Greenhouse operation decision-making method, related device, equipment and storage medium

Abstract

The invention discloses a greenhouse operation decision-making method, a related device, equipment and a storage medium, and the method comprises the steps: obtaining a greenhouse operation decision-making model obtained through reinforcement learning training based on the sample growth data of a target greenhouse crop; wherein the sample growth data comprises state parameters of the target greenhouse crop at a second sample moment after the target greenhouse crop adopts the action parameters under the state parameters of the first sample moment, and the state parameters comprise outdoor parameters, indoor parameters, water and fertilizer parameters, growth vigor parameters, pest and disease damage parameters and yield parameters; the action parameters comprise greenhouse control actions and farming activity actions; and processing the state parameter of the target greenhouse crop at the current moment based on the greenhouse operation decision model to obtain an action parameter to be executed on the target greenhouse crop at the current moment. According to the scheme, the greenhouse operation decision can be predicted in real time so as to realize greenhouse closed-loop decision, and yield and cost are considered at the same time.

Description

Technical Field

[0001] The present application relates to the field of artificial intelligence technology, and in particular to a greenhouse operation decision-making method and related devices, equipment and storage media. Background Art

[0002] At present, most greenhouses on the market are still managed manually, with low automation and intelligence. Due to the uneven technical level of individuals, the material resources and human resources are used irrationally during the planting process of greenhouse crops, resulting in waste of resources, and the greenhouse crops cannot get an ideal growth environment, resulting in low yields.

[0003] Based on this, currently, the backwardness and shortcomings of manual management are mainly improved through methods such as expert experience guidance and preset rule control. However, various existing technical methods are difficult to take into account both output and cost at the same time, and it is also difficult to form a greenhouse closed-loop decision. In view of this, how to predict greenhouse operation decisions in real time to achieve greenhouse closed-loop decision-making and take into account both output and cost at the same time has become an urgent problem to be solved. Summary of the invention

[0004] The main technical problem solved by the present application is to provide a greenhouse operation decision-making method and related devices, equipment and storage media, which can predict greenhouse operation decisions in real time to achieve greenhouse closed-loop decision-making while taking into account both output and cost.

[0005] In order to solve the above technical problems, the first aspect of the present application provides a greenhouse operation decision-making method, including: obtaining a greenhouse operation decision model based on sample growth data of target greenhouse crops and obtained through reinforcement learning training; wherein the sample growth data includes the state parameters of the target greenhouse crops at a second sample moment after adopting action parameters under the state parameters of the first sample moment, the state parameters include outdoor parameters, indoor parameters, water and fertilizer parameters, growth parameters, pest and disease parameters and yield parameters, and the action parameters include greenhouse control actions and farming activity actions; based on the greenhouse operation decision model, the state parameters of the target greenhouse crops at the current moment are processed to obtain the action parameters to be executed on the target greenhouse crops at the current moment.

[0006] In order to solve the above-mentioned technical problems, the second aspect of the present application provides a greenhouse operation decision-making device, including: an acquisition module and a decision module, the acquisition module is used to obtain a greenhouse operation decision model obtained by reinforcement learning training based on the sample growth data of the target greenhouse crop; wherein the sample growth data includes the state parameters of the target greenhouse crop at the second sample moment after the action parameters are adopted under the state parameters of the first sample moment, the state parameters include outdoor parameters, indoor parameters, water and fertilizer parameters, growth parameters, pest and disease parameters and yield parameters, and the action parameters include greenhouse control actions and farming activity actions; the decision module is used to process the state parameters of the target greenhouse crop at the current moment based on the greenhouse operation decision model, and obtain the action parameters to be executed on the target greenhouse crop at the current moment.

[0007] In order to solve the above technical problems, the third aspect of the present application provides an electronic device, which at least includes a memory and a processor coupled to each other, wherein the memory at least stores program instructions, and the processor is used to execute the program instructions to implement the greenhouse operation decision method in the above first aspect.

[0008] In order to solve the above technical problems, the fourth aspect of the present application provides a computer-readable storage medium storing program instructions that can be executed by a processor, and the program instructions are used to implement the greenhouse operation decision-making method of the first aspect.

[0009] The above scheme obtains a greenhouse operation decision model obtained by reinforcement learning training based on sample growth data of target greenhouse crops. The sample growth data includes the state parameters of the target greenhouse crops at the second sample moment after the action parameters are adopted under the state parameters of the first sample moment. The state parameters include outdoor parameters, indoor parameters, water and fertilizer parameters, growth parameters, pest and disease parameters and yield parameters. The action parameters include greenhouse control actions and farming activity actions. The state parameters of the target greenhouse crops at the current moment are processed based on the greenhouse operation decision model to obtain the action parameters to be executed on the target greenhouse crops at the current moment. On the one hand, the state parameters at the current moment can be processed by the greenhouse operation decision model trained by reinforcement learning to predict the action parameters to be executed, and it can be real-time. Predict greenhouse operation decisions, and because the action parameters include greenhouse control actions and agricultural activity actions, it can take into account multiple aspects of greenhouse operation decisions. Compared with single decisions such as greenhouse environmental control action decisions, it helps to achieve greenhouse closed-loop decisions. On the other hand, because the greenhouse operation decision model is trained with sample growth data, and the sample growth data includes the state parameters of the target greenhouse crop at the second sample moment after the action parameters are adopted under the state parameters of the first sample moment, reinforcement learning training can force the greenhouse operation decision model to learn how the execution of specific action parameters under specific state parameters affects subsequent state parameters, which helps the greenhouse operation decision model to take into account the yield represented by the yield parameter in the state parameters and the cost represented by the action parameters. Therefore, it is possible to predict greenhouse operation decisions in real time to achieve greenhouse closed-loop decisions and take into account both yield and cost. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] Figure 1 It is a flow chart of an embodiment of a greenhouse operation decision-making method of the present application; Figure 2a It is a schematic diagram of a process of obtaining growth parameters in a greenhouse operation decision-making process; Figure 2b is a schematic diagram of a process of obtaining pest and disease parameters in a greenhouse operation decision-making process; Figure 2c It is a schematic diagram of the principle of an embodiment of a greenhouse climate control simulation model; Figure 2d It is a schematic diagram of the principle of an embodiment of a greenhouse water and fertilizer control simulation model; Figure 2e It is a schematic diagram of the principle of an embodiment of a crop growth state simulation model; Figure 2f is a schematic diagram of another embodiment of a crop growth state simulation model; Figure 2g It is a schematic diagram of the principle of an embodiment of a farming activity simulation model; Figure 2hIt is a schematic diagram of the principle of an embodiment of a crop yield simulation model; Figure 2i It is a schematic diagram of the interaction between the greenhouse operation decision model and the growth simulation model; Figure 3 This is a process diagram of an embodiment of a greenhouse operation decision-making method of the present application; Figure 4 It is a schematic diagram of the framework of an embodiment of the greenhouse operation decision-making device of the present application; Figure 5 It is a schematic diagram of the framework of an embodiment of the electronic device of the present application; Figure 6 It is a schematic diagram of a framework of an embodiment of a computer-readable storage medium of the present application. DETAILED DESCRIPTION

[0011] The scheme of the embodiment of the present application is described in detail below in conjunction with the drawings of the specification.

