Greenhouse climate multi-objective intelligent control method fusing rule constraints and reinforcement learning

Abstract

The application discloses a greenhouse climate multi-objective intelligent control method fusing rule constraints and reinforcement learning, and relates to the technical field of agricultural environment control. The method comprises the following steps: constructing a temperature and humidity dynamic model based on the law of conservation of energy, and establishing a mapping relationship among temperature, relative humidity and actuator input data by coupling external environmental disturbance terms; generating a plurality of interval training data based on the mapping relationship; constructing an intelligent decision model; taking the deviation degree of the environmental target value and the interval range as an evaluation benchmark; screening an optimal actuator state vector; setting a difference mediation threshold of mutually exclusive operation for safety correction; taking the environmental adjustment difference and the energy consumption minimization as optimization objectives; designing a multi-objective reward function; determining an optimal control instruction; updating the intelligent decision model; and realizing accurate and efficient greenhouse climate control, so that the comprehensive control of temperature and humidity is more accurate and reasonable, and the energy consumption is reduced.

Description

Technical Field

[0001] This invention relates to the field of agricultural environmental control technology, specifically to a multi-objective intelligent control method for greenhouse climate that integrates rule constraints and reinforcement learning. Background Technology

[0002] Greenhouse agriculture is an important way to ensure global food security and achieve year-round, high-efficiency crop production. Its core objective is to create a controlled internal microclimate environment to isolate the effects of adverse external climatic conditions. In this environment, temperature and relative humidity are considered two of the most critical and interdependent environmental factors. They not only directly affect the physiological processes of crops, such as photosynthesis, respiration, and transpiration, but also jointly determine the growth, development, and final yield and quality of crops. Therefore, the precise control technology of greenhouse temperature and relative humidity has become a key link in modern agricultural engineering and is of great significance for improving agricultural production efficiency and sustainability.

[0003] In existing technologies, the widely used PID control method typically designs the temperature and humidity control loops independently, ignoring the physical coupling relationship between them: heating will reduce relative humidity due to the increase in temperature, while humidification will cause the temperature to drop due to the evaporative cooling effect. This coupling characteristic can easily cause control command conflicts, resulting in unreasonable operations such as "heating and cooling at the same time" or "humidifying and dehumidifying at the same time", which leads to energy waste, increased operating costs, and does not meet the requirements of sustainable agricultural development. Therefore, it is essential to develop an intelligent decision-making method that can respond to environmental changes in real time and optimize multi-objective control.

[0004] The information disclosed in the background section is only intended to enhance the understanding of the background of this disclosure, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0005] The purpose of this invention is to provide a multi-objective intelligent control method for greenhouse climate that integrates rule constraints and reinforcement learning, so as to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: A multi-objective intelligent control method for greenhouse climate that integrates rule constraints and reinforcement learning, comprising the following steps: Based on the law of conservation of energy, a temperature dynamic model and a humidity dynamic model of the greenhouse to be controlled are constructed. According to the dynamics of solar radiation and crop transpiration, an external environmental disturbance term is set and coupled with the temperature dynamic model and the humidity dynamic model to establish a mapping relationship between temperature, relative humidity and actuator input data. Based on the mapping relationship, multiple sets of training data are output, and the training data are discretized into intervals to generate multiple sets of interval training data and construct interval mapping relationships. The interval training data includes temperature intervals, relative humidity intervals, and actuator state vectors composed of each actuator control level. An intelligent decision-making model is constructed. The deviation between the environmental target value and the range is used as the evaluation benchmark. Based on the range mapping relationship, the current temperature range and the current relative humidity range are analyzed to determine the evaluation index of each actuator state vector and select the preferred actuator state vector. The difference adjustment threshold of the mutual exclusion operation of each actuator is set. Based on the difference adjustment threshold, the preferred actuator state vector is screened again to form the selectable actuator state vector. Based on the safety correction amount of the state vector of each selectable actuator, and with the optimization objectives of minimizing environmental adjustment differences and minimizing energy consumption, a multi-objective reward function is designed by combining the safety correction amount and environmental adjustment differences to determine the optimal control command, and the intelligent decision model is updated accordingly.

[0007] Furthermore, the actuators in the controlled greenhouse specifically include heaters, coolers, humidifiers, and dehumidifiers; The temperature dynamic model of the greenhouse to be controlled is specifically represented as follows: In the formula, Indicates in The internal temperature of the greenhouse needs to be controlled at all times. This indicates the temperature inside the greenhouse to be controlled at time k. The energy conversion coefficient of the heater. This is the heater control input at time k. The energy conversion coefficient of the refrigerator. This is the cooler control input at time k. The greenhouse heat exchange coefficient, The ambient temperature, The noise in the temperature model at time k is specifically the noise value that follows a Gaussian distribution, where k is the time variable of a day in 24-hour format. It is a quantity that changes over time; The humidity dynamic model of the greenhouse to be controlled is specifically represented as follows: In the formula, and express The relative humidity inside the greenhouse to be controlled at time k and time k. The humidifier's moisture conversion coefficient. The humidifier control input at time k is... This is the moisture conversion coefficient of the dehumidifier. The dehumidifier control input at time k is... The greenhouse moisture exchange coefficient, The humidity of the external environment, The noise of the humidity model at time k is specifically the noise value that follows a Gaussian distribution.

[0008] Furthermore, based on the dynamics of solar radiation and crop transpiration, an external environmental disturbance term is set, which includes dynamic disturbances of solar radiation and disturbances of crop transpiration. The specific logic is as follows: the dynamic disturbance of solar radiation is comprehensively characterized by solar radiation intensity and cloud cover, specifically expressed as follows: In the formula, Let K be the dynamic perturbation of solar radiation at time k. The solar radiation absorption coefficient. Let be the solar radiation intensity at time k; where the solar radiation intensity at time k is... The specific calculation method is as follows: In the formula, This is the theoretical value of the maximum solar radiation intensity. and These are sunrise and sunset times, respectively. This is the cloud cover coefficient. Let k be the cloud coverage rate at time k. Crop transpiration disturbance is determined based on an empirical model, specifically using the following formula: In the formula, For crop transpiration disturbance, This is the crop transpiration coefficient. Let be the crop transpiration rate at time k, where is the crop transpiration rate at time k. The specific formula used for the calculation is as follows: In the formula, This is the theoretical value of the maximum transpiration rate. Let be the response function of the temperature, humidity, and light intensity of the greenhouse to be controlled at time k to the transpiration rate. Factors related to crop growth stages; The external environmental disturbance term is coupled with the greenhouse temperature dynamic model and the greenhouse humidity dynamic model to obtain the accurate temperature dynamic model and the accurate humidity dynamic model. This establishes the mapping relationship between temperature, humidity, and actuator input data. The specific formula used is as follows: In the formula, This represents the precise internal temperature of the greenhouse at time k+1. The precise relative humidity inside the greenhouse at time k+1; The constraints of the greenhouse operating conditions are set. Within these constraints, temperature and humidity simulation data generated over multiple consecutive time steps are output using precise temperature and humidity dynamic models. Specifically, the constraints of the greenhouse operating conditions include: In the formula, and These are the minimum and maximum greenhouse temperature thresholds, respectively. and These are the minimum and maximum relative humidity thresholds, respectively.

