate the adverse effects of noise and outliers on modeling, this paper introduces a weighted penalty factor into Broad Learning System (BLS). By automatically assigning appropriate weights to each sample, samples with high reliability are given higher weights, while suspicious outlier samples receive lower weights, thereby reducing the contribution of suspected outliers to the modeling. The weighted ridge regression algorithm is used to solve the proposed algorithm, and the corresponding weighted incremental learning system (WIBLS) algorithm is used to achieve fast model reconstruction. The specific implementation process of the Weighted Width Learning System (WBLS) algorithm is as follows: The control time of the intelligent greenhouse temperature control system is divided into time windows, let the first... The width of the training data in each time window is BLS training input data includes control input. and , will the The training input data for each window is denoted as . , ; Group feature nodes; each group contains One node; Grouped enhancement nodes; each group contains One node; No. The output of the group feature nodes is: , No. The output of the group enhancement node is: , in, ; and From a given appropriate distribution range Randomly generated, and It is the activation function, and the output of all boosted nodes is represented as ; A generalized learning network is represented as: ,in, This is the output label vector; the WBLS used is built based on the standard BLS, and the connection weights of WBLS can be calculated using the weighted ridge regression algorithm as follows: ; ; in, It is a weighted penalty factor. This represents the extended input matrix. It is a regularization coefficient close to zero. It is the identity matrix; The weighted penalty factor is calculated using the Huber weighting function; drawing inspiration from iterative weighted least squares, the Huber weighting function is used to handle sample residuals; the Huber weighting function is expressed in the following form: ; in, Indicates the first A weighted penalty factor, Indicates an adjustable positive parameter; It is the first in the window data The standardized residuals of each sample are calculated using the following formula: ; in, Indicates the first The residuals between the predicted and measured values of each sample; the residuals of all samples form a residual matrix. ; It is a robust sizing estimator, calculated using the following formula: MAR Where MAR represents the median absolute residual, which can be calculated using the following formula: ; in, The median function is represented by the Huber weight function; WBLS based on the Huber weight function is called Huber-WBLS. In this algorithm, the connection weights and weighted penalty factors can be calculated iteratively to obtain higher accuracy. The maximum absolute value of the difference between consecutive solutions is represented as follows: ; Adding additional boosting nodes when adding new samples to a model is one of the convenient ways to improve the learning performance of the network. Using a weighted incremental learning algorithm, nodes can be added quickly without retraining the existing model, and the robustness and generalization ability of the model can be improved without repeating the entire training process. Assume the first The output matrices of the feature layer and enhancement layer for training data within each time window are respectively represented as follows: , ; in, Is it towards the first The weights are calculated after incremental updates are added to the training data within each time window; therefore, the output layer weights are updated as follows: ; ; ; ; in, It is being updated gradually. The system's predictions are calculated using training data from multiple time windows; these predictions can be obtained in the following ways: ; in, It is the expanded version Input matrix; by combining the generated connection matrix with Multiplication allows us to calculate the predicted data for each time period. The weighted incremental width learning system eliminates the negative impact of measurement disturbance estimation on the output, and the predicted values are used in subsequent control algorithms. Step 4: Construct a data-driven linearized model: Based on the discretized system input and output data obtained in Step 1, construct a model-free compact form dynamic linearized data model with an observer. This model dynamically approximates the complex nonlinear greenhouse temperature system as a linear system near each operating point. The model includes slowly time-varying pseudo-partial derivative parameters. The sampled data model with the observer is a model-free compact-format dynamic linearized data model, which will... Defined as The estimated value is obtained as follows: ; in, This represents the total uncertainty of the system. This indicates the parameter estimation error; Based on the model-free compact-form dynamic linearized data model of the sampled data, a proposed... estimation algorithm ; in, and It is the step size factor; The sampling data control law is: ; ; in, , These are the constraint values that control the input; It is the step size factor. It is a bounded expected temperature value; Since it is unknown, a sampled data extended state