Multi-objective daylight greenhouse ventilation decision-making method, device, electronic equipment and storage medium

Abstract

The present invention provides a multi-objective ventilation decision-making method, device, electronic device and storage medium for a solar greenhouse, including: obtaining real-time environmental information inside and outside the greenhouse to determine the greenhouse ventilation feature vector; generating a preliminary ventilation decision-making strategy according to the greenhouse ventilation feature vector; performing multi-objective optimization on the preliminary ventilation decision-making strategy; and dynamically adjusting the optimal ventilation decision-making strategy to automatically adapt to the meteorological changes inside and outside the greenhouse and the growth stage of the crops. The multi-objective ventilation decision-making method provided by the present invention comprehensively uses technologies such as multi-sensor fusion technology, multi-modal variational autoencoder algorithm, soft behavior cloning algorithm, meta-learning technology, multi-objective particle swarm optimization algorithm, multi-objective co-evolution algorithm, multi-agent system and transfer learning technology of reinforcement learning, etc., which can effectively improve the efficiency of greenhouse ventilation decision-making, reduce the energy consumption of greenhouse ventilation, and improve the intelligent control of the temperature inside the greenhouse and the growth quality of the crops.

Description

Technical Field

[0001] The present invention relates to the technical field of agricultural automation, and particularly to a multi-objective ventilation decision-making method, device, electronic device and storage medium for a solar greenhouse. Background Art

[0002] A solar greenhouse is a facility that uses solar energy and natural ventilation to adjust the indoor environment of the greenhouse. It has the advantages of energy conservation, low carbon, environmental protection, etc., and is an important tool in agricultural production. However, solar greenhouses also face challenges such as the uncertainty of the indoor and outdoor environment, the volatility of the indoor temperature, and the diversity of crops. The above factors will affect the ventilation effect of the greenhouse and the growth quality of the crops.

[0003] Traditional greenhouse ventilation decision-making methods are usually based on experience or fixed thresholds, lacking dynamic adaptation to the indoor and outdoor environment of the greenhouse, and not considering the multi-objective requirements such as the temperature, humidity, and energy consumption of the crops in the greenhouse. The above methods often cannot achieve the optimal control of the indoor temperature of the greenhouse, nor can they meet the growth requirements of the crops, resulting in low energy utilization efficiency of the greenhouse and poor growth quality of the crops.

[0004] Therefore, how to reasonably make greenhouse ventilation decisions based on the real-time environmental data inside and outside the greenhouse is the key issue to improve the production efficiency and effect of the greenhouse. Summary of the Invention

[0005] The present invention provides a multi-objective ventilation decision-making method, device, electronic device and storage medium for a solar greenhouse, which at least solves one of the technical problems such as low efficiency of greenhouse ventilation decision-making, high energy consumption, and poor growth quality of crops in the prior art or related technologies.

[0006] The present invention provides a multi-objective ventilation decision-making method for a solar greenhouse, including the following steps:

[0007] Obtain the real-time environmental information inside and outside the greenhouse;

[0008] Determine the greenhouse ventilation feature vector according to the real-time environmental information;

[0009] Generate a preliminary ventilation decision-making strategy according to the greenhouse ventilation feature vector;

[0010] Perform multi-objective optimization on the preliminary ventilation decision-making strategy so that the optimized optimal ventilation decision-making strategy takes into account the indoor temperature control, crop growth requirements, operating costs and effects of the greenhouse;

[0011] Dynamically adjust the optimal ventilation decision-making strategy to automatically adapt to the meteorological changes inside and outside the greenhouse and the growth stages of the crops.

[0012] A multi-objective solar greenhouse ventilation decision-making method provided by the present invention, wherein the real-time environmental information includes at least one of the following environmental data: indoor and outdoor temperature, indoor and outdoor humidity, indoor and outdoor light radiation value, wind speed, wind direction, indoor and outdoor images, greenhouse volume, greenhouse height-span ratio, greenhouse comprehensive thermal resistance, indoor and outdoor temperature difference, light transmittance, saturated water vapor pressure difference;

[0013] The obtaining of the real-time environmental information inside and outside the greenhouse includes:

[0014] Fusing the temperature measurement values inside and outside the greenhouse by the weighted average method to obtain the temperature inside and outside the greenhouse;

[0015] Fusing the temperature and humidity measurement values inside and outside the greenhouse by the weighted average method to obtain the temperature and humidity inside and outside the greenhouse;

[0016] Fusing the light radiation measurement values inside and outside the greenhouse by the weighted average method to obtain the light radiation inside and outside the greenhouse;

[0017] Fusing the wind speed measurement values inside and outside the greenhouse by the weighted average method to obtain the wind speed inside and outside the greenhouse;

[0018] Fusing the wind direction measurement values inside and outside the greenhouse by the vector average method to obtain the wind direction inside and outside the greenhouse;

[0019] Fusing the captured images inside and outside the greenhouse taken in real time at the pixel level or feature level to obtain the indoor and outdoor images;

[0020] Calculating the greenhouse volume according to the dimensions such as the length, width, and height of the greenhouse;

[0021] Calculating the greenhouse height-span ratio according to the height and span of the greenhouse;

[0022] Calculating the greenhouse comprehensive thermal resistance according to the material thermal resistance and surface area of each surface of the greenhouse;

[0023] Calculating the indoor and outdoor temperature difference according to the indoor and outdoor temperatures of the greenhouse;

[0024] Calculating the light transmittance according to the indoor and outdoor light radiation values of the greenhouse;

[0025] Calculating the saturated water vapor pressure difference according to the temperature inside the greenhouse and the humidity inside the greenhouse.

[0026] According to a multi-objective solar greenhouse ventilation decision-making method provided by the present invention, the determining of the greenhouse ventilation feature vector according to the real-time environmental information includes:

[0027] Inputting the real-time environmental information into a multi-modal variational autoencoder to obtain an initial feature vector related to the influencing factors of greenhouse ventilation and the temperature change law;

[0028] Using a multi-modal attention mechanism, adaptively weighted sum fusion is performed on the initial feature vector to obtain a greenhouse ventilation feature vector.

[0029] According to a multi-objective solar greenhouse ventilation decision-making method provided by the present invention, the step of using a multi-modal attention mechanism to perform adaptively weighted sum fusion on the initial feature vector to obtain a greenhouse ventilation feature vector includes:

[0030] Determine the attention weights and gating mechanisms of the real-time environmental information of different modalities that make up the initial feature vector;

[0031] According to the attention weights and gating mechanisms, perform adaptively weighted sum fusion on the data of different modalities to obtain the greenhouse ventilation feature vector.

[0032] According to a multi-objective solar greenhouse ventilation decision-making method provided by the present invention, the step of generating a preliminary ventilation decision-making strategy according to the greenhouse ventilation feature vector includes:

[0033] Based on the soft behavior cloning algorithm, maximize the reward and minimize the difference between the output action and the expert strategy to construct an objective function;

[0034] According to the greenhouse ventilation feature vector and the reward function determined by the objective function, determine the preliminary ventilation decision-making strategy;

[0035] According to the preliminary ventilation decision-making strategy, obtain the output action of greenhouse ventilation;

[0036] The objective function includes a policy network and a value network respectively constructed by using two neural networks; the policy network outputs Gaussian distribution parameters, and the Gaussian distribution parameters are used to characterize the probability distribution of the output action under the greenhouse ventilation feature vector; the value network outputs a scalar, and the scalar is used to characterize the expected reward value under the greenhouse ventilation feature vector.

[0037] According to a multi-objective solar greenhouse ventilation decision-making method provided by the present invention, after constructing the objective function, it further includes:

[0038] Optimize the objective function based on the entropy regularization method, including adding an entropy term to the objective function, and the expression of the entropy term is:

[0039] ;

[0040] The expression of the optimized objective function is:

[0041] ;

[0042] Wherein, Represents the expectation of a variable; The greenhouse ventilation eigenvector for greenhouse ventilation; The output action for greenhouse ventilation; The reward function for greenhouse ventilation; The policy network; and The model parameters; The expert demonstration dataset; and The hyperparameters; The entropy function; The KL divergence; The expert policy; The entropy term for entropy regularization; The objective function.

[0043] According to a multi-objective solar greenhouse ventilation decision-making method provided by the present invention, after determining the preliminary ventilation decision-making strategy, it further includes:

[0044] Using multi-task learning and meta-gradient update, adaptively adjust the greenhouse ventilation decision-making strategy according to the current greenhouse environment and crop variety.

[0045] According to a multi-objective solar greenhouse ventilation decision-making method provided by the present invention, multi-objective optimization of the preliminary ventilation decision-making strategy includes:

[0046] Using a multi-objective particle swarm optimization algorithm and a multi-objective co-evolution algorithm to perform multi-objective optimization on the preliminary decision-making actions, comprehensively considering objectives such as temperature control in the greenhouse, crop growth requirements, operating costs, and effects, to obtain the optimal ventilation decision-making strategy.

[0047] According to a multi-objective solar greenhouse ventilation decision-making method provided by the present invention, dynamically adjust the optimal ventilation decision-making strategy to automatically adapt to meteorological changes inside and outside the greenhouse and the growth stage of the crops, including:

[0048] Through a multi-agent system and transfer learning technology, utilize the similarity and difference between the source task and the target task to perform real-time optimization on the optimal ventilation decision-making strategy;

[0049] The target task is determined according to the meteorological changes inside and outside the greenhouse and the growth stage of the crops.

[0050] The present invention also provides a multi-objective solar greenhouse ventilation decision-making device, including the following modules:

[0051] An information acquisition unit for obtaining real-time environmental information inside and outside the greenhouse;

[0052] An information processing unit for determining a greenhouse ventilation feature vector according to the real-time environmental information;

[0053] A strategy generation unit for generating a preliminary ventilation decision strategy according to the greenhouse ventilation feature vector;

[0054] A strategy optimization unit for performing multi-objective optimization on the preliminary ventilation decision strategy so that the optimized optimal ventilation decision strategy takes into account temperature control in the greenhouse, crop growth requirements, operating costs and effects;

[0055] A strategy matching unit for dynamically adjusting the optimal ventilation decision strategy to automatically adapt to meteorological changes inside and outside the greenhouse and the growth stage of the crops.