[0009] Furthermore, the training data includes temperature and relative humidity data of the greenhouse to be controlled at several times, as well as the corresponding actuator input data; The specific logic behind performing interval discretization on the training data is as follows: based on the set greenhouse operating condition constraints, determine the temperature constraint interval and relative humidity constraint interval of the greenhouse to be controlled. For the temperature constraint range of the greenhouse to be controlled, the temperature constraint range is evenly divided with a preset temperature adjustment interval to form several temperature ranges of the greenhouse to be controlled. According to the number of temperature ranges, the relative humidity adjustment interval is set to form a corresponding number of relative humidity ranges. The temperature and relative humidity simulation data in the training data are mapped to the corresponding discrete ranges to obtain the intervalized training data of temperature and relative humidity data in the training data. The specific steps for performing interval discretization on actuator input data include: discretizing each actuator input data into 7 levels, which correspond to the control levels from off to maximum power, where level 0 power control level represents the actuator being off, level 6 power control level represents the actuator starting at maximum power, forming a complete action space with the Cartesian product of the four actuators, and determining several actuator state vectors in the action space; Using the current temperature range, relative humidity range, and actuator state vector as input, and the temperature range and relative humidity range corresponding to the time variable as output, a range mapping model is trained to construct a range mapping relationship.

[0010] Furthermore, the intelligent decision-making model is built based on the standard Q-learning algorithm; The environmental target values ​​specifically include the temperature target value and the relative humidity target value of the greenhouse to be controlled. The logic based on the interval mapping relationship analysis of the current temperature interval and the current relative humidity interval is as follows: input the current temperature interval and the current relative humidity interval, as well as the state vector of each actuator, into the interval mapping model that has been trained to obtain the current temperature interval and the current relative humidity interval. Under any actuator state vector, the temperature interval and the relative humidity interval predicted after the corresponding time variable. The specific method for selecting the preferred actuator state vector is as follows: the deviation between the predicted temperature range and relative humidity range and the environmental target value is used as the evaluation benchmark, and a deviation threshold is set. The actuator state vectors corresponding to the deviation between the predicted temperature range and relative humidity range and the environmental target value is not greater than the deviation threshold are marked as preferred actuator state vectors. The specific logic for determining the deviation between the predicted temperature range and relative humidity range and the environmental target value is as follows: if the environmental target value falls within the predicted temperature range and relative humidity range, then the deviation between the predicted temperature range and relative humidity range and the environmental target value is marked as 0. For the temperature target value of the greenhouse to be controlled in the environmental target value, if the temperature target value of the greenhouse to be controlled is less than the predicted minimum temperature range, the relative error between the predicted minimum temperature range and the temperature target value of the greenhouse to be controlled is taken as the degree of temperature deviation; if the temperature target value of the greenhouse to be controlled is greater than the predicted maximum temperature range, the relative error between the temperature target value of the greenhouse to be controlled and the predicted maximum temperature range is taken as the degree of temperature deviation. For the relative humidity of the greenhouse to be controlled, a target value is set. If the target value is less than the predicted minimum relative humidity range, the relative error between the predicted minimum relative humidity range and the target value is taken as the degree of relative humidity deviation. If the target value is greater than the predicted maximum relative humidity range, the relative error between the target value and the predicted maximum relative humidity range is taken as the degree of relative humidity deviation. The deviation of the predicted temperature range and relative humidity range from the environmental target value is defined as the maximum value of the relative humidity deviation and the temperature deviation.

[0011] Furthermore, temperature conflict and relative humidity conflict are detected sequentially on the preferred actuator state vector, and the preferred actuator state vector is further filtered according to the corresponding conflict. The specific method used for temperature conflict detection is as follows: if the difference between the median value of the current ambient temperature range and the target value of the greenhouse temperature to be controlled is greater than... If the absolute difference between the control levels of the heater and the cooler is greater than the control level difference threshold, then the temperature constraint condition is used. Based on the temperature constraint condition, all preferred actuator state vectors are screened a second time. If the temperature constraint condition is met, or the heater power control level in the preferred actuator state vector is level 0, then the preferred actuator state vector is marked. If the difference between the median value of the current ambient temperature range and the target value of the greenhouse temperature to be controlled is less than... If the absolute difference between the control levels of the heater and the cooler is greater than the control level difference threshold, then the temperature constraint condition is used. Based on the temperature constraint condition, all preferred actuator state vectors are screened a second time. If the temperature constraint condition is met, or the cooler power control level in the preferred actuator state vector is level 0, then the preferred actuator state vector is marked. If the difference between the median value of the current ambient temperature range and the target value of the greenhouse temperature to be controlled falls within the range... Within the interval, the preferred actuator state vectors for heaters and coolers with control levels both less than 1 are marked; The marked preferred actuator state vector is subjected to relative humidity conflict detection, specifically including: if the difference between the median value of the current relative humidity range and the target value of the relative humidity of the greenhouse to be controlled is greater than 100%, then the conflict is considered to be caused by a conflict. If the absolute difference between the control levels of the humidifier and the dehumidifier is greater than the control level difference threshold, then the humidity constraint condition is used. Based on the humidity constraint condition, the marked preferred actuator state vector is screened a second time. If the humidity constraint condition is met, or the humidifier power control level in the preferred actuator state vector is level 0, then the marked preferred actuator state vector is marked a second time. If the difference between the median value of the current relative humidity range and the target value of the relative humidity of the greenhouse to be controlled is less than... If the absolute difference between the control levels of the humidifier and the dehumidifier is greater than the control level difference threshold, then the humidity constraint condition is used. Based on the humidity constraint condition, the marked preferred actuator state vector is screened a second time. If the humidity constraint condition is met, or the dehumidifier power control level in the preferred actuator state vector is level 0, then the marked preferred actuator state vector is marked a second time. If the difference between the median value of the current relative humidity range and the target value of the relative humidity of the greenhouse to be controlled falls within the range... Within the range, the state vectors of preferred actuators with control levels of both humidifier and dehumidifier less than 1 are marked a second time. Based on the preferred actuator state vector that has been secondary-labeled, a set of optional actuator state vectors is constructed; The difference adjustment thresholds include a temperature difference adjustment threshold and a relative humidity difference adjustment threshold, wherein... Adjusting the threshold for temperature differences, The threshold for adjusting relative humidity difference.

[0012] Furthermore, the specific formula used to calculate the safety correction amount of the optional actuator state vector is as follows: In the formula, This is the safety correction amount for the v-th optional executor state vector. Let v be the state vector of the optional executor. This is the current executor state vector. This represents the Euclidean distance between the v-th optional executor state vector and the current executor state vector, where v is the index of the optional executor state vector; The specific formula upon which the multi-objective reward function is designed is as follows: In the formula, Let v be the multi-objective reward function for the v-th optional executor state vector. The energy-saving reward term for the v-th optional actuator state vector. The tracking reward term for the v-th optional executor state vector. , and Here are the weighting coefficients, where and , and All are greater than 0.