observer is proposed; The observer is a sampled data extended state observer, designed as follows: definition The expression for the sampled data extended state observer is: ; The sampled data extended state observer is proposed as follows: ; in, and They represent respectively to and The estimate, and It is the observer gain; The algorithm employs a sampling period to compensate for its impact on control performance and noise. A saturation function is designed for the control law, effectively handling greenhouse input constraints. Unlike traditional model-free adaptive control, Observer-Based Sampled-Data MFAC (ObSMFAC) uses an observer to estimate the total uncertainty as an extended state. Therefore, ObSMFAC can actively resist both internal uncertainties and external disturbances. ObSMFAC is almost entirely data-driven; aside from the sampling period and structural information from state to output, it does not rely on any other model knowledge of the continuous-time nonlinear system. Step 5: Output Feedback: In the controller design process, the noise-suppressed temperature prediction output value obtained in Step 3 is used to replace the direct measurement output value of the greenhouse temperature model in order to improve the robustness of the system. Step Six: Controller Design: Model-free adaptive control design for the smart greenhouse temperature system, including: online estimation of pseudo-partial derivatives in the compact form dynamic linearized data model, design of a pseudo-partial derivative resetting algorithm to prevent the estimated value from drifting or becoming too large, and comprehensive construction of the actual controller output of the smart greenhouse temperature system; Step 7: Temperature Control Execution: Using the actual controller obtained in Step 6, control commands are generated and applied to the actuators of the smart greenhouse, such as the ventilation system and heating equipment, so that the actual temperature of the greenhouse can be quickly tracked and stabilized near the desired temperature, and the control error is limited to the preset allowable range.
[0022] The specific steps in the temperature control process of a smart greenhouse system are as follows: Step A: Set the reference temperature. Initially set the temperature based on the rated temperature and wavelength-temperature tuning factor provided in the laser datasheet, and strictly adhere to the maximum permissible case temperature to prevent damage. Adjust the temperature according to the specific application scenario. Step B: Collect greenhouse thermodynamic parameters using sensors. Through multi-source sensors deployed in the greenhouse, key thermodynamic parameters such as temperature and light intensity are monitored and collected in real time, providing a data foundation for system modeling and control. Step C: Establish a model of the smart greenhouse temperature system. Based on the real-time monitoring data obtained in Step 2, construct a mechanism or data-driven model that reflects the dynamic changes in greenhouse temperature and describes the heat exchange and energy balance relationship within the system. Step D: Construct a dynamic linearization data model. Based on the nonlinear characteristics of the greenhouse temperature system, and building upon Steps 2 and 3, a dynamic linearization method is used to transform the original nonlinear system into a linear data model, which facilitates the subsequent controller design. Step E: Weighted incremental width learning is used to suppress sensor disturbances and reduce the impact of noise and disturbances in the sensor measurement process on prediction accuracy. The weighted incremental width learning method is introduced to estimate the system state, thereby outputting a more accurate and robust temperature prediction value. Step F: Design a model-free adaptive controller and a pseudo-partial derivative reset algorithm to estimate the pseudo-partial derivatives of the system online, thereby completing the structural design of the model-free adaptive controller; further, design a pseudo-partial derivative reset algorithm to realize real-time updating and correction of controller parameters, ensuring that the control behavior conforms to the actual physical constraints and can effectively track the reference temperature set in Step I. In summary, by combining the above six steps, especially the weighted incremental width learning technique used in step five, and the pseudo-partial derivative estimation and reset strategy based on model-free adaptive control in step six, this system can achieve precise control of greenhouse temperature, ensuring that the actual temperature stably tracks the set reference value and that the control error is always kept within the preset range.
[0023] Components not described in detail in this article are existing technologies.
[0024] While the specific embodiments of the present invention have been described in detail above, the present invention is not limited to the above embodiments. Within the scope of knowledge possessed by those skilled in the art, various changes can be made without departing from the spirit of the present invention, and modifications or variations without creative effort are still within the protection scope of the present invention.