[0012] In the following description, for the purpose of explanation rather than limitation, specific details such as specific system structures, interfaces, and technologies are provided to facilitate a thorough understanding of the present application.

[0013] The terms "system" and "network" are often used interchangeably in this article. The term "and / or" in this article is only a description of the association relationship of associated objects, indicating that there can be three relationships. For example, A and / or B can mean: A exists alone, A and B exist at the same time, and B exists alone. In addition, the fragment " / " in this article generally indicates that the associated objects before and after are in an "or" relationship. In addition, "many" in this article means two or more than two.

[0014] See also Figure 1 , Figure 1 This is a flow chart of an embodiment of the greenhouse operation decision-making method of the present application. Specifically, it may include the following steps: Step S11: obtaining a greenhouse operation decision model based on sample growth data of target greenhouse crops through reinforcement learning training.

[0015] In the disclosed embodiment, the sample growth data may include the state parameters of the target greenhouse crop at the second sample time after the action parameters are adopted under the state parameters of the first sample time, the state parameters include outdoor parameters, indoor parameters, water and fertilizer parameters, growth parameters, pest and disease parameters and yield parameters, and the action parameters include greenhouse control actions and farming activity actions. It should be noted that in the disclosed embodiment, "target greenhouse crops" may refer to the collection of all target crops in the greenhouse, or may refer to a single target crop in the greenhouse, or may refer to part of the target crops in the greenhouse, without limitation here.

[0016] In one implementation scenario, the target greenhouse crops can be set according to actual application conditions. For example, the target greenhouse crops can include but are not limited to tomatoes, leeks, cucumbers, etc., and the specific types of the target greenhouse crops are not limited here.

[0017] In one implementation scenario, as shown in Table 1, outdoor parameters may include but are not limited to: outdoor temperature, outdoor humidity, solar radiation, wind speed and direction, rainfall, total radiation, photosynthetically active radiation, etc. As a possible example, in the greenhouse operation decision-making process, outdoor parameters can be collected by a small meteorological station. Of course, the above example is only a possible way to collect outdoor parameters in actual application, and the way to collect outdoor parameters is not limited here.

[0018] Table 1 Schematic table of outdoor parameters and indoor parameters of an embodiment

[0019] In one implementation scenario, as shown in Table 1, indoor parameters may include but are not limited to: indoor temperature, indoor humidity, carbon dioxide concentration, vent status, shade curtain status, heating pipe temperature, carbon dioxide generator supplement value, photosynthetically active radiation, fill light radiation value, etc. As a possible example, in the greenhouse operation decision-making process, indoor parameters can be collected by temperature and humidity sensors and other related sensors. Of course, the above example is only a possible way to collect indoor parameters in actual application, and the way to collect indoor parameters is not limited here.

[0020] In one implementation scenario, as shown in Table 2, water and fertilizer parameters may include but are not limited to: irrigation water pH value, irrigation water EC value, irrigation water ion concentration, real-time irrigation status, irrigation water flow, etc. As a possible example, in the greenhouse operation decision-making process, water and fertilizer parameters can be collected by water and fertilizer sensors. Of course, the above example is only a possible way to collect water and fertilizer parameters in actual application, and the collection method of water and fertilizer parameters is not limited here.

[0021] Table 2 Schematic table of water and fertilizer parameters in an embodiment

[0022] In one implementation scenario, as shown in Table 3, the growth parameters may include but are not limited to: stem diameter, planting density, growth stage, leaf area index, fruit number, fruit size, etc. As a possible example, in the greenhouse operation decision-making process, the growth parameters can be identified by computer vision technology. Figure 2a , Figure 2a FIG. 1 is a schematic diagram of an embodiment of obtaining growth parameters in a greenhouse operation decision-making process. Figure 2aAs shown, crop images of target greenhouse crops can be obtained from a variety of data acquisition devices (such as the aforementioned various sensors and cameras), and then the crop images can be evaluated using the crop growth assessment model to obtain crop growth assessment results including growth parameters. It should be noted that the crop growth assessment model can include but is not limited to convolutional neural networks, etc., and the network structure of the crop growth assessment model is not limited here.

[0023] Table 3 Schematic table of growth parameters of an embodiment

[0024] In one implementation scenario, as shown in Table 4, the pest and disease parameters may include but are not limited to: disease occurrence degree, disease category, pest occurrence degree, pest category, etc. As a possible example, in the greenhouse operation decision-making process, the pest and disease parameters can be identified by computer vision technology. Figure 2b , Figure 2b FIG. 1 is a schematic diagram of an embodiment of obtaining pest and disease parameters in the greenhouse operation decision-making process. Figure 2b As shown, crop images can be obtained from a variety of RGB data acquisition devices, and the crop images can be identified using a pest and disease recognition model to obtain pest and disease recognition results containing pest and disease parameters. It should be noted that the pest and disease recognition model can include but is not limited to a convolutional neural network, etc., and the network structure of the pest and disease recognition model is not limited here.

[0025] Table 4 Schematic table of pest and disease parameter one embodiment

[0026] In one implementation scenario, the greenhouse control action includes but is not limited to: environmental control action, water and fertilizer control action, etc. For example, the environmental control action can be used to control the fans, water curtains and other equipment in the greenhouse to adjust the greenhouse temperature, and can also be used to control the skylights and other equipment in the greenhouse to adjust the greenhouse humidity, and can also be used to control the sunshades and other equipment in the greenhouse to adjust the greenhouse temperature, humidity and radiation, etc., which are not listed one by one here; or, the water and fertilizer control action can be used to control the various parameters of the irrigation liquid, the irrigation state, etc., for example, drip irrigation, sprinkler irrigation, etc. can be regulated, and the concentration, flow rate, and ratio of the irrigation liquid can also be regulated, which are not listed one by one here.

[0027] In one implementation scenario, as shown in Table 5, agricultural activity actions may include but are not limited to: thinning flowers and fruits, removing old leaves, dropping vines, harvesting, pest and disease management, etc. As a possible example, in the greenhouse operation decision-making process, agricultural activity actions can be obtained through the log data collection of the greenhouse agricultural management system, and can also be obtained by using computer vision technology to identify greenhouse images. The collection method of agricultural activity actions is not limited here.

[0028] Table 5 Schematic diagram of an embodiment of a farming activity action

[0029] In one implementation scenario, as a possible implementation method, the sample growth data can be actually collected during the process of planting the target greenhouse crop. As another possible implementation method, the sample growth data can be generated by a growth simulation model of the target greenhouse crop, so that the sample growth data of the target greenhouse crop can be obtained as quickly and accurately as possible without having to follow the actual growth of the target greenhouse crop and wait for a long time. As another possible implementation method, the sample growth data can also be obtained by combining the above two methods. The method for obtaining the sample growth data set is not limited here. For ease of description, the first sample moment is recorded as t, the second sample moment after the first sample moment is recorded as t+1, the outdoor parameters are recorded as W, the indoor parameters are recorded as C, the water and fertilizer parameters are recorded as N, the growth parameters are recorded as G, the pest and disease parameters are recorded as D, the yield parameters are recorded as P, the greenhouse control action is recorded as A, and the agricultural activity action is recorded as B.