[0013] Furthermore, the energy-saving reward term for the v-th optional actuator state vector The specific logic underlying the calculation is as follows: In the formula, Let j be the energy consumption coefficient of the j-th actuator in the v-th optional actuator state vector. Let j be the power control level of the j-th actuator in the state vector of the v-th selectable actuator; The tracking reward term for the v-th optional executor state vector The formula used for the calculation is: In the formula, The median of the temperature range of the greenhouse to be controlled after executing the v-th optional actuator state vector. The median of the relative humidity interval of the greenhouse to be controlled after executing the v-th optional actuator state vector. and Target values ​​are set for temperature and relative humidity, respectively. The specific temperature range and relative humidity range of the greenhouse observation value after the vth optional actuator state vector are determined by the interval mapping model. Traverse the state vector of the preferred executor without safety correction and all optional executor state vectors, and determine the multi-objective reward function value for any preferred executor state vector without safety correction and any optional executor state vector; Based on the multi-objective reward function, the optimal control command is selected from the Q-values ​​of the selectable actuator state vectors in the Bellman equation, and the intelligent decision model is updated accordingly.

[0014] Compared with the prior art, the beneficial effects of the present invention are: By integrating rule constraints and reinforcement learning into a multi-objective intelligent control method for greenhouse climate, this approach effectively solves the problems of neglecting coupling relationships and control conflicts in traditional temperature and humidity control methods. Through a dynamic temperature and humidity model built based on the law of conservation of energy, this scheme can more accurately simulate climate change inside the greenhouse and make real-time adjustments according to external environmental disturbances. This coupled model not only improves the accuracy of environmental data but also makes the comprehensive control of temperature and humidity more precise and reasonable, fundamentally reducing energy waste and unnecessary costs. Furthermore, this solution employs reinforcement learning technology combined with a multi-objective reward function to achieve intelligent decision-making regarding greenhouse climate. By dynamically adjusting the set of actionable constraints and calculating safety corrections, the system can respond to environmental changes in real time, optimize control commands, ensure coordinated operation between actuators, and avoid unreasonable behaviors such as "heating and cooling simultaneously" or "humidifying and dehumidifying simultaneously." This intelligent decision-making method not only improves crop growth efficiency and product quality in the greenhouse but also significantly reduces energy consumption, promotes the achievement of sustainable agricultural development goals, and provides a new solution for the advancement of modern agricultural technology. Attached Figure Description

[0015] Figure 1 This is a schematic diagram of the overall method flow of the present invention; Figure 2 A comparison chart of temperature and humidity control in intelligent decision-making models; Figure 3 A comparison chart of total energy consumption controlled by intelligent decision-making models; Figure 4 This is a comparison chart of control errors in intelligent decision-making models. Detailed Implementation

[0016] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0017] It should be noted that, unless otherwise defined, the technical or scientific terms used in this invention should have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "first," "second," and similar terms used in this invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed following the word and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0018] Example: Please see Figures 1-4 The present invention provides a technical solution: A multi-objective intelligent control method for greenhouse climate that integrates rule constraints and reinforcement learning, comprising the following steps: Step 1: Construct a temperature dynamic model and a humidity dynamic model of the greenhouse to be controlled based on the law of conservation of energy. According to the dynamics of solar radiation and crop transpiration, set an external environmental disturbance term and couple the external environmental disturbance term with the temperature dynamic model and the humidity dynamic model to establish a mapping relationship between temperature, relative humidity and actuator input data.

[0019] The actuators in the controlled greenhouse specifically include heaters, coolers, humidifiers, and dehumidifiers; The temperature dynamic model of the greenhouse to be controlled is specifically represented as follows: In the formula, Indicates in The internal temperature of the greenhouse needs to be controlled at all times. This indicates the temperature inside the greenhouse to be controlled at time k. The energy conversion coefficient of the heater. This refers to the heater control input at time k, specifically the heater power control input. The energy conversion coefficient of the refrigerator. This refers to the cooler control input at time k, specifically the cooler power control input. The greenhouse heat exchange coefficient, The ambient temperature, Let k be the process noise of the temperature model at time k, specifically the noise value following a Gaussian distribution, where k is the time variable of a day in 24-hour format. It is a quantity that changes over time; It should be noted that, This is used to represent the temperature inside the controlled greenhouse at the next time interval, where an increase in the heater control input power directly leads to an increase in the greenhouse temperature. The increase, the increase This will make Increase, thereby improve ; Increasing the power of the refrigeration unit will have a negative impact on the internal temperature of the greenhouse to be controlled, leading to a decrease in greenhouse temperature. The decrease, the increase Increase negative items Thus reducing ; Changes in the external ambient temperature affect heat exchange inside the greenhouse. Higher than At that time, heat exchange term The result is positive, leading to An increase; conversely, a decrease leads to a decrease in the internal temperature of the greenhouse to be controlled.

[0020] Process noise This term represents the influence of uncontrollable factors in the model; it is a random variable that can affect temperature positively or negatively. However, its impact is random and uncertain, and is usually assumed to be noise values ​​following a Gaussian distribution, process noise. It is used to represent uncontrollable random factors in the model, such as changes in equipment efficiency or sudden changes in the external environment, thereby increasing the robustness of the model. The specific method for obtaining the heater energy conversion coefficient is as follows: the coefficient can be calculated by actually measuring the heat output of the heater under different power inputs. Usually, it is necessary to conduct experiments in a controlled environment, record the power input of the heater and the corresponding temperature changes, or obtain its performance curve or energy conversion efficiency by consulting the heater's technical specifications. The specific method for obtaining the energy conversion coefficient of a refrigerator is as follows: similar to that of a heater, the performance of a refrigerator can also be measured experimentally. It is necessary to record the power input of the refrigerator and the change in room temperature to calculate the coefficient. The specific methods for obtaining the greenhouse heat exchange coefficient are as follows: the heat exchange coefficient can be estimated through heat conduction experiments, usually by combining measurements of the greenhouse's materials, structure, and ambient temperature, or by using heat transfer models, such as finite element analysis, for numerical simulation, which can help estimate the coefficient; The specific method for obtaining the process noise of the temperature model at time k is as follows: by statistically analyzing the temperature data, the random fluctuations in the temperature change process are evaluated, thereby obtaining the characteristics of the noise, including the mean and variance. When building the model, the error between the model output and the actual observation data can be minimized by adjusting the parameters of the process noise. Commonly used methods include Kalman filtering.