[0056] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the program, it implements the multi-objective sunlight greenhouse ventilation decision method as described in any one of the above.

[0057] The present invention also provides a non-transitory computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, it implements the multi-objective sunlight greenhouse ventilation decision method as described in any one of the above.

[0058] The present invention also provides a computer program product, including a computer program. When the computer program is executed by a processor, it implements the multi-objective sunlight greenhouse ventilation decision method as described in any one of the above.

[0059] The multi-objective sunlight greenhouse ventilation decision method, device, electronic device and storage medium provided by the present invention comprehensively utilize technologies such as multi-sensor fusion technology, multi-modal variational autoencoder algorithm, soft behavior cloning algorithm, meta-learning technology, multi-objective particle swarm optimization algorithm, multi-objective co-evolution algorithm, multi-agent system and transfer learning technology of reinforcement learning, etc., which can effectively improve the efficiency of greenhouse ventilation decision-making, reduce the energy consumption of greenhouse ventilation, and improve the intelligent control of the temperature in the greenhouse and the growth quality of crops. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the drawings in the following description are some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0061] Figure 1 It is a flowchart of the multi-objective sunlight greenhouse ventilation decision method provided by the present invention.

[0062] Figure 2 It is a schematic structural diagram of the multi-objective solar greenhouse ventilation decision-making device provided by the present invention.

[0063] Figure 3 It is a schematic structural diagram of the electronic device provided by the present invention. Detailed implementation manners

[0064] To make the objectives, technical solutions and advantages of the present invention clearer, the technical solutions in the present invention will be clearly and completely described below with reference to the accompanying drawings in the present invention. Obviously, the described embodiments are some but not all of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments in the present invention without creative efforts shall fall within the protection scope of the present invention.

[0065] It should be noted that in the description of the present invention, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements not only includes those elements but also includes other elements not explicitly listed, or further includes elements inherent to such process, method, article or device. Without further limitations, an element defined by the statement "including one..." does not exclude the existence of additional identical elements in the process, method, article or device including the said element. For those of ordinary skill in the art, the specific meanings of the above terms in the present invention can be understood according to specific situations.

[0066] Greenhouse ventilation decision-making is an important link in greenhouse production. The traditional technical methods mainly include the following several:

[0067] (1) Experience-based method: According to the experience and intuition of greenhouse producers, based on indicators such as temperature and humidity inside and outside the greenhouse, manually or regularly adjust the ventilation opening of the greenhouse to control the temperature inside the greenhouse. The advantage is that it is simple and easy to implement, and does not require complex equipment and algorithms. However, the disadvantage is that it lacks scientific basis, cannot adapt to the dynamic changes of the environment inside and outside the greenhouse, and cannot consider the multi-objective needs of the crops inside the greenhouse, which is likely to cause the temperature inside the greenhouse to be too high or too low, affecting the growth quality of the crops.

[0068] (2) Threshold-based method: According to indicators such as temperature and humidity inside and outside the greenhouse, set fixed thresholds. When the environmental data inside and outside the greenhouse exceeds or is lower than the above thresholds, automatically adjust the ventilation opening of the greenhouse to control the temperature inside the greenhouse. The advantage is that it can realize the automation of greenhouse ventilation and reduce manual intervention. The disadvantage is that the setting of the threshold is often subjective, cannot reflect the real-time changes of the environment inside and outside the greenhouse, and cannot consider the multi-objective needs of the crops inside the greenhouse, which is likely to cause fluctuations in the temperature inside the greenhouse, affecting the growth quality of the crops.

[0069] (3) Model-based methods: Based on the environmental data inside and outside the greenhouse, a mathematical model of the greenhouse is established, such as an energy balance model, a dynamic model, etc. Then, according to the greenhouse model, optimization algorithms such as linear programming, non-linear programming, and fuzzy control are used to solve the optimal ventilation opening of the greenhouse to achieve temperature control inside the greenhouse. The advantage is that it can optimize the greenhouse ventilation and improve the energy utilization efficiency of the greenhouse. However, the disadvantage is that establishing a greenhouse model is often complex, requires a large number of parameters and data, and the greenhouse model is often inaccurate, unable to fully reflect the complexity of the environment inside and outside the greenhouse, nor can it consider the multi-objective requirements of the crops inside the greenhouse, which is likely to cause decision errors and instability.

[0070] Existing greenhouse ventilation decision-making technologies are mainly applied in the field of facility agriculture, especially in the field of solar greenhouses, aiming to achieve reasonable temperature control inside the greenhouse and ensure the growth and quality of crops. However, the existing technologies still have the following deficiencies:

[0071] (1) Insufficient data acquisition and processing capabilities: Existing technologies often only use single or limited sensor data, unable to obtain various data types such as temperature, humidity, and light radiation inside and outside the greenhouse, nor can they effectively fuse and analyze the data, resulting in insufficient acquisition and processing capabilities of greenhouse environmental information and unable to reflect the real-time and comprehensive state of the greenhouse.

[0072] In response to this, the present invention uses multi-sensor fusion technology to collect and analyze various data types inside and outside the greenhouse to obtain real-time and comprehensive information on the greenhouse environment, improve the quality and reliability of the data, as well as the expression and discrimination capabilities of the data.

[0073] (2) Insufficient multi-objective optimization and coordination capabilities: Existing technologies often only consider single or main objectives, such as minimizing the temperature difference inside and outside the greenhouse, maximizing the yield or quality of crops, etc., ignoring the mutual influence and restriction between other objectives, resulting in insufficient multi-objective optimization and coordination capabilities of greenhouse ventilation and unable to balance temperature control inside the greenhouse and crop growth requirements, as well as energy consumption and utilization.

[0074] In response to this, the present invention uses multi-objective optimization algorithms to comprehensively consider multiple objectives such as meteorological parameters inside and outside the greenhouse, crop growth requirements, and energy benefits of active heat storage and release, to achieve multi-objective optimization and coordination of greenhouse ventilation, and balance temperature control inside the greenhouse and crop growth requirements, as well as energy consumption and utilization.

[0075] (3) Insufficient intelligence and adaptability: Existing technologies often rely on artificial or empirical set values, parameters, or models, unable to automatically learn and optimize greenhouse ventilation decision-making strategies, nor can they dynamically adjust greenhouse ventilation decision-making strategies, resulting in insufficient intelligence and adaptability of greenhouse ventilation and being unable to adapt to meteorological changes inside and outside the greenhouse and the growth stages of crops.

[0076] In response to this, the present invention utilizes reinforcement learning and meta-learning technologies to automatically learn and optimize greenhouse ventilation decision-making strategies, dynamically adjust greenhouse ventilation decision-making strategies, achieve the intelligence and adaptability of greenhouse ventilation, and adapt to meteorological changes inside and outside the greenhouse and the growth stages of crops.

[0077] Generally speaking, the multi-objective solar greenhouse ventilation decision-making method provided by the present invention has the following main improvements:

[0078] (1) Data-based method: According to the environmental data inside and outside the greenhouse, using data mining and machine learning technologies - such as clustering, classification, regression, neural networks, etc., analyze and predict the greenhouse ventilation data, and then according to the data characteristics of greenhouse ventilation, use decision trees, rules, expert systems, etc. to formulate decision rules for greenhouse ventilation to achieve the control of the temperature inside the greenhouse, and can utilize a large amount of data to improve the accuracy and sensitivity of greenhouse ventilation decision-making.

[0079] (2) Deep learning-based method: According to the environmental data inside and outside the greenhouse, using deep learning technologies such as convolutional neural networks, recurrent neural networks, autoencoders, etc., perform deep feature extraction and representation on the greenhouse ventilation data, and then according to the deep features of greenhouse ventilation, use deep reinforcement learning technologies such as deep Q networks, policy gradients, actor-critic, etc. to learn and optimize the greenhouse ventilation decision-making strategy to achieve the control of the temperature inside the greenhouse, and can utilize the powerful representation ability and generalization ability of deep learning to improve the intelligence level and efficiency of greenhouse ventilation decision-making.

[0080] (3) Multi-objective optimization-based method: According to the environmental data inside and outside the greenhouse and the multi-objective requirements of the crops inside the greenhouse, establish a multi-objective optimization problem for greenhouse ventilation temperature, humidity, light, etc., and then use multi-objective optimization technologies such as genetic algorithms, particle swarm optimization, co-evolution, etc. to solve the multi-objective optimization solution for greenhouse ventilation to achieve the control of the temperature inside the greenhouse, and can comprehensively consider the multi-objective requirements of greenhouse crops to improve the coordination and balance of decision-making.

[0081] The following combines Figures 1 - 3 to describe the multi-objective solar greenhouse ventilation decision-making method, device, electronic device, and storage medium provided by the present invention, aiming to achieve more accurate and detailed classification of the target area through the combination of panchromatic images and multispectral images, and can improve the classification efficiency to a certain extent.

[0082] Figure 1 This is one of the flow schematic diagrams of the multi-objective solar greenhouse ventilation decision-making method provided by the present invention. The execution subject can be an industrial control computer, a proximal computer, a cloud server, or other movable electronic terminals with data processing functions, such as an Ipad, a mobile phone, etc. The present invention does not make specific limitations on this. As Figure 1 shown, the method includes but is not limited to the following steps:

[0083] Step 101, obtain the real-time environmental information inside and outside the greenhouse.

[0084] Obtain various data types such as temperature, humidity, light radiation, wind speed, wind direction, images, etc. inside and outside the greenhouse as the input states for greenhouse ventilation decision-making. The role is to obtain the real-time information of the environment inside and outside the greenhouse and provide data support for greenhouse ventilation decision-making.

[0085] To improve the quality and reliability of the data, this step adopts the multi-sensor fusion technology (Multi-Sensor Fusion Technology, MSFT), uses sensors of different types and settings at different positions to comprehensively measure and calibrate the environmental parameters inside and outside the greenhouse, so as to eliminate the errors and noises of a single sensor, improve the accuracy and stability of the data, effectively solve the limitations and deficiencies of a single sensor, and improve the integrity and credibility of greenhouse environmental data.