[0030] In one implementation scenario, taking the example that the sample growth data is generated by at least a growth simulation model of a target greenhouse crop, the growth simulation model may include a greenhouse climate control simulation model for simulating indoor parameters in the state parameters.

[0031] In a specific implementation scenario, the greenhouse climate control simulation model may include but is not limited to multi-layer neural networks, such as convolutional neural networks, recurrent neural networks, long short-term memory networks, etc. The network structure of the greenhouse climate control simulation model is not limited here.

[0032] In a specific implementation scenario, please refer to Figure 2c , Figure 2c Schematic diagram of a greenhouse climate control simulation model. Figure 2c As shown, the indoor parameters at the first sample time can be obtained , Outdoor parameters And greenhouse control actions The environmental control actions in the first sample time are based on the indoor parameters , Outdoor parameters And greenhouse control actions The environmental control action in the prediction is used to obtain the indoor parameters at the second sample time. The above method first obtains the indoor parameters, outdoor parameters and environmental control actions in the greenhouse control action at the first sample moment, and then predicts the indoor parameters at the second sample moment based on the obtained indoor parameters. This method can fully consider the influence of outdoor parameters and environmental control actions on indoor parameters, that is, the future indoor parameters are affected by the current outdoor parameters and previous indoor parameters, as well as the environmental control actions, which helps to improve the accuracy of modeling indoor parameters.

[0033] In a specific implementation scenario, in order to improve the accuracy of the greenhouse climate control simulation model, the greenhouse climate control simulation model can also be trained in advance. Exemplarily, a plurality of sets of training sample data related to indoor parameters in a small amount of sample growth data can be collected first, and each set of training sample data can include outdoor parameters, indoor parameters, environmental control actions in greenhouse control actions, and indoor parameters at the next moment. On this basis, the greenhouse climate control simulation model can be trained using each set of training sample data. For example, the outdoor parameters, indoor parameters, and environmental control actions in greenhouse control actions at a certain moment in a certain set of training sample data can be input into the greenhouse climate control simulation model to predict the indoor parameters at the next moment, and then the network parameters of the greenhouse climate control simulation model can be adjusted based on the difference between the real indoor parameters at the next moment in the set of training sample data and the predicted indoor parameters at the next moment. This cycle is repeated until the greenhouse climate control simulation model training converges, and the greenhouse climate control simulation model can be forced to learn to accurately predict the indoor parameters at the next moment based on the outdoor parameters, indoor parameters, and environmental control actions in greenhouse control actions at the current moment.

[0034] In one implementation scenario, taking the example that the sample growth data is generated by at least a growth simulation model of a target greenhouse crop, the growth simulation model may include a greenhouse water and fertilizer control simulation model for simulating water and fertilizer parameters in the state parameters.

[0035] In a specific implementation scenario, the greenhouse water and fertilizer control simulation model may include but is not limited to multi-layer neural networks, such as convolutional neural networks, recurrent neural networks, long short-term memory networks, etc. The network structure of the greenhouse water and fertilizer control simulation model is not limited here.

[0036] In a specific implementation scenario, please refer to Figure 2d , Figure 2d Schematic diagram of the principle of an embodiment of a greenhouse water and fertilizer control simulation model. Figure 2d As shown, the water and fertilizer parameters at the first sample time can be obtained , Growth parameters And greenhouse control actions The water and fertilizer control action in the first sample moment is then based on the water and fertilizer parameters , Growth parameters And greenhouse control actions The water and fertilizer control actions in the prediction are used to obtain the water and fertilizer parameters at the second sample time. The above method first obtains the water-fertilizer parameters, growth parameters and water-fertilizer regulation actions in the greenhouse control action at the first sample moment, and then predicts the water-fertilizer parameters at the second sample moment based on the water-fertilizer parameters, growth parameters and water-fertilizer regulation actions at the first sample moment. This method can fully consider the factors affecting the growth parameters and water-fertilizer regulation actions on the water-fertilizer parameters, that is, the future water-fertilizer parameters are affected by the current water-fertilizer parameters and previous growth parameters, as well as the water-fertilizer regulation actions, which helps to improve the accurate modeling of water-fertilizer parameters.

[0037] In a specific implementation scenario, in order to improve the accuracy of the greenhouse water and fertilizer control simulation model, the greenhouse water and fertilizer control simulation model can also be trained in advance. Exemplarily, a small amount of sample growth data can be first collected in real terms and multiple groups of training sample data related to water and fertilizer parameters, each group of training sample data can include water and fertilizer parameters at a certain moment, growth parameters, water and fertilizer regulation actions in greenhouse control actions, and water and fertilizer parameters at the next moment. On this basis, each group of training sample data can be used to train the greenhouse water and fertilizer control simulation model. For example, the water and fertilizer parameters, growth parameters, and water and fertilizer regulation actions in greenhouse control actions at a certain moment in a certain group of training sample data can be input into the greenhouse water and fertilizer control simulation model to predict the water and fertilizer parameters at the next moment, and then the network parameters of the greenhouse water and fertilizer control simulation model can be adjusted based on the difference between the real water and fertilizer parameters at the next moment in the group of training sample data and the water and fertilizer parameters predicted at the next moment. In this way, the greenhouse water and fertilizer control simulation model is iterated until the training of the greenhouse water and fertilizer control simulation model converges, so that the greenhouse water and fertilizer control simulation model can be forced to learn to accurately predict the water and fertilizer parameters at the next moment based on the water and fertilizer parameters, growth parameters, and water and fertilizer regulation actions in greenhouse control actions at the current moment.

[0038] In one implementation scenario, taking the example that the sample growth data is generated by at least a growth simulation model of a target greenhouse crop, the growth simulation model may include a crop growth state simulation model for simulating growth potential parameters in the state parameters.

[0039] In a specific implementation scenario, the crop growth status simulation model may include but is not limited to multi-layer neural networks, such as convolutional neural networks, recurrent neural networks, long short-term memory networks, etc. The network structure of the crop growth status simulation model is not limited here.

[0040] In a specific implementation scenario, please refer to Figure 2e , Figure 2eSchematic diagram of the principle of an embodiment of a crop growth state simulation model. Figure 2d As shown, the indoor parameters at the first sample time can be obtained , water and fertilizer parameters and growth parameters , and then based on the indoor parameters at the first sample time , water and fertilizer parameters and growth parameters , predict the growth parameters at the second sample time In the above method, the indoor parameters, water-fertilizer parameters and growth parameters at the first sample time are first obtained, and then the growth parameters at the second sample time are predicted based on the indoor parameters, water-fertilizer parameters and growth parameters at the first sample time. This method can fully consider the influencing factors of water-fertilizer parameters and indoor parameters, that is, the future growth parameters are affected by the current growth parameters, water-fertilizer parameters and indoor parameters, which helps to improve the accuracy of modeling growth parameters.