[0021] The humidity dynamic model of the greenhouse to be controlled is specifically represented as follows: In the formula, and express The relative humidity inside the greenhouse to be controlled at time k and time k. The humidifier's moisture conversion coefficient. The humidifier control input at time k specifically refers to the humidifier power control input. This refers to the moisture conversion coefficient of the dehumidifier. The dehumidifier control input at time k specifically refers to the dehumidifier power control input. The greenhouse moisture exchange coefficient, The humidity of the external environment, The noise of the humidity model at time k is specifically the noise value that follows a Gaussian distribution.

[0022] The specific methods for obtaining the dehumidifier moisture conversion coefficient, greenhouse moisture exchange coefficient, humidifier moisture conversion coefficient, and humidity model process noise at time k are the same as those described above, and will not be repeated here.

[0023] In dynamic models of greenhouse internal temperature, the influence of external environmental factors, particularly solar radiation and crop transpiration, needs to be considered. These factors significantly affect the heat balance inside the greenhouse. Based on the dynamics of solar radiation and crop transpiration, an external environmental disturbance term is set, which includes dynamic disturbances of solar radiation and disturbances of crop transpiration. The specific logic is as follows: the dynamic disturbance of solar radiation is comprehensively characterized by solar radiation intensity and cloud cover, specifically expressed as follows: In the formula, Let K be the dynamic perturbation of solar radiation at time k. The solar radiation absorption coefficient. Let k be the solar radiation intensity at time k. It should be noted that the dynamic perturbation of solar radiation at time k Specifically used to represent the effect of solar radiation on the control of the internal temperature of a greenhouse, when A larger value indicates that the greenhouse will absorb more heat, resulting in an increase in internal temperature. Solar radiation is the primary external heat source for the greenhouse's internal temperature. The amount of heat entering the greenhouse varies with changes in solar radiation intensity, and different materials have varying absorption efficiencies for solar radiation. Solar radiation absorption coefficient. solar radiation intensity Converted to the actual impact on the internal temperature of the greenhouse, this coefficient takes into account the characteristics of the greenhouse materials and their response to radiation, ensuring that the formula accurately reflects the actual situation.

[0024] Where the solar radiation intensity at time k The specific calculation method is as follows: In the formula, This is the theoretical value of the maximum solar radiation intensity. and These are sunrise and sunset times, respectively. This is the cloud cover coefficient. Let k be the cloud coverage rate at time k. It should be noted that the impact of cloud cover on solar radiation is common knowledge in meteorology. The intensity of solar radiation is corrected by using the shading coefficient and coverage rate. The theoretical value of the maximum solar radiation intensity is usually measured under the condition of no clouds and the strongest sunshine, which reflects the radiation intensity under ideal conditions. The cloud cover coefficient represents the degree to which clouds affect solar radiation. This coefficient reflects the thickness and type of clouds, influencing the amount of radiation transmitted through them. The cloud cover rate at time k is also relevant. Indicates at time The cloud cover rate, with a value between 0 and 1, where 0 means no clouds at all and 1 means completely covered by clouds. In the formula Part of it is used to simulate the variation of solar radiation intensity during the day. As time progresses, the sun's altitude angle changes between sunrise and sunset, therefore the radiation intensity exhibits a typical sine wave shape. At that time, the sun had just risen and the radiation intensity was 0. At noon, when the sun's altitude angle is highest, the radiation intensity reaches its maximum. At that time, the sun sets, and the radiation intensity is 0 again; In the formula Part of it is used to correct the reduction in radiation intensity caused by cloud cover, when the cloud coverage is high. When the shielding effect increases, the radiation intensity decreases accordingly. This correction term is achieved through... Adjustments were made to ensure that the formula reflects the impact of actual meteorological conditions on solar radiation intensity. The specific method for determining the cloud obstruction coefficient is as follows: using professional equipment such as a radiation meter or solar radiometer, the radiation intensity received by the ground is measured under different cloud conditions, and then compared with the theoretical maximum radiation intensity to calculate the proportion of radiation obstruction by the cloud. In the experiment, the obstruction effect of different cloud types, including cirrus, cumulonimbus, and stratus, on solar radiation is measured step by step to determine the cloud obstruction coefficient value of each type of cloud. The specific methods for determining cloud coverage are as follows: cloud coverage of the target area can be obtained through meteorological satellite remote sensing observation, or real-time meteorological data can be obtained through meteorological service platforms, such as the China Meteorological Administration website or the International Meteorological Service System. Crop transpiration disturbance is determined based on an empirical model, specifically using the following formula: In the formula, For crop transpiration disturbance, This is the crop transpiration coefficient. Let k be the crop transpiration rate at time k; It should be noted that crop transpiration disturbance is the process by which plants evaporate water from the leaf surface into the environment through stomata. It is an important mechanism for normal plant growth and water regulation. Used to represent the impact of crop transpiration on environmental humidity; crop transpiration disturbance The higher the value, the greater the relative humidity of the environment that increases crop transpiration; The crop transpiration coefficient represents the transpiration efficiency of a crop under specific environmental conditions. The specific method for determining it is as follows: by measuring the conductance of the crop's stomata, the transpiration capacity of the plant is determined. This requires the use of gas exchange instruments to measure the gas exchange of the plant under certain environmental conditions. Under controlled conditions, the transpiration rate is calculated by quantitatively measuring the water lost by the plant over a certain period of time.

[0025] Where the crop transpiration rate at time k The specific formula used for the calculation is as follows: In the formula, This is the theoretical value of the maximum transpiration rate. Let be the response function of the temperature, humidity, and light intensity of the greenhouse to be controlled at time k to the transpiration rate. Factors related to crop growth stages; It should be noted that temperature has a direct impact on plant physiological activities. High temperatures usually promote transpiration, but extremely high temperatures may cause stomatal closure, thereby inhibiting transpiration. Therefore, using temperature response can more accurately reflect the effect of temperature on transpiration rate. Humidity affects the water vapor content in the air, which in turn affects the water loss of plants. Selective humidity response can effectively reflect the regulatory effect of air moisture content on transpiration rate. The higher the light intensity, the stronger the photosynthesis of plants, and the more stomata tend to open, thus increasing the transpiration rate. Using light response can better quantify the impact of light on transpiration. The specific method for obtaining the response function of temperature, humidity, and light intensity to transpiration rate in a controlled greenhouse is as follows: Under controlled conditions, experiments are conducted using a gas exchange instrument. The transpiration rate of the crop is recorded using gas exchange data under different temperature, humidity, and light conditions. The experimental data are then statistically analyzed, and regression analysis or a machine learning model is used to fit the data to obtain the response function. .

[0026] The growth stages of crops, including the seedling stage, vegetative growth stage, flowering stage, and maturity stage, affect their water requirements and transpiration characteristics. The physiological characteristics of plants are different at each growth stage, such as stomatal conductance and photosynthetic efficiency, which directly affect the transpiration rate. As plants grow, stomatal conductance generally increases, and transpiration capacity increases accordingly. Therefore, introducing crop growth stage factors can more accurately reflect the transpiration efficiency of crops at different stages. The specific crop growth stage factor values ​​can be determined by consulting relevant materials to determine the trend of transpiration rate changes.