[0086] Furthermore, it is also possible to calculate and derive the following data as co-modeling variables according to the greenhouse structure parameter indicators and the internal and external meteorological environment data: greenhouse volume, greenhouse height-span ratio, greenhouse comprehensive thermal resistance (according to the proportion of their respective surface areas), indoor-outdoor temperature difference, light transmittance, vapor pressure deficit (VPD), etc.

[0087] Step 102, determine the greenhouse ventilation feature vector according to the real-time environmental information.

[0088] According to the multi-source real-time environmental information (such as temperature, humidity, light radiation, wind speed, wind direction, images, etc.) obtained in the previous step, and combined with the calculated derived data (such as greenhouse volume, height-span ratio, comprehensive thermal resistance, indoor-outdoor temperature difference, light transmittance, vapor pressure deficit, etc.), construct a feature vector reflecting the greenhouse ventilation characteristics, which is called the greenhouse ventilation feature vector. This greenhouse ventilation feature vector is the basis for formulating subsequent greenhouse ventilation decisions, and it comprehensively combines various key factors affecting greenhouse ventilation decisions.

[0089] Step 103, generate a preliminary ventilation decision strategy according to the greenhouse ventilation feature vector.

[0090] A preset decision-making model or algorithm (such as a rule-based decision tree, fuzzy logic, neural network, etc.) can be used to analyze the priority and necessity of greenhouse ventilation under the current environmental conditions, so as to generate a preliminary ventilation decision-making strategy. This strategy may include the opening and closing degree of ventilation windows, ventilation time, ventilation method (natural ventilation or mechanical ventilation), etc.

[0091] Step 104, perform multi-objective optimization on the preliminary ventilation decision-making strategy so that the optimized optimal ventilation decision-making strategy takes into account temperature control in the greenhouse, crop growth requirements, operating costs, and effects.

[0092] Based on the generated preliminary ventilation decision-making strategy, it can be further optimized with multiple objectives. The goal of optimization is to balance multiple aspects such as temperature control in the greenhouse, crop growth requirements, operating costs, and effects, ensuring that the ventilation decision can not only meet the best environmental conditions for crop growth but also effectively control energy consumption and operating costs.

[0093] During the optimization process, multi-objective optimization algorithms (such as NSGA-II, MOEA / D, etc.) may be used to comprehensively consider the conflicts and trade-offs between various objectives.

[0094] Step 105, dynamically adjust the optimal ventilation decision-making strategy to automatically adapt to meteorological changes inside and outside the greenhouse and the growth stage of the crops.

[0095] After completing the multi-objective optimization, the optimized optimal ventilation decision-making strategy can also be dynamically adjusted according to the real-time monitored meteorological changes inside and outside the greenhouse and the growth stage of the crops. This dynamic adjustment ability enables the system to flexibly respond to various unpredictable environmental changes, ensuring that the greenhouse environment is always in the most suitable state for crop growth. At the same time, through continuous learning and adaptation, it can gradually optimize its decision-making logic and improve the accuracy and efficiency of ventilation decision-making.

[0096] The multi-objective daylight greenhouse ventilation decision-making method provided by the present invention can comprehensively apply technologies such as multi-sensor fusion technology, multi-modal variational autoencoder algorithm, soft behavior cloning algorithm, meta-learning technology, multi-objective particle swarm optimization algorithm, multi-objective co-evolution algorithm, multi-agent system, and transfer learning technology of reinforcement learning, etc., which can effectively improve the efficiency of greenhouse ventilation decision-making, reduce the energy consumption of greenhouse ventilation, and improve the intelligent control of the temperature inside the greenhouse and the growth quality of crops.

[0097] Based on the content of the above embodiments, as an alternative embodiment, the real-time environmental information includes at least one of the following environmental data: indoor and outdoor temperature, indoor and outdoor humidity, indoor and outdoor light radiation value, wind speed, wind direction, indoor and outdoor images, greenhouse volume, greenhouse height-span ratio, greenhouse comprehensive thermal resistance, indoor and outdoor temperature difference, light transmittance, saturated water vapor pressure difference;

[0098] The acquisition of real-time environmental information inside and outside the greenhouse includes:

[0099] Fusing the temperature measurement values inside and outside the greenhouse by the weighted average method to obtain the temperature inside and outside the greenhouse;

[0100] Fusing the temperature and humidity measurement values inside and outside the greenhouse by the weighted average method to obtain the temperature and humidity inside and outside the greenhouse;

[0101] Fusing the light radiation measurement values inside and outside the greenhouse by the weighted average method to obtain the light radiation inside and outside the greenhouse;

[0102] Fusing the wind speed measurement values inside and outside the greenhouse by the weighted average method to obtain the wind speed inside and outside the greenhouse;

[0103] Fusing the wind direction measurement values inside and outside the greenhouse by the vector average method to obtain the wind direction inside and outside the greenhouse;

[0104] Fusing the captured images inside and outside the greenhouse taken in real time at the pixel level or feature level to obtain the indoor and outdoor images;

[0105] Calculating the volume of the greenhouse according to the dimensions such as the length, width, and height of the greenhouse;

[0106] Calculating the height-span ratio of the greenhouse according to the height and span of the greenhouse;

[0107] Calculating the comprehensive thermal resistance of the greenhouse according to the material thermal resistance and surface area of each surface of the greenhouse;

[0108] Calculating the temperature difference between inside and outside the greenhouse according to the indoor and outdoor temperatures of the greenhouse;

[0109] Calculating the light transmittance according to the indoor and outdoor light radiation values of the greenhouse;

[0110] Calculating the saturation vapor pressure difference according to the temperature and humidity inside the greenhouse;

[0111] In the present invention, various types of sensors are respectively installed at multiple positions inside and outside the greenhouse, mainly including temperature sensors, humidity sensors, light radiation sensors, wind speed sensors, wind direction sensors, image sensors, etc., to respectively measure the environmental parameters inside and outside the greenhouse, such as temperature, humidity, light radiation, wind speed, wind direction, image, etc., and send the measurement data to the data processing center as the execution subject through the wireless transmission module, as the input state for greenhouse ventilation decision-making.

[0112] In order to improve the quality and reliability of data, the present invention adopts multi-sensor fusion technology, using sensors of different types and arranged at different positions to comprehensively measure and calibrate the environmental parameters inside and outside the greenhouse, eliminate sensor errors and noise, improve data accuracy and stability, effectively solve the limitations and shortcomings of a single sensor, and improve the integrity and credibility of greenhouse environmental data.

[0113] Specifically, this step adopts the following multi-sensor fusion technology:

[0114] Temperature sensor fusion technology: Several temperature sensors (of different types) are installed inside and outside the greenhouse to measure the temperature values ​​inside and outside the greenhouse (called temperature measurement values). The temperature measurement values ​​are fused by weighted averaging to obtain the temperature fusion values ​​inside and outside the greenhouse.

[0115] The weighted average calculation formula is as follows:

[0116] ;

[0117] in, is the temperature fusion value, i.e. the final temperature, For the The temperature measurement value of each temperature sensor, For the The weight of each temperature sensor, is the total number of temperature sensors. Weight It is determined and dynamically adjusted according to factors such as the accuracy, position, and stability of the temperature sensor to ensure the accuracy of the temperature fusion value.

[0118] Humidity sensor fusion technology: install several humidity sensors inside and outside the greenhouse, measure the humidity values ​​inside and outside the greenhouse respectively, fuse the humidity values ​​through weighted average method, and get the humidity fusion value inside and outside the greenhouse. The calculation formula of weighted average is the same as that of temperature sensor fusion technology. Replaced with humidity measurement , temperature fusion value Replaced with humidity fusion value That's it.

[0119] Light radiation sensor fusion technology: install several light radiation sensors inside and outside the greenhouse to measure the light radiation values ​​inside and outside the greenhouse respectively, and fuse the light radiation measurement values ​​through the weighted average method to obtain the light radiation fusion value inside and outside the greenhouse. The formula of the weighted average method is the same as that of the temperature fusion technology. Replaced with light radiation measurements , temperature fusion value Replaced with light radiation fusion value That's all.

[0120] Wind speed sensor fusion technology: Install several wind speed sensors inside and outside the greenhouse (near the ventilation openings) to measure the wind speed values inside and outside the greenhouse respectively. The measured values are fused by the weighted average method to obtain the wind speed fusion value inside and outside the greenhouse. The formula of the weighted average method is the same as that of the temperature sensor fusion technology, only need to replace the temperature value with the wind speed value and replace the temperature fusion value with the wind speed fusion value

[0121] Wind direction sensor fusion technology: Install several wind direction sensors inside and outside the greenhouse to measure the wind direction values inside and outside the greenhouse respectively. The measured values are fused by the vector average method to obtain the wind direction fusion value inside and outside the greenhouse. The formula of the vector average method is as follows:

[0122] ;

[0123] where, is the wind direction fusion value, is the measured value of the th wind direction sensor, is the total number of wind direction sensors.

[0124] Image sensor fusion technology: Install several image sensors inside and outside the greenhouse to collect the captured images inside and outside the greenhouse respectively. The captured images are fused by image processing algorithms to obtain the fused images inside and outside the greenhouse.

[0125] Among them, the image processing algorithms include steps such as image alignment, image fusion, and image enhancement, which are specifically as follows: Image alignment performs geometric transformation on the images inside and outside the greenhouse to make the perspectives, sizes, directions, etc. of the images consistent, facilitating subsequent image fusion. Image fusion performs pixel-level or feature-level fusion on the images inside and outside the greenhouse to make the brightness, contrast, color, etc. of the images uniform, improving the clarity and quality of the images. Image enhancement performs operations such as filtering, sharpening, and denoising on the images inside and outside the greenhouse to eliminate defects such as blurring, noise, and distortion in the images, improving the details and information content of the images.

[0126] Furthermore, the present invention can calculate and derive the following data as co-modeling variables according to the greenhouse structure parameter indicators and the above-mentioned fused meteorological environment data inside and outside the greenhouse.

[0127] Greenhouse volume: According to the dimensions such as the length, width, and height of the greenhouse, calculate the volume of the greenhouse, with the unit of cubic meters. The calculation formula of the greenhouse volume is as follows:

[0128] ;

[0129] where, is the greenhouse volume, is the greenhouse length, is the greenhouse width, is the greenhouse height.