[0041] In a specific implementation scenario, as a specific example, the indoor parameters, water and fertilizer parameters and growth parameters at the first sample moment can be processed based on the graph attention mechanism to obtain a first feature representation, and the indoor parameters, water and fertilizer parameters and growth parameters at the first sample moment can be processed based on the multi-head attention mechanism to obtain a second feature representation, and then the first feature representation and the second feature representation are fused to obtain a fused feature representation, so that autoregressive decoding can be performed based on the fused feature representation to obtain the growth parameters at the second sample moment. It should be noted that the processing process of the graph attention mechanism and the multi-head attention mechanism can refer to the technical details of the graph attention mechanism and the multi-head attention mechanism respectively, which will not be repeated here. In addition, the specific process of autoregressive decoding can refer to the technical details of autoregressive decoding in network architectures such as "encoder-decoder", which will not be repeated here. The above method, by combining the multi-head attention mechanism and the graph attention mechanism, can not only consider the direct interaction between multiple parameters in the crop growth state simulation model, but also consider the indirect interaction of multiple parameters with intermediate variables as the bridge, such as the impact of climate change and water and fertilizer changes on crop growth is reflected in different aspects and has a mutual influence. Furthermore, the indoor parameters and water and fertilizer parameters of multiple time periods can be used to accumulate and form growth parameters, which helps to improve the accuracy of growth parameters. Figure 2f , Figure 2f FIG. 1 is a schematic diagram of another embodiment of the crop growth state simulation model. Figure 2fAs shown in the figure, the crop growth state simulation model can specifically adopt the Transformer-based "encoder-decoder" architecture, that is, the "encoder-decoder" architecture, and replace the multi-head attention in the Transformer-based "encoder-decoder" with multi-head attention (Multi-Head Attention) and graph attention (Graph Attention). For the rest, please refer to the technical details of the Transformer-based "encoder-decoder", which will not be repeated here. Please continue to refer to Figure 2f Based on the graph attention mechanism, the indoor parameters, water and fertilizer parameters and growth parameters of the first sample moment are processed to obtain a feature vector. Based on the multi-head attention mechanism, the indoor parameters, water and fertilizer parameters and growth parameters of the first sample moment are processed to obtain another feature vector. The two feature vectors are then fused (e.g., concatenated) to be processed by subsequent network layers until the autoregressive decoding is completed, and the growth parameters of the second sample moment can be obtained. The specific connotation of the growth parameters can be found in the above-mentioned related description, which will not be repeated here. It should be noted that Figure 2f The crop growth state simulation model shown is only a possible example of the crop growth state simulation model in actual application, and does not limit the specific structure of the crop growth state simulation model, which will not be given one by one here.

[0042] In a specific implementation scenario, in order to improve the accuracy of the crop growth state simulation model, the crop growth state simulation model can also be trained in advance. For example, a small amount of sample growth data can be used to collect multiple groups of training sample data related to growth parameters, and each group of training sample data can include water and fertilizer parameters, growth parameters, indoor parameters at a certain moment, and growth parameters at the next moment. On this basis, the crop growth state simulation model can be trained using each group of training sample data. For example, the water and fertilizer parameters, growth parameters, and indoor parameters at a certain moment in a certain group of training sample data can be input into the crop growth state simulation model to predict the growth parameters at the next moment, and then the network parameters of the crop growth state simulation model can be adjusted based on the difference between the actual growth parameters at the next moment in the group of training sample data and the predicted growth parameters at the next moment. This cycle is repeated until the crop growth state simulation model training converges, and the crop growth state simulation model can be forced to learn to accurately predict the growth parameters at the next moment based on the water and fertilizer parameters, growth parameters, and indoor parameters at the current moment.

[0043] In one implementation scenario, taking the example that the sample growth data is generated by at least a growth simulation model of a target greenhouse crop, the growth simulation model may include an agricultural activity simulation model for simulating agricultural activity actions in the action parameters.

[0044] In a specific implementation scenario, the agricultural activity simulation model may include but is not limited to multi-layer neural networks, such as convolutional neural networks, recurrent neural networks, long short-term memory networks, etc. The network structure of the agricultural activity simulation model is not limited here.

[0045] In a specific implementation scenario, please refer to Figure 2g , Figure 2g Schematic diagram of the principle of an embodiment of a farming activity simulation model. Figure 2g As shown, the pest and disease parameters at the first sample moment can be obtained and growth parameters , and then based on the pest and disease parameters at the first sample moment and growth parameters , predict the agricultural activity action at the second sample time The above method first obtains the growth parameters and yield parameters at the first sample time, and then predicts the yield parameters at the second sample time based on the growth parameters and yield parameters at the first sample time. This method can fully consider the influencing factors of pest and disease parameters and growth parameters, that is, future agricultural activities are affected by current growth parameters and pest and disease parameters, which helps to improve the accuracy of modeling agricultural activities.

[0046] In a specific implementation scenario, in order to improve the accuracy of the agricultural activity simulation model, the agricultural activity simulation model can also be trained in advance. Exemplarily, a small amount of sample growth data can be first collected in real terms, and multiple groups of training sample data related to growth parameters, each group of training sample data can include pest and disease parameters and growth parameters at a certain moment and agricultural activity actions at the next moment. On this basis, each group of training sample data can be used to train the agricultural activity simulation model. For example, the pest and disease parameters and growth parameters at a certain moment in a certain group of training sample data can be input into the agricultural activity simulation model to predict the agricultural activity actions at the next moment, and then the network parameters of the agricultural activity simulation model can be adjusted based on the difference between the actual agricultural activity actions at the next moment in the group of training sample data and the agricultural activity actions predicted at the next moment. This cycle is repeated until the agricultural activity simulation model training converges, and the agricultural activity simulation model can be forced to learn to accurately predict the agricultural activity actions at the next moment based on the pest and disease parameters and growth parameters at the current moment.

[0047] In one implementation scenario, taking the example that the sample growth data is generated by at least a growth simulation model of a target greenhouse crop, the growth simulation model may include a crop yield simulation model for simulating a yield parameter in a state parameter.

[0048] In a specific implementation scenario, the crop yield simulation model may include but is not limited to multi-layer neural networks, such as convolutional neural networks, recurrent neural networks, long short-term memory networks, etc. The network structure of the crop yield simulation model is not limited here.

[0049] In a specific implementation scenario, please refer to Figure 2h , Figure 2h Schematic diagram of the principle of an embodiment of a crop yield simulation model. Figure 2h As shown, the growth parameters at the first sample moment can be obtained and yield parameters , based on the growth parameters of the first sample moment and yield parameters , predict the production parameters at the second sample time The above method first obtains the growth parameters and yield parameters of the first sample moment, and then predicts the yield parameters of the second sample moment based on the growth parameters and yield parameters of the first sample moment. This method can fully consider the influencing factors of the growth parameters, that is, the future yield parameters are affected by the current yield parameters and growth parameters, which helps to improve the accuracy of modeling yield parameters.