Claims
1. A data-driven smart greenhouse temperature control method based on sampled data, characterized in that, Includes the following steps: Step 1: System Modeling and Discretization: The continuous-time temperature model of the smart greenhouse is sampled and discretized to obtain the input and output sample data of the system, and a discrete-time temperature model of the smart greenhouse is established. The discrete-time dynamic temperature model of the smart greenhouse satisfies the preset assumptions and lemmas. Step 2: Set the desired temperature: Based on the equipment's operating status, set the desired temperature value that the smart greenhouse needs to achieve and maintain; Step 3: Sensor noise suppression: Based on the noise model of sensor measurement data, the weighted width learning system (WBLS) algorithm is used to process the noisy measurement data. Reliable samples and abnormal samples are distinguished by weighted penalty factors. Combined with the weighted incremental learning algorithm, the model is quickly reconstructed to suppress the influence of noise and outliers, and output the temperature prediction value after noise suppression. Step 4: Construct a data-driven linearized model: Based on the discretized system input and output data obtained in Step 1, construct a model-free compact form dynamic linearized data model with an observer. This model dynamically approximates the complex nonlinear greenhouse temperature system as a linear system near each operating point. The model includes slowly time-varying pseudo-partial derivative parameters. Step 5: Output Feedback: In the controller design process, the noise-suppressed temperature prediction output value obtained in Step 3 is used to replace the direct measurement output value of the greenhouse temperature model; Step Six: Controller Design: Model-free adaptive control design for the smart greenhouse temperature system, including: online estimation of pseudo-partial derivatives in the compact form dynamic linearized data model, design of a pseudo-partial derivative resetting algorithm to prevent the estimated value from drifting or becoming too large, and comprehensive construction of the actual controller output of the smart greenhouse temperature system; Step 7: Temperature control execution: Using the actual controller obtained in Step 6, control commands are generated and applied to the actuators of the smart greenhouse, so that the actual temperature of the greenhouse can be quickly tracked and stabilized near the desired temperature, and the control error is limited to the preset allowable range.
2. The data-driven smart greenhouse temperature control method based on sampled data according to claim 1, characterized in that: In step one, the temperature model of the smart greenhouse is as follows: Define external disturbance variables, where, Solar radiation, Outdoor temperature; , , All are system thermodynamic parameters; the standardized representation of the control input is: ,in, This indicates the input to the greenhouse heating system, controlled by the percentage of the heating equipment's power opening. This indicates the input to the greenhouse's natural ventilation system, controlled by the percentage of window opening. It is about A linear function, Indicates about The row vector.
3. The data-driven smart greenhouse temperature control method based on sampled data according to claim 2, characterized in that: In step one, after sampling and discretizing the smart greenhouse temperature model to obtain sampled data, the discrete-time smart greenhouse temperature model is obtained as follows: ； in, Indicates the greenhouse The actual temperature at that moment; It is about A linear function; Indicates the first The actual temperature inside the greenhouse at all times; Indicates about The row vector; the sampling time of the smart greenhouse temperature controller is The intelligent greenhouse temperature model is only used to generate the system's input and output data and does not participate in the controller design.
4. The data-driven smart greenhouse temperature control method based on sampled data according to claim 3, characterized in that: In step one, the preset assumptions and lemmas are as follows: set up , , For smart greenhouses in ( The actual change in temperature at any given moment; Assumption 1: Except for a finite number of time points, the function right , and It has continuous partial derivatives; Assumption 2: Function Satisfying the generalized Lipschitz condition, that is, for any and ,have ,in and It is the Lipschitz constant; Lemma 1: For a smart greenhouse temperature system that satisfies Assumption 1 and Assumption 2, when At that time, there exists a time-varying parameter vector that transforms the system into a sampled data input / output model, and uses replace The model is as follows: ； in, , For slow time-varying parameters, The symbol for any moment All remain unchanged. Always greater than 0 yes Constantly control the amount of change in the input. ; It is a nonlinear uncertainty term.