[0130] Greenhouse height-span ratio: According to the height and span of the greenhouse, calculate the greenhouse height-span ratio, with the unit being dimensionless. The formula for the greenhouse height-span ratio is as follows:

[0131] ;

[0132] where, is the greenhouse height-span ratio, is the greenhouse height, is the greenhouse span, that is, the maximum horizontal width of the greenhouse.

[0133] Greenhouse overall thermal resistance: According to the parameters such as the material, thickness, and area of each surface of the greenhouse, calculate the greenhouse overall thermal resistance, with the unit being square meter·Kelvin / watt. The formula for the greenhouse overall thermal resistance is as follows:

[0134] ;

[0135] where, is the greenhouse overall thermal resistance, is the thermal resistance of the th surface of the greenhouse, is the area of the th surface of the greenhouse, is the total number of surfaces of the greenhouse.

[0136] Among them, the calculation of the thermal resistance is carried out according to Fourier's law, and the formula is as follows:

[0137] ;

[0138] where, is the thickness of the th surface of the greenhouse, is the thermal conductivity of the th surface of the greenhouse.

[0139] Indoor-outdoor temperature difference: According to the temperature integration values inside and outside the greenhouse, calculate the indoor-outdoor temperature difference of the greenhouse, with the unit being degree Celsius. The formula for the indoor-outdoor temperature difference is as follows:

[0140] ;

[0141] where, is the indoor-outdoor temperature difference, is the temperature integration value inside the greenhouse, is the temperature integration value outside the greenhouse.

[0142] Light transmittance: Calculate the light transmittance of the greenhouse according to the integrated value of light radiation inside and outside the greenhouse, with the unit of percentage. The formula for light transmittance is as follows:

[0143] ;

[0144] wherein, is the light transmittance, is the integrated value of light radiation inside the greenhouse, is the integrated value of light radiation outside the greenhouse.

[0145] Vapor Pressure Deficit (VPD): Calculate the vapor pressure deficit of the greenhouse according to the integrated values of temperature and humidity inside and outside the greenhouse, with the unit of kilopascal.

[0146] The calculation formula of VPD is as follows:

[0147] ;

[0148] wherein, is the vapor pressure deficit, is the saturated vapor pressure inside the greenhouse, is the actual vapor pressure inside the greenhouse, is the integrated value of temperature inside the greenhouse, is the integrated value of humidity inside the greenhouse.

[0149] The saturated vapor pressure and the actual vapor pressure are calculated according to the following empirical formulas:

[0150] ;

[0151] .

[0152] Based on the content of the above embodiments, as an alternative embodiment, determining the greenhouse ventilation feature vector according to the real-time environmental information includes:

[0153] Input the real-time environmental information into the multi-modal variational autoencoder to obtain an initial feature vector related to the influencing factors of greenhouse ventilation and the temperature change law;

[0154] Use the multi-modal attention mechanism to perform adaptive weighted sum and fusion on the initial feature vector to obtain the greenhouse ventilation feature vector.

[0155] The present invention can adopt a multi-modal variational autoencoder (MVAE) to jointly model and generate real-time environmental information of multiple data types collected inside and outside the greenhouse by using variational inference and latent variable models, extract and represent the influencing factors related to greenhouse ventilation and the temperature change law, and obtain an initial feature vector characterizing greenhouse ventilation. Its function is to perform deep learning and feature extraction on greenhouse environmental data and provide an input feature representation for greenhouse ventilation decision-making.

[0156] To improve the expression ability and information content of the feature vector, a multi-modal attention mechanism is adopted. By using attention weights and gating mechanisms, the real-time environmental information of different modalities is adaptively weighted and fused, highlighting important information and related information, and suppressing redundant information and noisy information.

[0157] The multi-modal attention mechanism effectively improves the fusion effect of multi-modal data and the representation ability of the feature vector. Its specific implementation steps include:

[0158] Adopt a multi-modal variational autoencoder to jointly model and generate multiple data types inside and outside the greenhouse by using variational inference and latent variable models, extract and represent the influencing factors of greenhouse ventilation and the temperature change law, and obtain the feature vector of greenhouse ventilation.

[0159] The multi-modal variational autoencoder deep generation model can simultaneously process data of multiple modalities and can perform conversion and complementarity between different modalities. In the structure of the multi-modal variational autoencoder, For kinds of real-time environmental information of different modalities; is the latent variable, that is, the output feature vector of greenhouse ventilation; is the encoder; is the decoder; And are the model parameters.

[0160] The goal of the multi-modal variational autoencoder is to maximize the marginal log-likelihood of the data, that is:

[0161] ;

[0162] Since the integral in the above formula is unsolvable, a variational inference method can be adopted to introduce an approximate posterior distribution , and maximize the variational lower bound (ELBO), that is:

[0163] ;

[0164] Among them, represents the expectation; represents the KL divergence; is the prior distribution of the latent variable, usually assumed to be a standard normal distribution; is the variational lower bound.

[0165] For convenient calculation, the approximate posterior distribution is constructed in the following way:

[0166] ;

[0167] where is the normalization constant; denotes K the product of the posterior distributions of the latent variable z given by single-modal encoders; the single-modal encoders are usually assumed to be Gaussian distributions, i.e.:

[0168] ;

[0169] where and are the outputs of the encoder, representing the mean and variance of the latent variable; denotes the Gaussian distribution; the decoder is designed according to the types of real-time environmental information of different modalities. For example, for numerical data, a Gaussian distribution is used; for text data, a Bernoulli distribution is used; for image data, a multinomial distribution is used, etc.

[0170] By inputting the real-time environmental information into the multi-modal variational autoencoder, an initial feature vector related to the influencing factors of greenhouse ventilation and the temperature change law can be obtained , which can contain information of various data types inside and outside the greenhouse, reflect the influencing factors of greenhouse ventilation and the temperature change law, and provide a feature representation for greenhouse ventilation decision-making.

[0171] As an alternative embodiment, the present invention adopts a multi-modal attention mechanism, which uses attention weights and a gating mechanism to adaptively weight and fuse the real-time environmental information of different modalities, highlighting important information and related information, and suppressing redundant information and noisy information. The multi-modal attention mechanism effectively improves the fusion effect of multi-modal data and the representation ability of the feature vector.

[0172] Specifically, the following multi-modal attention mechanism can be adopted, which mainly includes:

[0173] First, consider the attention weights: for the real-time environmental information of each modality, calculate its contribution degree to the initial feature vector, that is, the attention weight, denoted as where . The attention weight k of the real-time environmental information of the th modality is calculated by the following formula:

[0174] ;

[0175] Among them, is a learnable function used to evaluate the importance of the real-time environmental information of the th modality, which can be implemented using a neural network or other methods. The attention weight The larger the value, the greater the contribution of the real-time environmental information of the th modality to the initial feature vector, and it should be considered and fused more.

[0176] Then, consider the gating mechanism: For the real-time environmental information of each modality, calculate its reliability for the initial feature vector, that is, the gating mechanism, denoted as , where . The calculation of the gating mechanism is carried out in the following way:

[0177] ;

[0178] Among them, is the sigmoid function, is a learnable function used to evaluate the reliability of the real-time environmental information of the th modality, which can be implemented using a neural network or other methods. The gating mechanism The closer the value is to 1, the more reliable the real-time environmental information of the th modality, and it should be retained and fused more.

[0179] Finally, based on the above-determined attention weights and gating mechanisms, adaptively weighted sum and fusion of the real-time environmental information of different modalities are performed to obtain the final greenhouse ventilation feature vector, denoted as:

[0180] ;

[0181] Among them, is the latent variable obtained by the single-modal encoder for the data of the th modality, that is:

[0182] ;

[0183] Among them, is Gaussian noise, is element-wise multiplication; represents the k th modality data x ( k ) after passing through the single-modal encoder, the mean (or expectation) of the obtained latent variable z ( k ); Indicating the data of the k th modality x ( k ) After passing through the single-modal encoder, the obtained latent variable z ( k ) The covariance matrix (or variance, if the covariance matrix is assumed to be diagonal or scalar).

[0184] Through multi-modal fusion, a more fused feature vector is obtained , which can synthesize the information of data from different modalities, highlight important information and associated information, suppress redundant information and noise information, and provide a more effective feature representation for ventilation decision-making.

[0185] The multi-objective solar greenhouse ventilation decision-making method provided by the present invention uses the multi-modal variational autoencoder (MVAE) algorithm to extract and represent features from data from multiple sensors, and generate a feature vector for greenhouse ventilation. The innovative data processing method can effectively reduce the dimension and complexity of the data, improve the expression and discrimination ability of the data, and has higher accuracy and adaptability compared with traditional data-level fusion and feature-level fusion methods, and can adapt to different types, incomplete and uncertain data.

[0186] As an alternative embodiment, in the multi-objective solar greenhouse ventilation decision-making method provided by the present invention, a preliminary ventilation decision-making strategy is generated according to the greenhouse ventilation feature vector, which mainly includes:

[0187] Based on the soft behavior cloning algorithm, maximize the reward and minimize the difference between the output action and the expert strategy to construct an objective function;

[0188] Determine the preliminary ventilation decision-making strategy according to the greenhouse ventilation feature vector and the reward function determined by the objective function;

[0189] According to the preliminary ventilation decision-making strategy, obtain the output action of greenhouse ventilation.

[0190] Among them, the objective function includes a policy network and a value network respectively constructed by using two neural networks; the policy network outputs Gaussian distribution parameters, and the Gaussian distribution parameters are used to characterize the probability distribution of the output action under the greenhouse ventilation feature vector; the value network outputs a scalar, and the scalar is used to characterize the expected reward value under the greenhouse ventilation feature vector.

[0191] As an alternative embodiment, the present invention adopts the Soft Actor-Critic (SAC) algorithm, which uses a dual neural network to automatically learn and optimize the greenhouse ventilation decision-making strategy according to the feature vector and reward function of greenhouse ventilation, and obtain the output action of greenhouse ventilation. Its function is to perform decision learning and strategy optimization on the feature vector of greenhouse ventilation, and provide an output action for the greenhouse ventilation decision.