[0050] In a specific implementation scenario, in order to improve the accuracy of the crop yield simulation model, the crop yield simulation model can also be trained in advance. For example, a plurality of groups of training sample data related to the yield parameters in a small amount of sample growth data can be collected first, and each group of training sample data can include the yield parameters and growth parameters at a certain moment and the yield parameters at the next moment. On this basis, the crop yield simulation model can be trained using each group of training sample data. For example, the yield parameters and growth parameters at a certain moment in a certain group of training sample data can be input into the crop yield simulation model to predict the yield parameters at the next moment, and then the network parameters of the crop yield simulation model can be adjusted based on the difference between the actual yield parameters at the next moment in the group of training sample data and the predicted yield parameters at the next moment. This cycle is repeated until the crop yield simulation model training converges, and the crop yield simulation model can be forced to learn to accurately predict the yield parameters at the next moment based on the yield parameters and growth parameters at the current moment.

[0051] It should be noted that the above is a separate description of each model in the growth simulation model, namely the greenhouse climate control simulation model, the greenhouse water and fertilizer control simulation model, the crop growth state simulation model, the farming activity simulation model and the crop yield simulation model. When simulating sample growth data using the growth simulation model, the growth data can be continuously simulated and generated from the initial stage of the target greenhouse crop until the iteration end condition is met. Exemplarily, the iteration end condition may include, but is not limited to: the accumulated sample growth data is not less than a preset quantity threshold, the "growth stage" in the growth parameters of the target greenhouse crop is "death", etc., and the iteration end condition is not limited here. Specifically, the current sample moment can be selected as the first sample moment, and the sample prediction parameters predicted at the previous sample moment at the current sample moment can be selected as the sample current parameters at the current sample moment, and the outdoor parameters, pest and disease parameters and greenhouse control actions at the current sample moment can be simulated and generated. Then, based on the sample current parameters, outdoor parameters, pest and disease parameters and greenhouse control actions at the current sample moment, the sample prediction parameters at the second sample moment are predicted, and the sample prediction parameters include indoor parameters, water and fertilizer parameters, growth parameters, yield parameters and farming activities. Thus, the second sample moment can be used as the new current sample moment, and the above step of selecting the current sample moment as the first sample moment is returned to iterate until the iteration end condition is met. In the above method, the relevant parameters of the current sample moment are predicted by combining the relevant parameters predicted at the previous sample moment and the relevant parameters that need to be introduced at the current sample moment in each iteration process. In this way, the sample growth data can be continuously simulated and generated by cyclic iteration, which helps to improve the convenience of obtaining sample growth data.

[0052] In an implementation scenario, when the sample growth data is generated for the first time, since the sample prediction parameters have not been predicted at the previous sample moment at the current sample moment, the sample prediction parameters can be generated by simulation. Exemplarily, the simulation generation can be initialized by random initialization, selection from a relevant parameter set (such as outdoor parameters, which can be set with a parameter set including "temperature", "humidity", "solar radiation", etc., and parameters can be selected from the parameter set as the initialized outdoor parameters), etc., and the specific method of simulating the generation of parameters is not limited here.

[0053] In an implementation scenario, taking the current sample time t as an example, as the first sample time, the sample prediction parameter obtained by predicting a sample time t-1 at the current sample time t can be selected as the sample current parameter at the current sample time: indoor parameter , water and fertilizer parameters , Growth parameters , Output parameters , Farming Activities , and simulate the outdoor parameters at the current sample time , pest and disease parameters , Greenhouse control action , so that the current sample parameters at the current sample time can be used: indoor parameters , water and fertilizer parameters , Growth parameters , Output parameters , Farming Activities And the outdoor parameters at the current sample time , pest and disease parameters , Greenhouse control action , predict the sample prediction parameters at the second sample time t+1: indoor parameters , water and fertilizer parameters , Growth parameters , Output parameters , Farming Activities The specific process of parameter prediction can refer to the relevant description of the aforementioned growth simulation model, which will not be repeated here. On this basis, the second sample time t+1 can be used as the new current sample time, and the sample prediction parameters (i.e., indoor parameters) obtained by predicting a sample time t at the new current sample time t+1 can be selected. , water and fertilizer parameters , Growth parameters , Output parameters , Farming Activities ), as the sample current parameter of the new current sample time t+1, and simulate the outdoor parameters of the new current sample time , pest and disease parameters , Greenhouse control action , based on the current sample parameters at the new current sample time t+1: indoor parameters , water and fertilizer parameters , Growth parameters , Output parameters , Farming Activities And the new outdoor parameters at the current sample time , pest and disease parameters , Greenhouse control action , predict the sample prediction parameters at the second sample time t+2: indoor parameters , water and fertilizer parameters , Growth parameters , Output parameters , Farming Activities , so that the second sample time t+2 is used as the new current sample time, and new sample growth data can be continuously generated by repeating the cycle. Examples are not given one by one here.

[0054] In an implementation scenario, after obtaining the sample growth data in the above manner, the greenhouse operation decision model can be trained by reinforcement learning training using the sample growth data. As a possible example, reinforcement learning training can be performed using, including but not limited to, AC (Actor-Critic) algorithms. The specific process can refer to the technical details of reinforcement learning algorithms such as AC algorithms, which will not be repeated here. In addition, reinforcement learning training can specifically aim to maximize model rewards, and the model rewards include the difference between the planting income and the planting cost of the target greenhouse crop, and the planting income is obtained based on the yield parameter, and the planting cost is obtained based on the cost consumed by the execution action parameter. Exemplarily, the planting income can be obtained by the product of the yield value in the yield parameter and the unit price of the target greenhouse crop; the planting cost can be obtained by summing up the electricity cost, labor cost, resource cost (water and fertilizer, etc.) consumed by the environmental control action, and the machinery cost (such as machinery rental, depreciation cost of purchasing machinery, etc.), labor cost, etc. consumed by the agricultural activity action.