[0027] By taking into account the effects of temperature, humidity, light, and growth stage, the maximum transpiration rate can be reflected in the true transpiration behavior of plants under different conditions.

[0028] The external environmental disturbance term is coupled with the greenhouse temperature dynamic model and the greenhouse humidity dynamic model to obtain the accurate temperature dynamic model and the accurate humidity dynamic model. This establishes the mapping relationship between temperature, humidity, and actuator input data. The specific formula used is as follows: In the formula, for Precise temperature inside the greenhouse at any given moment. for Precise relative humidity inside the greenhouse at any given time; The constraints of the greenhouse operating conditions are set. Within these constraints, temperature and humidity simulation data generated over multiple consecutive time steps are output using precise temperature and humidity dynamic models. Specifically, the constraints of the greenhouse operating conditions include: In the formula, and These are the minimum and maximum greenhouse temperature thresholds, respectively. and These are the minimum and maximum relative humidity thresholds, respectively.

[0029] To ensure the reasonableness of the generated data, constraints were set that conform to the actual operating conditions of the greenhouse: temperature ,humidity ,Right now for , for , for , for This constraint processing ensures the physical rationality of the generated data, enabling it to better reflect the operational characteristics of a real greenhouse.

[0030] Step 2: Output multiple sets of training data based on the mapping relationship, and perform interval discretization operation on the training data to generate multiple sets of interval-based training data and construct interval-based mapping relationship. The interval-based training data includes temperature interval, relative humidity interval, and actuator state vector composed of each actuator control level.

[0031] The training data includes temperature and relative humidity data of the greenhouse to be controlled at several times, as well as the corresponding actuator input data; The specific logic behind performing interval discretization on the training data is as follows: based on the set greenhouse operating condition constraints, determine the temperature constraint interval and relative humidity constraint interval of the greenhouse to be controlled. For the temperature constraint range of the greenhouse to be controlled, the temperature constraint range is evenly divided with a preset temperature adjustment interval to form several temperature ranges of the greenhouse to be controlled. According to the number of temperature ranges, the relative humidity adjustment interval is set to form a corresponding number of relative humidity ranges. The temperature and relative humidity simulation data in the training data are mapped to the corresponding discrete ranges to obtain the intervalized training data of temperature and relative humidity data in the training data. The temperature adjustment interval is set as follows: The relative humidity adjustment interval is This choice of discretization granularity is based on an in-depth analysis of greenhouse control requirements: The temperature accuracy can meet the growth requirements of most crops, while The relative humidity accuracy ensures control effectiveness while avoiding excessive expansion of the state space. The discretized state space retains necessary environmental feature information while keeping the size of the intelligent decision-making model within a trainable range, laying a solid foundation for subsequent policy learning.

[0032] The specific steps for performing interval discretization on actuator input data include: discretizing each actuator input data into 7 levels, which correspond to the control levels from off to maximum power, where level 0 power control level represents the actuator being off, level 6 power control level represents the actuator starting at maximum power, forming a complete action space with the Cartesian product of the four actuators, and determining several actuator state vectors in the action space; This discretization design maintains sufficient control precision while keeping the motion space within the range that the algorithm can handle. By pre-computing and generating a complete motion lookup table, it significantly improves the efficiency of online decision-making and provides technical support for real-time control.

[0033] Using the current temperature range, relative humidity range, and actuator state vector as input, and the temperature and relative humidity ranges corresponding to the time variables as output, an interval mapping model is trained to construct an interval mapping relationship. Specifically, the interval mapping model is constructed using a convolutional neural network model. The specific training steps include: collecting relevant input data for training, including the current temperature range, relative humidity range, and actuator state vector; simultaneously recording the temperature and relative humidity ranges corresponding to the time variables as output data; to simultaneously optimize the upper and lower bounds or center values ​​and amplitudes of the output intervals, mean squared error or mean absolute error is used to minimize the error between the predicted upper and lower bounds and the true values; and selecting a suitable optimizer, such as Adam or SGD, to optimize the model hyperparameters, including adjusting the learning rate, convolutional kernel size, network depth, batch size, and other hyperparameters to optimize model performance.

[0034] Step 3: Construct an intelligent decision-making model. Using the deviation between the environmental target value and the range as the evaluation benchmark, the model analyzes the current temperature range and the current relative humidity range based on the interval mapping relationship. It determines the evaluation index of each actuator state vector and selects the preferred actuator state vector. It sets the difference adjustment threshold for the mutual exclusion operation of each actuator and performs a second screening based on the difference adjustment threshold to form the optional actuator state vector.

[0035] The intelligent decision-making model is built based on the standard Q-learning algorithm; The environmental target values ​​specifically include the temperature target value and the relative humidity target value of the greenhouse to be controlled. The logic based on the interval mapping relationship analysis of the current temperature interval and the current relative humidity interval is as follows: input the current temperature interval and the current relative humidity interval, as well as the state vector of each actuator, into the interval mapping model that has been trained to obtain the current temperature interval and the current relative humidity interval. Under any actuator state vector, the temperature interval and the relative humidity interval predicted after the corresponding time variable. The specific method for selecting the preferred actuator state vector is as follows: the deviation between the predicted temperature range and relative humidity range and the environmental target value is used as the evaluation benchmark, and a deviation threshold is set. The actuator state vectors corresponding to the deviation between the predicted temperature range and relative humidity range and the environmental target value is not greater than the deviation threshold are marked as preferred actuator state vectors. The specific logic for determining the deviation between the predicted temperature range and relative humidity range and the environmental target value is as follows: if the environmental target value falls within the predicted temperature range and relative humidity range, then the deviation between the predicted temperature range and relative humidity range and the environmental target value is marked as 0. For the temperature target value of the greenhouse to be controlled in the environmental target value, if the temperature target value of the greenhouse to be controlled is less than the predicted minimum temperature range, the relative error between the predicted minimum temperature range and the temperature target value of the greenhouse to be controlled is taken as the degree of temperature deviation; if the temperature target value of the greenhouse to be controlled is greater than the predicted maximum temperature range, the relative error between the temperature target value of the greenhouse to be controlled and the predicted maximum temperature range is taken as the degree of temperature deviation. For the relative humidity of the greenhouse to be controlled, a target value is set. If the target value is less than the predicted minimum relative humidity range, the relative error between the predicted minimum relative humidity range and the target value is taken as the degree of relative humidity deviation. If the target value is greater than the predicted maximum relative humidity range, the relative error between the target value and the predicted maximum relative humidity range is taken as the degree of relative humidity deviation. The deviation of the predicted temperature range and relative humidity range from the environmental target value is defined as the maximum value of the relative humidity deviation and the temperature deviation.

[0036] The deviation threshold is set based on expert experience, and is generally set between 0.01 and 0.1.

[0037] By detecting the relationship between environmental target values ​​and prediction intervals, the current execution effect can be effectively judged, and then the optimal actuator state can be selected. This can reduce the deviation of environmental target values, achieve more precise temperature and humidity control, and improve the growth environment of greenhouse crops. By optimizing the actuator state, unnecessary operations can be effectively reduced, the energy utilization efficiency of the system can be improved, and the operating cost of the greenhouse can be reduced.