5. The data-driven smart greenhouse temperature control method based on sampled data according to claim 1, characterized in that: In step two, the desired temperature is set by initially setting it based on the rated temperature and wavelength-temperature tuning coefficient provided in the laser datasheet.
6. The data-driven smart greenhouse temperature control method based on sampled data according to claim 1, characterized in that: In step three, the noise model of the sensor measurement data is: ,in The noise is unbiased white Gaussian noise with a small fluctuation amplitude.
7. The data-driven smart greenhouse temperature control method based on sampled data according to claim 6, characterized in that: In step three, the specific implementation process of the weighted width learning system (WBLS) algorithm is as follows: The control time of the intelligent greenhouse temperature control system is divided into time windows, let the first... The width of the training data in each time window is The training input data includes control input. and , will the The training input data for each window is denoted as . , ; Group feature nodes; each group contains One node; Grouped enhancement nodes; each group contains One node; No. The output of the group feature nodes is: , No. The output of the group enhancement node is: , The output of the feature node is represented as ; and From the distribution range Randomly generated, and It is the activation function, and the output of all boosted nodes is represented as ; A width-based learning network is represented as: ,in, This is the output label vector; the WBLS used is built based on the standard BLS, and the connection weights of WBLS can be calculated using the weighted ridge regression algorithm as follows: ； ； in, It is a weighted penalty factor. This represents the extended input matrix. It is a regularization coefficient close to zero. It is the identity matrix; The weighted penalty factor is calculated using the Huber weighting function: ； in, Indicates the first A weighted penalty factor, Indicates an adjustable positive parameter; It is the first in the window data The standardized residuals of each sample are calculated using the following formula: ； in, Indicates the first The residuals between the predicted and measured values of each sample; the residuals of all samples form a residual matrix. ; It is a robust sizing estimator, calculated using the following formula: MAR Where MAR represents the median absolute residual, which can be calculated using the following formula: ； in, The median function is represented by the Huber weight function; WBLS based on the Huber weight function is called Huber-WBLS. The connection weights and weighted penalty factors can be calculated iteratively to obtain higher accuracy. The maximum absolute value of the difference between consecutive solutions is represented as follows: ； When adding new samples to the model, a weighted incremental learning algorithm is used, assuming the first... The output matrices of the feature layer and enhancement layer for training data within each time window are respectively represented as follows: , ； in, Is it towards the first The weights are calculated after incremental updates are added to the training data within each time window; therefore, the output layer weights are updated as follows: ； ； ； ； in, It is being updated gradually. The system's predictions are calculated using training data from multiple time windows; these predictions can be obtained in the following ways: ； in, It is the expanded version Input matrix; by combining the generated connection matrix with Multiplication allows us to calculate the predicted data for each time period. .
8. The data-driven smart greenhouse temperature control method based on sampled data according to claim 1, characterized in that: In step four, the sampled data with the observer is model-free, compact-format, dynamically linearized data model, which will... Defined as The estimated value is obtained as follows: ； in, This represents the total uncertainty of the system. This indicates the parameter estimation error; Based on the model-free compact-form dynamic linearized data model of the sampled data, a proposed... estimation algorithm ； in, and It is the step size factor; The sampling data control law is: ； ； in, It is the step size factor. It is a bounded expected temperature value; Since it is unknown, a sampled data extended state observer is proposed. These are the constraint values that control the input, where .
9. The data-driven smart greenhouse temperature control method based on sampled data according to claim 8, characterized in that: The observer is a sampled data extended state observer, designed as follows: definition The expression for the sampled data extended state observer is: ； The sampled data extended state observer is proposed as follows: ； in, and They represent respectively to and The estimate, and It is the observer gain.