[0055] In another implementation scenario, different from the above implementation, as another possible implementation, the growth simulation model of the target greenhouse crop can interact with the greenhouse operation decision model in a closed loop, sample growth data is generated during the interaction, and the model reward is calculated based on the sample growth data during each interaction, and the model reward is used to adjust the parameters of the greenhouse operation decision model. Figure 2i , Figure 2i This is a schematic diagram of the interaction between the greenhouse operation decision model and the growth simulation model. Figure 2i As shown in the figure, the greenhouse operation decision model can make decisions based on the relevant parameters of the growth simulation model (such as outdoor parameters, indoor parameters, growth parameters, water and fertilizer parameters, pest and disease parameters, and yield parameters), and obtain environmental control actions, water and fertilizer regulation actions, and agricultural activity actions, which will affect the relevant parameters of the growth simulation model (such as growth parameters, water and fertilizer parameters, pest and disease parameters, and indoor parameters). For example, water and fertilizer regulation actions will affect water and fertilizer parameters, agricultural activity parameters will affect growth parameters and pest and disease parameters, and environmental control parameters will affect indoor parameters. On this basis, the growth simulation model will re-simulate and generate relevant parameters based on the affected relevant parameters (i.e. Figure 2i ), and calculate the model reward (i.e. Figure 2i On the one hand, the greenhouse operation decision model can be adjusted based on the model reward; on the other hand, the greenhouse operation decision model can make decisions based on the relevant parameters newly generated by the growth model, and obtain environmental control actions, water and fertilizer regulation actions, and agricultural activity actions. In this way, sample growth data can be continuously generated in the iterative cycle and the parameters of the greenhouse operation decision model can be adjusted.

[0056] Step S12: Processing the state parameters of the target greenhouse crop at the current moment based on the greenhouse operation decision model to obtain the action parameters to be executed on the target greenhouse crop at the current moment.

[0057] In an implementation scenario, after obtaining a greenhouse operation decision model trained by reinforcement learning, the greenhouse operation decision model can be used for auxiliary decision-making. Specifically, the state parameters of the target greenhouse crop at the current moment (such as outdoor parameters, indoor parameters, water and fertilizer parameters, growth parameters, pest and disease parameters, and yield parameters, etc.) can be obtained first, and then the state parameters at the current moment are input into the greenhouse operation decision model to obtain the action parameters output by the greenhouse operation decision model (such as greenhouse control actions that at least include environmental control actions, water and fertilizer regulation actions, and agricultural activity actions, etc.). It should be noted that after the action parameters are executed, the state parameters will be affected (for details, please refer to the above-mentioned related descriptions, which will not be repeated here). Then, at the next moment, new state parameters can be obtained and input into the greenhouse operation decision model to obtain new action parameters. In this way, the greenhouse operation decision model can continuously provide auxiliary decisions during the planting process of the target greenhouse crops.

[0058] In an implementation scenario, please refer to Figure 3 , Figure 3 Schematic diagram of the process of an embodiment of the greenhouse operation decision-making method of the present application. Figure 3 As shown, firstly, by collecting data from the greenhouse, relevant sample data for training the growth simulation model is obtained, so that the growth simulation model can be forced to perform a growth model for the target greenhouse crop through training. At this point, on the one hand, the growth simulation model can provide relevant sample data for the greenhouse operation decision model, and on the other hand, the greenhouse operation decision model can be provided with relevant sample data by collecting data from the greenhouse, and then the greenhouse operation decision model can be obtained through reinforcement learning training. On this basis, the greenhouse operation decision model can determine the action parameters according to the state parameters at the current moment. After executing the action parameters, the greenhouse environment can be updated (such as indoor parameters, water and fertilizer parameters, etc.), and the greenhouse data can continue to be collected to provide the state parameters for the greenhouse operation decision model to determine the action parameters at the next moment. In this way, a full range of auxiliary decisions can be provided in the planting process of the target greenhouse crop, and the difference between the final benefit and cost can be maximized.

[0059] The above scheme obtains a greenhouse operation decision model obtained by reinforcement learning training based on sample growth data of target greenhouse crops. The sample growth data includes the state parameters of the target greenhouse crops at the second sample moment after the action parameters are adopted under the state parameters of the first sample moment. The state parameters include outdoor parameters, indoor parameters, water and fertilizer parameters, growth parameters, pest and disease parameters and yield parameters. The action parameters include greenhouse control actions and farming activity actions. The state parameters of the target greenhouse crops at the current moment are processed based on the greenhouse operation decision model to obtain the action parameters to be executed on the target greenhouse crops at the current moment. On the one hand, the state parameters at the current moment can be processed by the greenhouse operation decision model trained by reinforcement learning to predict the action parameters to be executed, and it can be real-time. Predict greenhouse operation decisions, and because the action parameters include greenhouse control actions and agricultural activity actions, it can take into account multiple aspects of greenhouse operation decisions. Compared with single decisions such as greenhouse environmental control action decisions, it helps to achieve greenhouse closed-loop decisions. On the other hand, because the greenhouse operation decision model is trained with sample growth data, and the sample growth data includes the state parameters of the target greenhouse crop at the second sample moment after the action parameters are adopted under the state parameters of the first sample moment, reinforcement learning training can force the greenhouse operation decision model to learn how the execution of specific action parameters under specific state parameters affects subsequent state parameters, which helps the greenhouse operation decision model to take into account the yield represented by the yield parameter in the state parameters and the cost represented by the action parameters. Therefore, it is possible to predict greenhouse operation decisions in real time to achieve greenhouse closed-loop decisions and take into account both yield and cost.

[0060] See also Figure 4 , Figure 4 It is a framework diagram of an embodiment of a greenhouse operation decision device of the present application. The greenhouse operation decision device 40 includes: an acquisition module 41 and a decision module 42, the acquisition module 41 is used to obtain a greenhouse operation decision model obtained by reinforcement learning training based on sample growth data of target greenhouse crops; wherein the sample growth data includes the state parameters of the target greenhouse crops at the second sample moment after the action parameters are adopted under the state parameters of the first sample moment, the state parameters include outdoor parameters, indoor parameters, water and fertilizer parameters, growth parameters, pest and disease parameters and yield parameters, and the action parameters include greenhouse control actions and farming activity actions; the decision module 42 is used to process the state parameters of the target greenhouse crops at the current moment based on the greenhouse operation decision model, and obtain the action parameters to be executed on the target greenhouse crops at the current moment.

[0061] In the above scheme, the greenhouse operation decision device 40 obtains a greenhouse operation decision model obtained by reinforcement learning training based on sample growth data of the target greenhouse crop. The sample growth data includes the state parameters of the target greenhouse crop at the second sample moment after the action parameters are adopted under the state parameters of the first sample moment. The state parameters include outdoor parameters, indoor parameters, water and fertilizer parameters, growth parameters, pest and disease parameters and yield parameters. The action parameters include greenhouse control actions and farming activity actions. Then, the state parameters of the target greenhouse crop at the current moment are processed based on the greenhouse operation decision model to obtain the action parameters to be executed on the target greenhouse crop at the current moment. On the one hand, the state parameters at the current moment can be processed by the greenhouse operation decision model trained by reinforcement learning to predict the action parameters to be executed. , it can predict greenhouse operation decisions in real time, and because the action parameters include greenhouse control actions and agricultural activity actions, it can take into account multiple aspects of greenhouse operation decisions. Compared with single decisions such as greenhouse environmental control action decisions, it helps to achieve greenhouse closed-loop decisions. On the other hand, because the greenhouse operation decision model is trained with sample growth data, and the sample growth data includes the state parameters of the target greenhouse crop at the second sample moment after the action parameters are adopted under the state parameters of the first sample moment, the greenhouse operation decision model can be forced to learn how the execution of specific action parameters under specific state parameters affects subsequent state parameters through reinforcement learning training, which helps the greenhouse operation decision model to take into account the yield represented by the yield parameter in the state parameters and the cost represented by the action parameters. Therefore, it is possible to predict greenhouse operation decisions in real time to achieve greenhouse closed-loop decisions and take into account both yield and cost.