[0038] Temperature conflict and relative humidity conflict detection are performed sequentially on the preferred actuator state vector, and the preferred actuator state vector is further filtered according to the corresponding conflict. The specific method used for temperature conflict detection is as follows: if the difference between the median value of the current ambient temperature range and the target value of the greenhouse temperature to be controlled is greater than... If the absolute difference between the control levels of the heater and the cooler is greater than the control level difference threshold, then the temperature constraint condition is used. Based on the temperature constraint condition, all preferred actuator state vectors are screened a second time. If the temperature constraint condition is met, or the heater power control level in the preferred actuator state vector is level 0, then the preferred actuator state vector is marked. If the difference between the median value of the current ambient temperature range and the target value of the greenhouse temperature to be controlled is less than... If the absolute difference between the control levels of the heater and the cooler is greater than the control level difference threshold, then the temperature constraint condition is used. Based on the temperature constraint condition, all preferred actuator state vectors are screened a second time. If the temperature constraint condition is met, or the cooler power control level in the preferred actuator state vector is level 0, then the preferred actuator state vector is marked. If the difference between the median value of the current ambient temperature range and the target value of the greenhouse temperature to be controlled falls within the range... Within the interval, the preferred actuator state vectors for heaters and coolers with control levels both less than 1 are marked; The marked preferred actuator state vector is subjected to relative humidity conflict detection, specifically including: if the difference between the median value of the current relative humidity range and the target value of the relative humidity of the greenhouse to be controlled is greater than 100%, then the conflict is considered to be caused by a conflict. If the absolute difference between the control levels of the humidifier and the dehumidifier is greater than the control level difference threshold, then the humidity constraint condition is used. Based on the humidity constraint condition, the marked preferred actuator state vector is screened a second time. If the humidity constraint condition is met, or the humidifier power control level in the preferred actuator state vector is level 0, then the marked preferred actuator state vector is marked a second time. If the difference between the median value of the current relative humidity range and the target value of the relative humidity of the greenhouse to be controlled is less than... If the absolute difference between the control levels of the humidifier and the dehumidifier is greater than the control level difference threshold, then the humidity constraint condition is used. Based on the humidity constraint condition, the marked preferred actuator state vector is screened a second time. If the humidity constraint condition is met, or the dehumidifier power control level in the preferred actuator state vector is level 0, then the marked preferred actuator state vector is marked a second time. If the difference between the median value of the current relative humidity range and the target value of the relative humidity of the greenhouse to be controlled falls within the range... Within the range, the state vectors of preferred actuators with control levels of both humidifier and dehumidifier less than 1 are marked a second time. Based on the preferred actuator state vector that has been secondary-labeled, a set of optional actuator state vectors is constructed; The difference adjustment thresholds include a temperature difference adjustment threshold and a relative humidity difference adjustment threshold, wherein... Adjusting the threshold for temperature differences, The relative humidity difference adjustment threshold; The difference adjustment thresholds include a temperature difference adjustment threshold and a relative humidity difference adjustment threshold, wherein... The temperature difference adjustment threshold is generally set to [value]. - Between; relative humidity difference adjustment threshold Normal settings - between; Step 4: Based on the safety correction amount of the state vector of each selectable actuator, and with the optimization objectives of minimizing environmental adjustment differences and minimizing energy consumption, design a multi-objective reward function in combination with the safety correction amount and environmental adjustment differences, determine the optimal control command, and update the intelligent decision model accordingly.

[0039] The specific formula used to calculate the safety correction amount of the optional actuator state vector is as follows: In the formula, This is the safety correction amount for the v-th optional executor state vector. Let v be the state vector of the optional executor. This is the current executor state vector. This represents the Euclidean distance between the v-th optional executor state vector and the current executor state vector, where v is the index of the optional executor state vector; It should be noted that, This represents the Euclidean distance between the v-th optional actuator state vector and the current actuator state vector, specifically the Euclidean distance between the optional actuator state vector and each power control level in the current actuator state vector; The specific formula upon which the multi-objective reward function is designed is as follows: In the formula, Let v be the multi-objective reward function for the v-th optional executor state vector. The energy-saving reward term for the v-th optional actuator state vector. The tracking reward term for the v-th optional executor state vector. , and Here are the weighting coefficients, where and , and All are greater than 0.

[0040] It should be noted that the multi-objective reward function of the v-th optional executor state vector Used to represent the priority of the v-th optional executor state vector, The larger the value, the higher the priority of the v-th optional executor state vector; Safety is the most important objective, carrying the largest weight in the reward function to ensure the system's safety and stability during operation. Tracking accuracy is secondary, reflecting the system's need to maintain good tracking capabilities to cope with dynamic environments during action execution. Although energy efficiency is important, it may be prioritized over safety and performance requirements in certain situations, hence the setting... and , and All are greater than 0.

[0041] Energy-saving reward term for the v-th optional actuator state vector The specific logic underlying the calculation is as follows: In the formula, Let j be the energy consumption coefficient of the j-th actuator in the v-th optional actuator state vector. Let j be the power control level of the j-th actuator in the state vector of the v-th selectable actuator; It should be noted that, It is the first The energy-saving reward term for each optional actuator state vector is defined as the negative value of energy consumption, reflecting the energy cost of the action. That is, the higher the energy consumption, the lower the reward term and the greater the penalty. The tracking reward term for the v-th optional executor state vector The formula used for the calculation is: In the formula, The median of the temperature range of the greenhouse to be controlled after executing the v-th optional actuator state vector. The median of the relative humidity interval of the greenhouse to be controlled after executing the v-th optional actuator state vector. and Target values ​​are set for temperature and relative humidity, respectively. The specific temperature range and relative humidity range of the greenhouse observation value after the vth optional actuator state vector are determined by the interval mapping model. It should be noted that, Indicates the first The tracking reward term for each optional actuator state vector is designed to evaluate how close the observed temperature and humidity values ​​in the greenhouse are to the target setpoints after executing the instruction. The larger the value, the more likely the execution of the [number]th [stage]. After selecting the actuator state vectors, the closer the temperature and relative humidity inside the greenhouse to be controlled are to the target values; Using an exponential decay function The reward is represented in a form that allows even small deviations to produce significant changes in reward. This sensitivity helps to respond quickly to minute changes in the greenhouse environment, thereby promoting better control strategies, particularly for the regulation of relative humidity. Similarly.