[0192] In the structure of soft behavior cloning, is the feature vector of greenhouse ventilation for greenhouse ventilation, is the output action of greenhouse ventilation, is the reward function of greenhouse ventilation, is the policy network, is the value network, and are model parameters. The goal of soft behavior cloning is to maximize the following objective function:

[0193] ;

[0194] where, represents the expectation of the variable ; is the expert demonstration dataset; and are hyperparameters; is the entropy function; is the KL divergence; is the expert policy.

[0195] The meaning of the objective function is to minimize the difference from the expert policy while maximizing the reward and policy entropy. The role of policy entropy is to increase the exploration and robustness of the policy and avoid overfitting to expert data. The role of KL divergence is to constrain the consistency of the policy with expert data and avoid deviating from expert data.

[0196] It should be noted that in order to optimize the objective function, the present invention adopts a dual neural network, that is, two neural networks are used to represent the policy network and the value network respectively. The policy network outputs the parameters of a Gaussian distribution, representing the probability distribution of the output action given the feature vector; the value network outputs a scalar, representing the expected reward value given the feature vector and the output action.

[0197] The parameters of the two networks are updated using the policy gradient method and the least squares method respectively, as follows:

[0198] ;

[0199] ;

[0200] where, and are the parameters of the policy network and the value network respectively; η is the learning rate of the policy network; is the learning rate of the value network; is the objective function of the policy network, which is used to measure the performance of the policy network under the given parameters θ and ϕ ; is the objective function of the value network, and its output under the given parameters ϕ , state x and action a represents the expected return after starting to execute the action x from the state a ; is the expected reward value obtained after executing the action x from the state a ;

[0201] For the policy network, the parameter θ is updated through the policy gradient . The policy gradient is the gradient of the objective function with respect to θ , which indicates how to adjust θ to maximize .

[0202] For the value network, the parameter ϕ is updated by minimizing the squared error between the predicted Q-value and the true reward . The gradient of this error indicates how to adjust ϕ to reduce this error. Note that here it is assumed that is part of the true or target Q-value, and a target network can also be used to stabilize the learning process.

[0203] The multi-objective solar greenhouse ventilation decision-making method provided by the present invention adopts the method of soft behavior cloning imitation learning, which can simultaneously utilize expert demonstration data and environmental feedback data to learn a randomized policy, so that while maximizing the reward, it is also as close as possible to the expert policy.

[0204] Based on the content of the above embodiments, as an optional example, after constructing the objective function, it further includes:

[0205] Optimizing the objective function based on the entropy regularization method includes adding an entropy term to the objective function. The expression of the entire entropy term calculates the negative expectation of these logarithmic probabilities, that is, the entropy of the policy. The larger the entropy, the more evenly the policy network selects different actions in a given state, that is, the higher the uncertainty of the policy, which helps exploration. By adding the entropy term to the objective function and multiplying it by the hyperparameter α , we can control the trade-off between exploration and exploitation.

[0206] The expression of the entropy term is:

[0207] ;

[0208] The expression of the optimized objective function is:

[0209] ;

[0210] where, represents the expectation of the variable, represents the expectation of the action a ; is the greenhouse ventilation feature vector for greenhouse ventilation; is the output action of greenhouse ventilation, which is sampled from the distribution of the policy network in the given state x ; is the reward function for greenhouse ventilation; is the policy network; and are the model parameters; is the expert demonstration dataset; and are the hyperparameters; is the entropy function; is the KL divergence; is the expert policy; is the entropy term of entropy regularization; is the objective function; is the logarithmic probability that the policy network selects the action x in the given state a .

[0211] The objective function also includes the reward function and the KL divergence terms. The reward function is used to evaluate the quality of performing a specific action in a given state, while the KL divergence is used to measure the difference between the policy network and the expert policy , which helps the policy network imitate the expert behavior. The hyperparameter controls the degree of imitating the expert behavior.

[0212] Based on the content of the above embodiments, as an alternative embodiment, after determining the preliminary ventilation decision-making strategy, it further includes:

[0213] Using multi-task learning and meta-gradient update, adaptively adjust the greenhouse ventilation decision-making strategy according to the current greenhouse environment and crop types.

[0214] Based on the above embodiments, the present invention adopts meta-learning to utilize multi-task learning and meta-gradient update to quickly adapt and adjust the greenhouse ventilation decision-making strategy, adapt to the uncertainty and variability of the indoor and outdoor environments of the greenhouse, and achieve long-term optimal control of the temperature in the greenhouse.

[0215] Meta-learning can utilize data from multiple related tasks to learn a general initial model, which can be quickly adjusted and optimized on new tasks, improving learning efficiency and adaptability.

[0216] Regarding different scenarios or conditions related to greenhouse ventilation, such as different seasons, climates, crops, etc., as different tasks, each task has corresponding feature vectors (such as temperature, humidity, light intensity, crop growth stage, etc.), output actions (such as ventilation volume, ventilation time, etc.), and reward functions (such as rewards defined based on crop growth conditions, energy consumption, etc.), constituting a multi-task learning problem.

[0217] The purpose of multi-task learning is to utilize the similarity and complementarity between different tasks to learn a general initial model, which performs well on different tasks and is also adjusted individually according to the characteristics of different tasks.

[0218] Using the method of meta-gradient update, optimize the parameters of the initial model so that it can quickly adapt and optimize on new tasks.

[0219] The method of meta-gradient update means using a meta-optimizer to calculate the gradient of the initial model parameters according to the loss functions of different tasks, and then using the meta-learning rate to update the initial model parameters to obtain an updated model. This model continues to perform gradient descent on new tasks to obtain the final model, which achieves the optimal performance on new tasks. The following details how to adaptively adjust the greenhouse ventilation decision-making strategy according to the current greenhouse environment and crop types by using multi-task learning and meta-gradient update, specifically including:

[0220] First, construct a multi-task learning framework, including: designing a feature vector that can comprehensively reflect the greenhouse environment and crop status, which will serve as the input to the model; constructing a shared model structure (such as a neural network) with multiple layers to extract useful information from the input features; after the shared model, adding specific layers (such as fully connected layers) for each task to output specific actions for that task; defining loss functions for each task, which are calculated based on the reward function of the task and the actions predicted by the model.

[0221] Then, perform meta-learning initialization, including: training a general initial model using meta-learning methods ϕ . This initial model ϕ is pre-trained on multiple tasks to learn the commonalities and basic knowledge between different tasks. Select a meta-optimizer (such as MAM, Reptile, etc.) to optimize the parameters of the initial model on multiple tasks.

[0222] Next, perform meta-gradient update. When facing a new greenhouse ventilation task, use the meta-gradient update method to adjust the initial model ϕ , so as to quickly adapt to the new task, including: starting from the initial model ϕ , creating a task-specific model for each new task θ (usually obtained by copying ϕ ); on each new task, use the task-specific model θ to train to minimize the loss function of this task L ( θ ); use the meta-optimizer to calculate the gradient of the initial model parameters ϕ according to the loss functions on all tasks. This gradient takes into account the influence of all tasks on the initial model ϕ and aims to optimize the initial model ϕ to better adapt to new tasks; use the meta-learning rate α to update the initial model parameters ϕ , that is, ; use the updated initial model ϕ as the starting point to further optimize each new task to obtain the final model θ , that is, .

[0223] Among them, are the initial model parameters, are the updated model parameters, is the loss function of the new task, is the meta-learning rate, is the ordinary learning rate.

[0224] Through training meta - learning, a general initial model is obtained. This initial model is pre - trained on multiple tasks to learn the commonalities and basic knowledge between different tasks. This initial model quickly adapts and optimizes on different greenhouse ventilation tasks, improving the robustness and generalization of the greenhouse ventilation decision - making strategy.

[0225] Finally, according to the ventilation indicators of greenhouse ventilation, the following data can be selected as the following output actions:

[0226] Window - opening time, represented by a binary variable indicating whether greenhouse ventilation is carried out where represents window - opening, represents not window - opening, represents not opening the window.

[0227] Opening degree, represented by a continuous percentage variable indicating the window - opening degree of greenhouse ventilation where represents fully closed, represents fully open, The larger the value of, the larger the window - opening and the larger the ventilation volume.

[0228] Duration, represented by a continuous variable indicating the window - opening time of greenhouse ventilation, with the unit of seconds, where represents not opening the window, The larger the value of, the longer the window - opening time and the larger the ventilation volume.

[0229] The multi - objective solar greenhouse ventilation decision - making method provided by the present invention uses the Soft Actor - Critic (SAC) algorithm and combines meta - learning technology to automatically learn and optimize the greenhouse ventilation decision - making strategy. The innovative decision - making optimization method can effectively balance the temperature control in the greenhouse and the growth needs of crops, as well as energy consumption and utilization. Compared with the traditional set - value method and mathematical model method, it has higher efficiency and performance and can automatically adapt to the meteorological changes inside and outside the greenhouse and the growth stage of crops.

[0230] In summary, through multi - sensor fusion technology, multi - modal variational auto - encoder algorithm, soft behavior cloning algorithm, meta - learning technology, etc., the present invention comprehensively considers multiple objectives such as meteorological parameters inside and outside the greenhouse, physiological needs of crops, and energy benefits of active heat storage and release, realizes multi - objective optimization and coordination of greenhouse ventilation, can effectively balance the temperature control in the greenhouse and the growth needs of crops, as well as energy consumption and utilization, realizes long - term optimal control of the temperature in the greenhouse, improves energy utilization efficiency and crop growth quality, increases the economic benefits of agriculture, realizes long - term optimal control of the temperature in the greenhouse, and improves energy utilization efficiency and crop growth quality.

[0231] Based on the content of the above embodiments, as an alternative embodiment, multi-objective optimization is performed on the preliminary ventilation decision-making strategy, which mainly includes:

[0232] Using the multi-objective particle swarm optimization algorithm and the multi-objective co-evolution algorithm to perform multi-objective optimization on the preliminary decision-making actions, comprehensively considering objectives such as temperature control in the greenhouse, crop growth requirements, operating costs, and effects, to obtain the optimal ventilation decision-making strategy.