[0062] In some disclosed embodiments, the sample growth data is generated by at least a growth simulation model of a target greenhouse crop, and the growth simulation model includes: a greenhouse climate control simulation model, a greenhouse water and fertilizer control simulation model, a crop growth status simulation model, an agricultural activity simulation model, and a crop yield simulation model.

[0063] In some disclosed embodiments, the greenhouse operation decision device 40 includes an indoor simulation module for obtaining indoor parameters, outdoor parameters and environmental control actions in greenhouse control actions at a first sample moment; based on the indoor parameters, outdoor parameters and environmental control actions at the first sample moment, the indoor parameters at the second sample moment are predicted.

[0064] In some disclosed embodiments, the greenhouse operation decision device 40 includes a water and fertilizer simulation module for obtaining water and fertilizer parameters, growth parameters and water and fertilizer regulation actions in greenhouse control actions at a first sample moment; based on the water and fertilizer parameters, growth parameters and water and fertilizer regulation actions at the first sample moment, the water and fertilizer parameters at the second sample moment are predicted.

[0065] In some disclosed embodiments, the greenhouse operation decision device 40 includes a growth simulation module for obtaining indoor parameters, water and fertilizer parameters and growth parameters at a first sample moment; based on the indoor parameters, water and fertilizer parameters and growth parameters at the first sample moment, the growth parameters at the second sample moment are predicted.

[0066] In some disclosed embodiments, the growth simulation module includes an attention submodule for processing indoor parameters, water and fertilizer parameters, and growth parameters at a first sample moment based on a graph attention mechanism to obtain a first feature representation, and processing the indoor parameters, water and fertilizer parameters, and growth parameters at the first sample moment based on a multi-head attention mechanism to obtain a second feature representation; the growth simulation module includes a fusion submodule for fusing the first feature representation and the second feature representation to obtain a fused feature representation; the growth simulation module includes a decoding submodule for performing autoregressive decoding based on the fused feature representation to obtain the growth parameters at the second sample moment.

[0067] In some disclosed embodiments, the greenhouse operation decision device 40 includes an agricultural simulation module, which is used for the action acquisition step including: acquiring the pest and disease parameters and growth parameters at the first sample moment; based on the pest and disease parameters and growth parameters at the first sample moment, predicting the agricultural activity actions at the second sample moment.

[0068] In some disclosed embodiments, the greenhouse operation decision device 40 includes a yield simulation module for obtaining growth parameters and yield parameters at a first sample moment; based on the growth parameters and yield parameters at the first sample moment, the yield parameters at a second sample moment are predicted.

[0069] In some disclosed embodiments, the greenhouse operation decision device 40 includes a selection module for selecting the current sample moment as the first sample moment, and selecting the sample prediction parameters predicted at the previous sample moment at the current sample moment as the sample current parameters at the current sample moment; the greenhouse operation decision device 40 includes a simulation module for simulating and generating outdoor parameters, pest and disease parameters and greenhouse control actions at the current sample moment; the greenhouse operation decision device 40 includes a prediction module for predicting the sample prediction parameters at the second sample moment based on the sample current parameters, outdoor parameters, pest and disease parameters and greenhouse control actions at the current sample moment; wherein the sample prediction parameters include indoor parameters, water and fertilizer parameters, growth parameters, yield parameters and agricultural activity actions; the greenhouse operation decision device 40 includes a loop module for taking the second sample moment as the new current sample moment, and returning to the step of selecting the current sample moment as the first sample moment for iteration until the iteration end condition is met.

[0070] In some disclosed embodiments, reinforcement learning training aims to maximize model rewards, where the model rewards include the difference between the planting income and planting costs of the target greenhouse crop, and the planting income is obtained based on the yield parameters, and the planting cost is obtained based on the cost consumed by executing the action parameters.

[0071] In some disclosed embodiments, outdoor parameters include: at least one of outdoor temperature, outdoor humidity, solar radiation, wind speed and direction, rainfall, total radiation, and photosynthetically active radiation; and / or, indoor parameters include: at least one of indoor temperature, indoor humidity, carbon dioxide concentration, vent status, shade curtain status, heating pipe temperature, carbon dioxide generator supplement value, photosynthetically active radiation, and fill light radiation value; and / or, water and fertilizer parameters include: at least one of irrigation water pH value, irrigation water EC value, irrigation water ion concentration, real-time irrigation status, and irrigation water flow rate; and / or, growth parameters include: at least one of stem diameter, planting density, growth stage, leaf area index, fruit number, and fruit size; and / or, pest and disease parameters include: at least one of disease occurrence degree, disease category, pest occurrence degree, and pest category; and / or, greenhouse control actions include: at least one of environmental control actions and water and fertilizer regulation actions; and / or, agricultural activity actions include: at least one of thinning flowers and fruits, removing old leaves, falling vines, harvesting, and pest and disease management.

[0072] See also Figure 5 , Figure 5 : is a schematic diagram of the framework of an embodiment of an electronic device of the present application. The electronic device 50 includes at least a memory 51 and a processor 52 coupled to each other, the memory 51 at least stores program instructions, and the processor 52 is used to execute the program instructions to implement the steps in any of the above greenhouse operation decision-making method embodiments. For details, please refer to the aforementioned disclosed embodiments, which will not be repeated here. As a possible example, the electronic device 50 may include but is not limited to a server, etc., and the specific type of the electronic device 50 is not limited here.

[0073] Specifically, the processor 52 is used to control itself and the memory 51 to implement the steps in any of the above-mentioned greenhouse operation decision-making method embodiments. The processor 52 can also be called a CPU (Central Processing Unit). The processor 52 may be an integrated circuit chip with signal processing capabilities. The processor 52 can also be a general-purpose processor, a digital signal processor (Digital Signal Processor, DSP), an application-specific integrated circuit (Application Specific Integrated Circuit, ASIC), a field-programmable gate array (Field-Programmable Gate Array, FPGA) or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components. The general-purpose processor can be a microprocessor or the processor can also be any conventional processor, etc. In addition, the processor 52 can be implemented by an integrated circuit chip.