[0042] Based on the multi-objective reward function, the Q value of the optional actuator state vector is determined by the Bellman equation. Specifically, the instantaneous reward value in the Bellman equation is replaced with the multi-objective reward function, the optional actuator state vector corresponding to the maximum Q value is selected as the optimal control command, and the intelligent decision model is updated accordingly. Multi-objective reward functions can quantify the trade-offs between different objectives into a single reward value, thereby guiding intelligent decision-making models to find the optimal solution in different actuator state vectors. In reinforcement learning, the Q-value table reflects the system's gains under different states and actions. Selecting the control instruction corresponding to the maximum Q-value and updating the intelligent decision-making model helps the system gradually learn the best decision-making strategy. At the same time, updating the intelligent decision-making model enables it to directly retrieve the Q-value table and select the optimal control instruction when encountering the same adjustment objective. Since the Bellman equation is a common existing formula, it will not be elaborated here.

[0043] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0044] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented in software, the above embodiments can be implemented, in whole or in part, as a computer program product. Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented by electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution.

[0045] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment, depending on actual needs.

[0046] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A multi-objective intelligent control method for greenhouse climate that integrates rule constraints and reinforcement learning, characterized in that, The specific steps include: Based on the law of conservation of energy, a temperature dynamic model and a humidity dynamic model of the greenhouse to be controlled are constructed. According to the dynamics of solar radiation and crop transpiration, an external environmental disturbance term is set and coupled with the temperature dynamic model and the humidity dynamic model to establish a mapping relationship between temperature, relative humidity and actuator input data. Based on the mapping relationship, multiple sets of training data are output, and the training data are discretized into intervals to generate multiple sets of interval training data and construct interval mapping relationships. The interval training data includes temperature intervals, relative humidity intervals, and actuator state vectors composed of each actuator control level. An intelligent decision-making model is constructed. The deviation between the environmental target value and the range is used as the evaluation benchmark. Based on the range mapping relationship, the current temperature range and the current relative humidity range are analyzed to determine the evaluation index of each actuator state vector and select the preferred actuator state vector. The difference adjustment threshold of the mutual exclusion operation of each actuator is set. Based on the difference adjustment threshold, the preferred actuator state vector is screened again to form the selectable actuator state vector. Based on the safety correction amount of the state vector of each selectable actuator, and with the optimization objectives of minimizing environmental adjustment differences and minimizing energy consumption, a multi-objective reward function is designed by combining the safety correction amount and environmental adjustment differences to determine the optimal control command, and the intelligent decision model is updated accordingly. The logic for the secondary screening of the preferred actuator state vector is as follows: temperature conflict and relative humidity conflict are detected sequentially for the preferred actuator state vector, and the preferred actuator state vector is screened secondaryly based on the corresponding conflicts. The specific method used for temperature conflict detection is as follows: if the difference between the median value of the current ambient temperature range and the target value of the greenhouse temperature to be controlled is greater than... If the absolute difference between the control levels of the heater and the cooler is greater than the control level difference threshold, then the temperature constraint condition is used. Based on the temperature constraint condition, all preferred actuator state vectors are screened a second time. If the temperature constraint condition is met, or the heater power control level in the preferred actuator state vector is level 0, then the preferred actuator state vector is marked. If the difference between the median value of the current ambient temperature range and the target value of the greenhouse temperature to be controlled is less than... If the absolute difference between the control levels of the heater and the cooler is greater than the control level difference threshold, then the temperature constraint condition is used. Based on the temperature constraint condition, all preferred actuator state vectors are screened a second time. If the temperature constraint condition is met, or the cooler power control level in the preferred actuator state vector is level 0, then the preferred actuator state vector is marked. If the difference between the median value of the current ambient temperature range and the target value of the greenhouse temperature to be controlled falls within the range... Within the interval, the preferred actuator state vectors for heaters and coolers with control levels both less than 1 are marked; The marked preferred actuator state vector is subjected to relative humidity conflict detection, specifically including: if the difference between the median value of the current relative humidity range and the target value of the relative humidity of the greenhouse to be controlled is greater than 100%, then the conflict is considered to be caused by a conflict. If the absolute difference between the control levels of the humidifier and the dehumidifier is greater than the control level difference threshold, then the humidity constraint condition is used. Based on the humidity constraint condition, the marked preferred actuator state vector is screened a second time. If the humidity constraint condition is met, or the humidifier power control level in the preferred actuator state vector is level 0, then the marked preferred actuator state vector is marked a second time. If the difference between the median value of the current relative humidity range and the target value of the relative humidity of the greenhouse to be controlled is less than [the median value of the relative humidity range], then [the ... If the absolute difference between the control levels of the humidifier and the dehumidifier is greater than the control level difference threshold, then the humidity constraint condition is used. Based on the humidity constraint condition, the marked preferred actuator state vector is screened a second time. If the humidity constraint condition is met, or the dehumidifier power control level in the preferred actuator state vector is level 0, then the marked preferred actuator state vector is marked a second time. If the difference between the median value of the current relative humidity range and the target value of the relative humidity of the greenhouse to be controlled falls within the range... Within the range, the state vectors of preferred actuators with control levels of both humidifier and dehumidifier less than 1 are marked a second time. Based on the preferred actuator state vector that has been secondary-labeled, a set of optional actuator state vectors is constructed; The difference adjustment thresholds include a temperature difference adjustment threshold and a relative humidity difference adjustment threshold, wherein... Adjusting the threshold for temperature differences, The threshold for adjusting relative humidity difference.

2. The multi-objective intelligent control method for greenhouse climate that integrates rule constraints and reinforcement learning according to claim 1, characterized in that: The actuators in the controlled greenhouse specifically include heaters, coolers, humidifiers, and dehumidifiers; The temperature dynamic model of the greenhouse to be controlled is specifically represented as follows: In the formula, Indicates in The internal temperature of the greenhouse needs to be controlled at all times. This indicates the temperature inside the greenhouse to be controlled at time k. The energy conversion coefficient of the heater. This is the heater control input at time k. The energy conversion coefficient of the refrigerator. This is the cooler control input at time k. The greenhouse heat exchange coefficient, The ambient temperature, Let k be the process noise of the temperature model at time k, specifically the noise value following a Gaussian distribution, where k is the time variable of a day in 24-hour format. It is a quantity that changes over time; The humidity dynamic model of the greenhouse to be controlled is specifically represented as follows: In the formula, and express The relative humidity inside the greenhouse to be controlled at time k and time k. The humidifier's moisture conversion coefficient. The humidifier control input at time k is... This is the moisture conversion coefficient of the dehumidifier. The dehumidifier control input at time k is... The greenhouse moisture exchange coefficient, The humidity of the external environment, The noise of the humidity model at time k is specifically the noise value that follows a Gaussian distribution.