[0233] The present invention adopts the multi-objective particle swarm optimization algorithm (MOPSO) algorithm, utilizes particle swarm intelligence and leader selection mechanism, comprehensively considers temperature control in the greenhouse and crop growth requirements, as well as the operating costs and effects of window ventilation and active heat storage and release devices, establishes and solves the multi-objective function of the greenhouse ventilation optimization problem, and obtains the optimal solution for greenhouse ventilation.

[0234] Specifically, in MOPSO, each particle represents a potential solution, and each particle has its own position and velocity. In the structure of the multi-objective particle swarm optimization algorithm, is kinds of time environmental information of different modes; is a latent variable; is an encoder; is a decoder; and are model parameters.

[0235] The goal of the multi-objective particle swarm optimization algorithm is to maximize the marginal log-likelihood of the data, that is:

[0236] ;

[0237] Since the integral in the above formula is unsolvable, a variational inference method is adopted, an approximate posterior distribution is introduced, and the variational lower bound (ELBO) is maximized, that is:

[0238] ;

[0239] Among them, represents the KL divergence, is the prior distribution of the latent variable, usually assumed to be a standard normal distribution.

[0240] For the convenience of calculation, the approximate posterior distribution is constructed in the following way:

[0241] ;

[0242] Among them, is a normalization constant, is a single-modal encoder, usually assumed to be Gaussian distribution, that is:

[0243] ;

[0244] wherein, and are the outputs of the encoder, representing the mean and variance of the latent variables. The decoder is designed according to different modal data types. For example, for numerical data, Gaussian distribution is used; for text data, Bernoulli distribution is used; for image data, multinomial distribution is used, etc.

[0245] By training the multi-objective particle swarm optimization algorithm, the initial feature vector of greenhouse ventilation is obtained. This initial feature vector can contain information of various data types inside and outside the greenhouse, reflect the influencing factors of greenhouse ventilation and the temperature change law, and provide a feature representation for greenhouse ventilation decision-making.

[0246] This step adopts the multi-objective co-evolution algorithm (MOCCA), and uses the co-evolution and decomposition strategy to decompose the multi-objective function and decision variables of the greenhouse ventilation optimization problem into multiple sub-problems and sub-components, and optimize and cooperate respectively to achieve the efficient solution of the greenhouse ventilation optimization problem and the maintenance of high-quality solutions.

[0247] MOCCA is based on the multi-objective optimization method of co-evolution. By decomposing the multi-objective optimization problem into multiple mutually cooperative sub-problems and using different populations to solve different sub-problems, the search efficiency and the diversity of solutions are improved.

[0248] In the structure of MOCCA, is objective functions, is decision variables, is co-evolved populations, is the external archive for saving non-dominated solutions.

[0249] The goal of the multi-objective co-evolution algorithm is to solve the following multi-objective optimization problem:

[0250] ;

[0251] ;

[0252] wherein, is the decision space, is the objective vector.

[0253] For the convenience of calculation, the multi-objective co-evolution algorithm decomposes the multi-objective optimization problem into the following form:

[0254] ;

[0255] ;

[0256] wherein, is the number of the sub - problem, is the th decision variable of the sub - problem, is the decision variable other than the th sub - problem, is the decision space of the th sub - problem, is the decision space other than the th sub - problem, is the objective vector of the th sub - problem, is the th th objective function of the

[0257] Each sub - problem is solved by a population , and the size of the population is , where is a hyper - parameter used to control the search space of each sub - problem. Each individual in the population is represented as , where is the number of the individual, is the th decision variable of the individual. Each individual has its own fitness value, represented as , where is the th objective vector of the individual, that is:

[0258] ;

[0259] wherein, is the collaborative variable of the th individual, that is, the decision variable other than the th sub - problem, and the value of is provided by the individuals of other populations, specifically as follows:

[0260] ;

[0261] wherein, is the number of the individual of other populations, , where and In this way, each individual collaborates with individuals from other populations to form a complete solution, thereby performing fitness evaluation and optimization.

[0262] Furthermore, a co-evolution method is used to enable different populations to cooperate and compete with each other, thereby improving the search efficiency and the diversity of solutions. The co-evolution method includes the following steps:

[0263] (1) Initialization:

[0264] For each population , randomly generate individuals, that is , where , and initialize them as random numbers following a uniform distribution, that is:

[0265] ;

[0266] Among them, is the decision space of the th sub-problem, and is a uniform distribution.

[0267] Then, for each individual, randomly select an individual from other populations as its collaborative variable, that is:

[0268] ;

[0269] Among them, is the number of the randomly selected individual; and .

[0270] Then, according to the decision variable and collaborative variable of the individual, calculate its objective vector and fitness value, that is:

[0271] ;

[0272] (2) Evolution:

[0273] For each population , perform the following operations:

[0274] Use the method of Tournament Selection to select two individuals from population as parental individuals, that is:

[0275] ;

[0276] ;

[0277] Among them, is the function of Tournament Selection, and its process is as follows:

[0278] Randomly select two individuals from the population, namely and , where . and .

[0279] Compare the fitness values of the two individuals, namely and . According to the relationship of Pareto Dominance, select the individual as the winner and return its decision variable, that is:

[0280] ;

[0281] where represents the relationship of Pareto Dominance, that is:

[0282] ;

[0283] represents the relationship of Pareto non - dominance, that is:

[0284] ;

[0285] (3) Mutation:

[0286] Use the method of Gaussian Mutation to randomly perturb the decision variables of the parent individuals to obtain the decision variables of the offspring individuals, that is:

[0287] ;

[0288] ;

[0289] where and are random numbers that follow a Gaussian distribution, that is:

[0290] ;

[0291] ;

[0292] where and are the standard deviations of the mutation, which are hyperparameters used to control the intensity of the mutation.

[0293] Then, perform a boundary check on the decision variables of the offspring individuals to make them satisfy the constraints of the decision space, that is:

[0294] ;

[0295] ;

[0296] Among them, and are the upper and lower bounds of the decision space of the th sub - problem, that is:

[0297] ;

[0298] (4) Crossover:

[0299] Use uniform crossover. Copy the genes at the above - mentioned positions of the parental chromosome 1 to the same positions of the offspring 1, and then fill in the genes missing in the offspring 1 from the parental chromosome 2 in order. The other offspring is obtained in a similar way.

[0300] Furthermore, according to the optimal solution of greenhouse ventilation determined by the above steps, determine the operating parameters of the window ventilation and the active heat storage and release device, such as the window - opening time, opening degree, and duration, as well as the temperature difference between the inlet and outlet, soil heat storage, water - wall heat storage, etc., to achieve the coordinated control and balance of the window ventilation and the active heat storage and release device. The function is to perform operation control and cooperation on the window ventilation and the active heat storage and release device according to the optimal solution of greenhouse ventilation, so as to meet the temperature control in the greenhouse and the growth requirements of crops.

[0301] To improve the effect of the coordinated control and balance of the window ventilation and the active heat storage and release device, the present invention adopts a multi - agent system (MAS), and uses multiple agents and a coordination mechanism to perform distributed control and cooperation on the window ventilation and the active heat storage and release device, so as to achieve the coordination and satisfaction of the temperature control in the greenhouse and the growth requirements of crops. MAS effectively solves the complexity and dynamics of the coordinated control and balance of the window ventilation and the active heat storage and release device.

[0302] Based on the content of the above - mentioned embodiments, as an alternative embodiment, dynamically adjusting the optimal ventilation decision - making strategy to automatically adapt to the meteorological changes inside and outside the greenhouse and the growth stage of crops includes:

[0303] Through the multi - agent system and transfer learning technology, using the similarity and difference between the source task and the target task, the optimal ventilation decision - making strategy is optimized in real - time;

[0304] The target task is determined according to the meteorological changes inside and outside the greenhouse and the growth stage of crops.

[0305] Specifically, according to the optimal solution of greenhouse ventilation, the present invention determines the operating parameters of window ventilation and the active heat storage and release device, such as the window opening time, opening degree, and duration, as well as the temperature difference between the inlet and outlet, soil heat storage, water wall heat storage, etc., to achieve the coordinated control and balance of window ventilation and the active heat storage and release device. The function is to perform operating control and cooperation on window ventilation and the active heat storage and release device according to the optimal solution of greenhouse ventilation, so as to meet the temperature control in the greenhouse and the growth requirements of crops.

[0306] To improve the effect of the coordinated control and balance of window ventilation and the active heat storage and release device, this step adopts a multi-agent system (MAS), and uses multiple agents and coordination mechanisms to perform distributed control and cooperation on window ventilation and the active heat storage and release device, so as to achieve the coordination and satisfaction of temperature control in the greenhouse and the growth requirements of crops. MAS effectively solves the complexity and dynamics of the coordinated control and balance of window ventilation and the active heat storage and release device. The specific content includes:

[0307] (1) Transfer learning of reinforcement learning

[0308] By using the method of transfer learning of reinforcement learning, and utilizing the similarities and differences between the source task and the target task, the greenhouse ventilation decision-making strategy is effectively transferred and adapted, so as to achieve the rapid update and improvement of the greenhouse ventilation decision-making strategy.

[0309] Transfer learning is a learning method based on reinforcement learning. By applying the knowledge learned in an environment to another related environment, the learning efficiency and performance can be improved.

[0310] In the structure of transfer learning of reinforcement learning, is the source task, is the target task, is the data in the source task, is the data in the target task, is the strategy in the source task, is the strategy in the target task, is the function of transfer learning, is the loss function of transfer learning. The goal of transfer learning of reinforcement learning is to learn the function of transfer learning by using the data and strategy in the source task, as well as the data in the target task, so that the strategy in the target task can achieve the optimal performance in the target task, that is:

[0311] ;

[0312] ;

[0313] Among them, is the loss function of transfer learning, which is used to measure the policy in the target task and the optimal policy The gap between is the optimal policy in the target task, that is:

[0314] ;

[0315] Among them, is the initial state distribution in the target task, is the state-action value function in the target task, which represents the state After performing the action According to the policy The expected return that can be obtained.

[0316] By training the transfer learning of reinforcement learning, the policy in the target task is obtained. This policy can utilize the knowledge in the source task and the data in the target task to achieve the rapid update and improvement of the greenhouse ventilation decision-making strategy.

[0317] (2) Greenhouse ventilation decision-making strategy

[0318] Model the greenhouse ventilation decision-making problem as a reinforcement learning problem, that is, a Markov decision process (MDP), which is represented by the quadruple Among them, is the state space, which represents the real-time environmental information inside and outside the greenhouse, including the temperature, humidity, light, wind speed, etc. inside and outside the greenhouse, as well as the crop growth indicators inside the greenhouse, etc.; is the action space, which represents the control actions of greenhouse ventilation, including the window opening time, opening degree, and duration, as well as the operating parameters of the active heat storage and release device, etc.; is the state transition probability, which represents the probability of transferring to the next state After performing the action That is ; ; is the reward function, which represents the immediate reward that can be obtained after performing the action in the state That is Among them is the reward signal, which is usually related to the error of temperature control in the greenhouse, the degree of satisfaction of crop growth requirements, the degree of energy consumption savings, etc.

[0319] Greenhouse ventilation decision-making strategy is the mapping from state to action, which represents selecting the action in the state The probability, that is .

[0320] The goal of the greenhouse ventilation decision-making strategy is to maximize the expected cumulative reward, that is:

[0321] ;

[0322] wherein, is the initial state distribution, is the discount factor, is the th time step reward signal.

[0323] The present invention dynamically adjusts the greenhouse ventilation decision-making strategy through the transfer learning technology of multi-agent system and reinforcement learning, can automatically adapt to the meteorological changes inside and outside the greenhouse and the growth stage of crops, improve the adaptability and robustness of the greenhouse, reduce the operation risk of the greenhouse, realize the energy conservation and environmental protection of the greenhouse, realize the intelligence and adaptability of greenhouse ventilation decision-making, and improve the adaptability and robustness of the greenhouse.

[0324] As an alternative embodiment, an embodiment is provided below to illustrate the multi-objective solar greenhouse ventilation decision-making method provided by the present invention, which mainly includes but is not limited to the following three steps:

[0325] (1) Data acquisition

[0326] Obtain the real-time environmental information inside and outside the greenhouse, including:

[0327] Temperature is an important factor affecting greenhouse ventilation. Monitor the temperature changes inside and outside the greenhouse and the temperature distribution in different areas of the greenhouse. Use temperature sensors DS18B20, LM35, etc. to collect temperature data. Distribute the temperature sensors at appropriate positions inside and outside the greenhouse, such as the top, middle and bottom of the greenhouse, and the four directions of southeast, northwest of the greenhouse. It is also necessary to consider the number and density of temperature sensors to ensure data coverage and accuracy. The unit symbol of temperature data is degree Celsius (°C), the value range is -40°C to 125°C, and the collection frequency is once per minute.

[0328] Humidity is another important factor affecting greenhouse ventilation. Use humidity sensors DHT11 or HIH-4000, etc. to collect humidity data. Distribute the humidity sensors at appropriate positions inside and outside the greenhouse. The unit symbol of humidity data is percentage (%), the value range is 0% to 100%, and the collection frequency is per minute.

[0329] Monitor the changes in light radiation inside and outside the greenhouse, as well as the light radiation distribution in different areas of the greenhouse. Use light sensors to collect light radiation data. Distribute the light sensors at appropriate positions inside and outside the greenhouse. The unit symbol is lux (lx), the value range is 0 lx to 65535 lx, and the collection frequency is once per minute.

[0330] In addition to the above three data types, other data types are also selected for acquisition according to the actual situation, such as:

[0331] Carbon dioxide concentration: The unit symbol is parts per million (ppm), the value range is 0 ppm to 10000 ppm, and the collection frequency is once per minute.

[0332] Wind speed: The unit symbol is meters per second (m / s), the value range is 0 m / s to 30 m / s, and the collection frequency is once per minute.

[0333] Obtain the above data through the following methods or channels:

[0334] Build a sensor network by yourself: Purchase and install sensor devices by yourself, and transmit sensor data to the data center or cloud server wirelessly or wired. Consider factors such as the cost, maintenance, compatibility, stability, and security of sensor devices.

[0335] Use third-party data services: Use third-party data service providers, such as Huawei Cloud, Alibaba Cloud, or Tencent Cloud, to obtain sensor data. Consider factors such as the cost, quality, reliability, availability, and privacy of data services.

[0336] (2)Data processing

[0337] For the obtained real-time environmental information, perform the following processing to be used as the input of the subsequent model:

[0338] Data cleaning, including quality inspection of data related to real-time environmental information, and removing invalid, abnormal, duplicate, missing, incorrect, etc. data to improve the accuracy and integrity of the data. Use Python language and related libraries such as pandas or numpy to perform data cleaning. Use the pandas.DataFrame.dropna() function to delete data rows containing null values, and use the pandas.DataFrame.drop_duplicates() function to delete duplicate data rows.

[0339] Data conversion, including converting data in terms of format, type, range, scale, etc., to meet the input requirements of the model. Use Python libraries such as pandas or scikit-learn for data conversion. Use the pandas.DataFrame.astype() function to change the data type, and use the sklearn.preprocessing.MinMaxScaler() function to normalize the data to the interval [0, 1].

[0340] Data fusion, including fusing data from multiple sources, multiple modalities, multiple time series, etc., to improve the information content and expression ability of the data. Use pandas or PyTorch for data fusion. Use the pandas.DataFrame.merge() function to merge data from different sources, and use the torch.cat() function to concatenate data of different modalities.

[0341] Data storage, including storing data in a structured or unstructured manner, using libraries such as sqlite3 or h5py for data storage. Use the sqlite3.connect() function to create a database file, and use the h5py.File() function to create an HDF5 file.

[0342] (3)Model construction

[0343] To implement the multi-objective greenhouse ventilation decision-making of the present invention, mainly construct the following three model components:

[0344] Multi-modal variational autoencoder (MVAE), as a deep generative model, jointly learns and extracts features from multi-modal data to generate feature vectors for greenhouse ventilation. Use the PyTorch language and related libraries such as torch and torchvision to construct the MVAE model. Define the network structures of the encoder and decoder of MVAE, as well as the loss function and optimizer of MVAE. The core parameter configurations of MVAE are shown in the following table:

[0345]

[0346] Soft actor-critic (SAC) model, an algorithm based on policy-based reinforcement learning, automatically learns and optimizes the greenhouse ventilation decision-making strategy. Use the PyTorch language and related libraries such as torch and stable-baselines3 to construct the SAC model. Define the network structures of the policy network and value function network of SAC, as well as the reward function and optimizer of SAC. The core parameter configurations of SAC are shown in the following table:

[0347]

[0348] Meta-Learning enables the model to quickly adapt to new tasks or environments, improving the model's generalization ability and robustness. Use PyTorch languages such as torch and related libraries to implement meta-learning techniques. Select appropriate meta-learning algorithms such as MAML, Reptile, etc., and appropriate meta-learning tasks such as greenhouse ventilation decision-making tasks or greenhouse temperature prediction tasks. The core parameter configuration of meta-learning is shown in the following table:

[0349]

[0350] Figure 2 is a schematic structural diagram of the multi-objective solar greenhouse ventilation decision-making device provided by the present invention, as Figure 2 shown, mainly including:

[0351] An information acquisition unit 21, used to obtain real-time environmental information inside and outside the greenhouse;

[0352] An information processing unit 22, used to determine the greenhouse ventilation feature vector according to the real-time environmental information;

[0353] A policy generation unit 23, used to generate a preliminary ventilation decision-making policy according to the greenhouse ventilation feature vector;

[0354] A policy optimization unit 24, used to perform multi-objective optimization on the preliminary ventilation decision-making policy so that the optimized optimal ventilation decision-making policy takes into account greenhouse temperature control, crop growth requirements, operating costs and effects;

[0355] A policy matching unit 25, used to dynamically adjust the optimal ventilation decision-making policy to automatically adapt to meteorological changes inside and outside the greenhouse and the growth stage of crops.

[0356] It should be noted that the multi-objective solar greenhouse ventilation decision-making device provided by the present invention, during specific operation, can execute the multi-objective solar greenhouse ventilation decision-making method described in any of the above embodiments, and this embodiment will not be elaborated here.

[0357] The multi-objective solar greenhouse ventilation decision-making device provided by the present invention comprehensively uses technologies such as multi-sensor fusion technology, multi-modal variational autoencoder algorithm, soft behavior cloning algorithm, meta-learning technology, multi-objective particle swarm optimization algorithm, multi-objective co-evolution algorithm, multi-agent system and reinforcement learning transfer learning technology, etc., which can effectively improve the greenhouse ventilation decision-making efficiency, reduce the greenhouse ventilation energy consumption, and improve the intelligent control of the temperature inside the greenhouse and the growth quality of crops.

[0358] Figure 3 is a schematic structural diagram of the electronic device provided by the present invention, as Figure 3As shown in the figure, the electronic device may include: a processor 310, a communications interface 320, a memory 330, and a communication bus 340. Among them, the processor 310, the communications interface 320, and the memory 330 complete communication with each other through the communication bus 340. The processor 310 can call the logical instructions in the memory 330 to execute the multi-objective solar greenhouse ventilation decision-making method, which includes: obtaining the real-time environmental information inside and outside the greenhouse; determining the greenhouse ventilation feature vector according to the real-time environmental information; generating a preliminary ventilation decision-making strategy according to the greenhouse ventilation feature vector; performing multi-objective optimization on the preliminary ventilation decision-making strategy so that the optimized optimal ventilation decision-making strategy takes into account the temperature control inside the greenhouse, the growth requirements of crops, the operating cost and effect; dynamically adjusting the optimal ventilation decision-making strategy to automatically adapt to the meteorological changes inside and outside the greenhouse and the growth stage of crops.

[0359] In addition, when the logical instructions in the above-mentioned memory 330 are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on such an understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or a part of this technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in various embodiments of the present invention. The foregoing storage medium includes: various media such as USB flash drives, mobile hard disks, read-only memories (ROM, Read-Only Memory), random access memories (RAM, Random Access Memory), magnetic disks, or optical discs that can store program codes.