3. The multi-objective intelligent control method for greenhouse climate that integrates rule constraints and reinforcement learning according to claim 2, characterized in that: Based on the dynamics of solar radiation and crop transpiration, an external environmental disturbance term is set, which includes dynamic disturbances of solar radiation and disturbances of crop transpiration. The specific logic is as follows: the dynamic disturbance of solar radiation is comprehensively characterized by solar radiation intensity and cloud cover, specifically expressed as follows: In the formula, Let K be the dynamic perturbation of solar radiation at time k. The solar radiation absorption coefficient. Let be the solar radiation intensity at time k; where the solar radiation intensity at time k is... The specific calculation method is as follows: In the formula, This is the theoretical value of the maximum solar radiation intensity. and These are sunrise and sunset times, respectively. This is the cloud cover coefficient. Let k be the cloud coverage rate at time k. Crop transpiration disturbance is determined based on an empirical model, specifically using the following formula: In the formula, For crop transpiration disturbance, This is the crop transpiration coefficient. Let be the crop transpiration rate at time k, where is the crop transpiration rate at time k. The specific formula used for the calculation is as follows: In the formula, This is the theoretical value of the maximum transpiration rate. Let be the response function of the temperature, humidity, and light intensity of the greenhouse to be controlled at time k to the transpiration rate. Factors related to crop growth stages; The external environmental disturbance term is coupled with the greenhouse temperature dynamic model and the greenhouse humidity dynamic model to obtain the accurate temperature dynamic model and the accurate humidity dynamic model. This establishes the mapping relationship between temperature, humidity, and actuator input data. The specific formula used is as follows: In the formula, for Precise temperature inside the greenhouse at any given moment. for Precise relative humidity inside the greenhouse at any given time; The constraints of the greenhouse operating conditions are set. Within these constraints, simulated temperature and humidity data generated over multiple consecutive time steps are output using precise temperature and humidity dynamic models. Specifically, the constraints of the greenhouse operating conditions include: In the formula, and These are the minimum and maximum greenhouse temperature thresholds, respectively. and These are the minimum and maximum relative humidity thresholds, respectively.

4. The multi-objective intelligent control method for greenhouse climate that integrates rule constraints and reinforcement learning according to claim 3, characterized in that: The training data includes temperature and relative humidity data of the greenhouse to be controlled at several times, as well as the corresponding actuator input data; The specific logic behind performing interval discretization on the training data is as follows: based on the set greenhouse operating condition constraints, determine the temperature constraint interval and relative humidity constraint interval of the greenhouse to be controlled. For the temperature constraint range of the greenhouse to be controlled, the temperature constraint range is evenly divided with a preset temperature adjustment interval to form several temperature ranges of the greenhouse to be controlled. According to the number of temperature ranges, the relative humidity adjustment interval is set to form a corresponding number of relative humidity ranges. The temperature and relative humidity simulation data in the training data are mapped to the corresponding discrete ranges to obtain the intervalized training data of temperature and relative humidity data in the training data. The specific steps for performing interval discretization on actuator input data include: discretizing each actuator input data into 7 levels, which correspond to the control levels from off to maximum power, where level 0 power control level represents the actuator being off, level 6 power control level represents the actuator starting at maximum power, forming a complete action space with the Cartesian product of the four actuators, and determining several actuator state vectors in the action space; Using the current temperature range, relative humidity range, and actuator state vector as input, and the temperature range and relative humidity range corresponding to the time variable as output, a range mapping model is trained to construct a range mapping relationship.

5. The multi-objective intelligent control method for greenhouse climate that integrates rule constraints and reinforcement learning according to claim 4, characterized in that: The intelligent decision-making model is built based on the standard Q-learning algorithm; The environmental target values ​​specifically include the temperature target value and the relative humidity target value of the greenhouse to be controlled. The logic based on the interval mapping relationship analysis of the current temperature interval and the current relative humidity interval is as follows: input the current temperature interval and the current relative humidity interval, as well as the state vector of each actuator, into the interval mapping model that has been trained to obtain the current temperature interval and the current relative humidity interval. Under any actuator state vector, the temperature interval and the relative humidity interval predicted after the corresponding time variable. The specific method for selecting the preferred actuator state vector is as follows: the deviation between the predicted temperature range and relative humidity range and the environmental target value is used as the evaluation benchmark, and a deviation threshold is set. The actuator state vectors corresponding to the deviation between the predicted temperature range and relative humidity range and the environmental target value is not greater than the deviation threshold are marked as preferred actuator state vectors. The specific logic for determining the deviation between the predicted temperature range and relative humidity range and the environmental target value is as follows: if the environmental target value falls within the predicted temperature range and relative humidity range, then the deviation between the predicted temperature range and relative humidity range and the environmental target value is marked as 0. For the temperature of the greenhouse to be controlled in the environmental target values, if the temperature target value of the greenhouse to be controlled is less than the predicted minimum temperature range, the relative error between the predicted minimum temperature range and the temperature target value of the greenhouse to be controlled is taken as the degree of temperature deviation. If the target temperature setting value of the greenhouse to be controlled is greater than the maximum value of the predicted temperature range, then the relative error between the target temperature setting value and the maximum value of the predicted temperature range will be used as the degree of temperature deviation. For the relative humidity of the greenhouse to be controlled, a target value is set. If the target value is less than the predicted minimum relative humidity range, the relative error between the predicted minimum relative humidity range and the target value is taken as the degree of relative humidity deviation. If the target value is greater than the predicted maximum relative humidity range, the relative error between the target value and the predicted maximum relative humidity range is taken as the degree of relative humidity deviation. The deviation of the predicted temperature range and relative humidity range from the environmental target value is defined as the maximum value of the relative humidity deviation and the temperature deviation.

6. The multi-objective intelligent control method for greenhouse climate that integrates rule constraints and reinforcement learning according to claim 5, characterized in that: The specific formula used to calculate the safety correction amount of the optional actuator state vector is as follows: In the formula, This is the safety correction amount for the v-th optional executor state vector. Let v be the state vector of the optional executor. This is the current executor state vector. This represents the Euclidean distance between the v-th optional executor state vector and the current executor state vector, where v is the index of the optional executor state vector; The specific formula upon which the multi-objective reward function is designed is as follows: In the formula, Let v be the multi-objective reward function for the v-th optional executor state vector. The energy-saving reward term for the v-th optional actuator state vector. The tracking reward term for the v-th optional executor state vector. , and Here are the weighting coefficients, where and , and All are greater than 0.

7. The multi-objective intelligent control method for greenhouse climate that integrates rule constraints and reinforcement learning according to claim 6, characterized in that: Energy-saving reward term for the v-th optional actuator state vector The specific logic underlying the calculation is as follows: In the formula, Let j be the energy consumption coefficient of the j-th actuator in the v-th optional actuator state vector. Let j be the power control level of the j-th actuator in the state vector of the v-th selectable actuator; The tracking reward term for the v-th optional executor state vector The formula used for the calculation is: In the formula, The median of the temperature range of the greenhouse to be controlled after executing the v-th optional actuator state vector. The median of the relative humidity interval of the greenhouse to be controlled after executing the v-th optional actuator state vector. and Target values ​​are set for temperature and relative humidity, respectively. The specific temperature range and relative humidity range of the greenhouse observation value after the vth optional actuator state vector are determined by the interval mapping model. Based on a multi-objective reward function, the Q-value of the optional actuator state vector is determined by the Bellman equation. The optional actuator state vector corresponding to the maximum Q-value is selected as the optimal control command, and the intelligent decision model is updated accordingly.

Images