[0360] On the other hand, the present invention also provides a computer program product. The computer program product includes a computer program stored on a non-transitory computer-readable storage medium. The computer program includes program instructions. When the program instructions are executed by a computer, the computer can execute the multi-objective solar greenhouse ventilation decision-making method provided in the above-mentioned various embodiments. The method includes: obtaining the real-time environmental information inside and outside the greenhouse; determining the greenhouse ventilation feature vector according to the real-time environmental information; generating a preliminary ventilation decision-making strategy according to the greenhouse ventilation feature vector; performing multi-objective optimization on the preliminary ventilation decision-making strategy so that the optimized optimal ventilation decision-making strategy takes into account the temperature control inside the greenhouse, the growth requirements of crops, the operating cost and effect; dynamically adjusting the optimal ventilation decision-making strategy to automatically adapt to the meteorological changes inside and outside the greenhouse and the growth stage of crops.

[0361] In another aspect, the present invention also provides a non-transitory computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, it is configured to execute the multi-objective daylight greenhouse ventilation decision-making method provided in the above embodiments. The method includes: obtaining real-time environmental information inside and outside the greenhouse; determining a greenhouse ventilation feature vector according to the real-time environmental information; generating a preliminary ventilation decision-making strategy according to the greenhouse ventilation feature vector; performing multi-objective optimization on the preliminary ventilation decision-making strategy so that the optimized optimal ventilation decision-making strategy takes into account temperature control inside the greenhouse, crop growth requirements, operating costs and effects; and dynamically adjusting the optimal ventilation decision-making strategy to automatically adapt to meteorological changes inside and outside the greenhouse and the growth stage of the crops.

[0362] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed to multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the solution of this embodiment. A person of ordinary skill in the art can understand and implement it without creative effort.

[0363] Through the description of the above embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus a necessary general hardware platform, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. The computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disc, etc., and includes several instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments.

[0364] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions described in the foregoing embodiments or perform equivalent replacements on some of the technical features. These modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A multi-objective solar greenhouse ventilation decision-making method, characterized in that: include: Get real-time environmental information inside and outside the greenhouse; Determining a greenhouse ventilation characteristic vector according to the real-time environmental information; generating a preliminary ventilation decision strategy according to the greenhouse ventilation feature vector; Performing multi-objective optimization on the preliminary ventilation decision strategy so that the optimized optimal ventilation decision strategy takes into account the temperature control in the greenhouse, the crop growth requirements, the operating cost and the effect; Dynamically adjust the optimal ventilation decision strategy to automatically adapt to meteorological changes inside and outside the greenhouse and the growth stage of crops; The generating a preliminary ventilation decision strategy according to the greenhouse ventilation feature vector comprises: Based on the soft behavior cloning algorithm, the objective function is constructed by maximizing the reward and minimizing the difference between the output action and the expert strategy; Determining a preliminary ventilation decision strategy according to the greenhouse ventilation feature vector and a reward function determined by the objective function; According to the preliminary ventilation decision strategy, an output action for greenhouse ventilation is obtained; The objective function includes a policy network and a value network respectively constructed by using two neural networks; the policy network outputs a Gaussian distribution parameter, and the Gaussian distribution parameter is used to characterize the probability distribution of the output action under the greenhouse ventilation feature vector; the value network outputs a scalar, and the scalar is used to characterize the expected reward value under the greenhouse ventilation feature vector; After constructing the objective function, it also includes: The objective function is optimized based on the entropy regularization method, including adding an entropy term to the objective function, and the expression of the entropy term is: The expression of the optimized objective function is: in, represents the expectation of the variable; x is the greenhouse ventilation feature vector of greenhouse ventilation; a is the output action of greenhouse ventilation; r is the reward function of greenhouse ventilation; π θ (a|x) is the policy network; θ and φ are model parameters; D is the expert demonstration dataset; α and β are hyperparameters; is the entropy function; D KL is the KL divergence; π E (·|x) is the expert strategy; is the entropy term of entropy regularization; J(θ,φ) is the objective function.

2. The multi-objective solar greenhouse ventilation decision-making method according to claim 1 is characterized in that: The real-time environmental information includes at least one of the following environmental data: indoor and outdoor temperature, indoor and outdoor humidity, indoor and outdoor light radiation value, wind speed, wind direction, indoor and outdoor images, greenhouse volume, greenhouse height-span ratio, greenhouse comprehensive thermal resistance, indoor and outdoor temperature difference, light transmittance, and saturated water vapor pressure difference; The real-time environmental information inside and outside the greenhouse is obtained, including: The temperature measurements inside and outside the greenhouse are combined by weighted average method to obtain the temperature inside and outside the greenhouse; The temperature and humidity measurement values ​​inside and outside the greenhouse are integrated by weighted average method to obtain the temperature and humidity inside and outside the greenhouse; The light radiation measurement values ​​inside and outside the greenhouse are merged by weighted average method to obtain the light radiation inside and outside the greenhouse; The wind speed measurements inside and outside the greenhouse are combined by weighted average method to obtain the wind speed inside and outside the greenhouse; The wind direction measurements inside and outside the greenhouse are merged by vector averaging method to obtain the wind direction inside and outside the greenhouse; Fusing the real-time images of the inside and outside of the greenhouse at the pixel level or feature level to obtain the indoor and outdoor images; Calculate the volume of the greenhouse according to the length, width, height and other dimensions of the greenhouse; Calculating the height-to-span ratio of the greenhouse according to the height and span of the greenhouse; Calculating the comprehensive thermal resistance of the greenhouse based on the material thermal resistance and surface area of ​​each surface of the greenhouse; Calculating the indoor and outdoor temperature difference according to the indoor and outdoor temperatures of the greenhouse; Calculating the light transmittance according to the indoor and outdoor light radiation values ​​of the greenhouse; The saturated water vapor pressure difference is calculated according to the temperature in the greenhouse and the humidity in the greenhouse.

3. The multi-objective solar greenhouse ventilation decision-making method according to claim 1 is characterized in that: Determining the greenhouse ventilation feature vector according to the real-time environmental information includes: Inputting the real-time environmental information into a multimodal variational autoencoder to obtain an initial feature vector related to the influencing factors of greenhouse ventilation and the law of temperature change; The initial feature vector is adaptively weighted and fused using a multimodal attention mechanism to obtain a greenhouse ventilation feature vector.

4. The multi-objective solar greenhouse ventilation decision-making method according to claim 3 is characterized in that: The multimodal attention mechanism is used to adaptively weight and fuse the initial feature vector to obtain a greenhouse ventilation feature vector, including: Determining attention weights and gating mechanisms for different modalities of real-time environmental information constituting the initial feature vector; According to the attention weight and gating mechanism, the real-time environmental information of different modalities is adaptively weighted and fused to obtain the greenhouse ventilation feature vector.

5. The multi-objective solar greenhouse ventilation decision-making method according to claim 1 is characterized in that: After determining the initial ventilation decision strategy, it also includes: Multi-task learning and meta-gradient updating are used to adapt the greenhouse ventilation decision strategy according to the current greenhouse environment and crop types.

6. The multi-objective solar greenhouse ventilation decision-making method according to claim 1 is characterized in that: The preliminary ventilation decision strategy is optimized by multiple objectives, including: The multi-objective particle swarm optimization algorithm and the multi-objective co-evolution algorithm are used to perform multi-objective optimization on the preliminary decision-making actions, and the objectives of greenhouse temperature control, crop growth requirements, operating costs and effects are comprehensively considered to obtain the optimal ventilation decision strategy.

7. The multi-objective solar greenhouse ventilation decision-making method according to claim 1 is characterized in that: The optimal ventilation decision strategy is dynamically adjusted to automatically adapt to the meteorological changes inside and outside the greenhouse and the growth stage of the crops, including: By using a multi-agent system and transfer learning technology, the optimal ventilation decision strategy is optimized in real time by utilizing the similarities and differences between the source task and the target task; The target task is determined based on the meteorological changes inside and outside the greenhouse and the growth stage of the crops.

8. A multi-objective solar greenhouse ventilation decision-making device, characterized in that: include: Information collection unit, used to obtain real-time environmental information inside and outside the greenhouse; An information processing unit, used for determining a greenhouse ventilation characteristic vector according to the real-time environmental information; A strategy generating unit is used to generate a preliminary ventilation decision strategy according to the greenhouse ventilation feature vector, specifically comprising: Based on the soft behavior cloning algorithm, the objective function is constructed by maximizing the reward and minimizing the difference between the output action and the expert strategy; Determining a preliminary ventilation decision strategy according to the greenhouse ventilation feature vector and a reward function determined by the objective function; According to the preliminary ventilation decision strategy, an output action for greenhouse ventilation is obtained; The objective function includes a policy network and a value network respectively constructed by using two neural networks; the policy network outputs a Gaussian distribution parameter, and the Gaussian distribution parameter is used to characterize the probability distribution of the output action under the greenhouse ventilation feature vector; the value network outputs a scalar, and the scalar is used to characterize the expected reward value under the greenhouse ventilation feature vector; After constructing the objective function, it also includes: The objective function is optimized based on the entropy regularization method, including adding an entropy term to the objective function, and the expression of the entropy term is: The expression of the optimized objective function is: in, represents the expectation of the variable; x is the greenhouse ventilation feature vector of greenhouse ventilation; a is the output action of greenhouse ventilation; r is the reward function of greenhouse ventilation; π θ (a|x) is the policy network; θ and φ are model parameters; D is the expert demonstration dataset; α and β are hyperparameters; is the entropy function; D KL is the KL divergence; π E (·|x) is the expert strategy; is the entropy term of entropy regularization; J(θ,φ) is the objective function; A strategy optimization unit, used for performing multi-objective optimization on the preliminary ventilation decision strategy, so that the optimized optimal ventilation decision strategy takes into account the temperature control in the greenhouse, the crop growth requirements, the operation cost and the effect; The strategy matching unit is used to dynamically adjust the optimal ventilation decision strategy to automatically adapt to the meteorological changes inside and outside the greenhouse and the growth stage of the crops.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the computer program, the multi-objective solar greenhouse ventilation decision method as described in any one of claims 1 to 7 is implemented.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the multi-objective solar greenhouse ventilation decision method as described in any one of claims 1 to 7 is implemented.

11. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the multi-objective solar greenhouse ventilation decision method as described in any one of claims 1 to 7 is implemented.

Images