[0074] In the above scheme, the electronic device 50 obtains a greenhouse operation decision model obtained by reinforcement learning training based on sample growth data of the target greenhouse crop. The sample growth data includes the state parameters of the target greenhouse crop at the second sample moment after the action parameters are adopted under the state parameters of the first sample moment. The state parameters include outdoor parameters, indoor parameters, water and fertilizer parameters, growth parameters, pest and disease parameters and yield parameters. The action parameters include greenhouse control actions and farming activity actions. Then, the state parameters of the target greenhouse crop at the current moment are processed based on the greenhouse operation decision model to obtain the action parameters to be executed on the target greenhouse crop at the current moment. On the one hand, the state parameters at the current moment can be processed by the greenhouse operation decision model trained by reinforcement learning to predict the action parameters to be executed, and on the other hand, It can predict greenhouse operation decisions in real time, and because the action parameters include greenhouse control actions and agricultural activity actions, it can take into account multiple aspects of greenhouse operation decisions. Compared with single decisions such as greenhouse environmental control action decisions, it helps to achieve greenhouse closed-loop decisions. On the other hand, because the greenhouse operation decision model is trained with sample growth data, and the sample growth data includes the state parameters of the target greenhouse crop at the second sample moment after the action parameters are adopted under the state parameters of the first sample moment, reinforcement learning training can force the greenhouse operation decision model to learn how the execution of specific action parameters under specific state parameters affects subsequent state parameters, which helps the greenhouse operation decision model to take into account the yield represented by the yield parameter in the state parameters and the cost represented by the action parameters. Therefore, it is possible to predict greenhouse operation decisions in real time to achieve greenhouse closed-loop decisions and take into account both yield and cost.

[0075] See also Figure 6 , Figure 6It is a schematic diagram of a framework of an embodiment of a computer-readable storage medium 60 of the present application. The computer-readable storage medium 60 stores program instructions 61 that can be executed by a processor, and the program instructions 61 are used to implement the steps in any of the above greenhouse operation decision-making method embodiments.

[0076] In the above scheme, the computer-readable storage medium 60 obtains a greenhouse operation decision model obtained by reinforcement learning training based on sample growth data of the target greenhouse crop, the sample growth data includes the state parameters of the target greenhouse crop at the second sample moment after the action parameters are adopted under the state parameters of the first sample moment, the state parameters include outdoor parameters, indoor parameters, water and fertilizer parameters, growth parameters, pest and disease parameters and yield parameters, the action parameters include greenhouse control actions and farming activity actions, and then the state parameters of the target greenhouse crop at the current moment are processed based on the greenhouse operation decision model to obtain the action parameters to be executed on the target greenhouse crop at the current moment. On the one hand, the state parameters at the current moment can be processed by the greenhouse operation decision model trained by reinforcement learning to predict the action parameters to be executed. The greenhouse operation decision model can be predicted in real time. Since the action parameters include greenhouse control actions and agricultural activity actions, it can take into account multiple aspects of greenhouse operation decisions. Compared with single decisions such as greenhouse environmental control action decisions, it helps to achieve greenhouse closed-loop decision-making. On the other hand, since the greenhouse operation decision model is trained with sample growth data, and the sample growth data includes the state parameters of the target greenhouse crop at the second sample moment after the action parameters are adopted under the state parameters of the first sample moment, the greenhouse operation decision model can be forced to learn how the execution of specific action parameters under specific state parameters affects subsequent state parameters through reinforcement learning training, which helps the greenhouse operation decision model to take into account the yield represented by the yield parameter in the state parameters and the cost represented by the action parameters. Therefore, it is possible to predict greenhouse operation decisions in real time to achieve greenhouse closed-loop decision-making and take into account both yield and cost.

[0077] In some embodiments, the functions or modules included in the device provided by the embodiments of the present disclosure can be used to execute the method described in the above method embodiments. The specific implementation can refer to the description of the above method embodiments, and for the sake of brevity, it will not be repeated here.

[0078] The above description of various embodiments tends to emphasize the differences between the various embodiments. The same or similar aspects can be referenced to each other, and for the sake of brevity, they will not be repeated herein.

[0079] In the several embodiments provided in the present application, it should be understood that the disclosed methods and devices can be implemented in other ways. For example, the device implementation described above is only schematic. For example, the division of modules or units is only a logical function division. There may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, and the indirect coupling or communication connection of devices or units can be electrical, mechanical or other forms.

[0080] 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 on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the present embodiment.

[0081] In addition, each functional unit in each embodiment of the present application may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit. The above-mentioned integrated unit may be implemented in the form of hardware or in the form of software functional units.

[0082] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application is essentially or the part that contributes to the prior art or all or part of the technical solution can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions to enable a computer device (which can be a personal computer, server, or network device, etc.) or a processor to execute all or part of the steps of each implementation method of the present application. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk and other media that can store program code.

[0083] If the technical solution of this application involves personal information, the product using the technical solution of this application has clearly informed the personal information processing rules and obtained the individual's voluntary consent before processing the personal information. If the technical solution of this application involves sensitive personal information, the product using the technical solution of this application has obtained the individual's separate consent before processing the sensitive personal information, and at the same time meets the "explicit consent" requirement. For example, on personal information collection devices such as cameras, clear and prominent signs are set to inform that the personal information collection scope has been entered and personal information will be collected. If the individual voluntarily enters the collection scope, it is deemed that he or she agrees to the collection of his or her personal information; or on the device that processes personal information, the personal information processing rules are notified by obvious signs / information, and the individual's authorization is obtained through pop-up information or by asking the individual to upload his or her personal information; among them, the personal information processing rules may include information such as the personal information processor, the purpose of personal information processing, the processing method, and the type of personal information processed.

Claims

1. A greenhouse operation decision-making method, characterized in that: include: A greenhouse operation decision model is obtained by reinforcement learning training based on sample growth data of target greenhouse crops; wherein the sample growth data includes state parameters of the target greenhouse crops at a second sample time after adopting action parameters under state parameters at a first sample time, the state parameters include outdoor parameters, indoor parameters, water and fertilizer parameters, growth parameters, pest and disease parameters, and yield parameters, and the action parameters include greenhouse control actions and farming activity actions; The state parameters of the target greenhouse crop at the current moment are processed based on the greenhouse operation decision model to obtain the action parameters to be executed on the target greenhouse crop at the current moment.

2. The method according to claim 1, characterized in that The sample growth data is generated by at least a growth simulation model of a target greenhouse crop, and the growth simulation model includes: a greenhouse climate control simulation model, a greenhouse water and fertilizer control simulation model, a crop growth state simulation model, a farming activity simulation model and a crop yield simulation model.

3. The method according to claim 1, characterized in that The step of acquiring indoor parameters at the second sample time comprises: Acquire indoor parameters, outdoor parameters, and environmental control actions in greenhouse control actions at the first sample time; Based on the indoor parameters, outdoor parameters and environmental control actions at the first sample moment, the indoor parameters at the second sample moment are predicted.

4. The method according to claim 1, characterized in that The step of obtaining the water and fertilizer parameters at the second sample moment includes: Acquire water and fertilizer parameters, growth parameters, and water and fertilizer regulation actions in greenhouse control actions at the first sample time; Based on the water and fertilizer parameters, growth parameters and water and fertilizer regulation actions at the first sample moment, the water and fertilizer parameters at the second sample moment are predicted.

5. The method according to claim 1, characterized in that The step of acquiring the growth parameters at the second sample time comprises: Acquiring indoor parameters, water and fertilizer parameters, and growth parameters at the first sample time; Based on the indoor parameters, water and fertilizer parameters and growth parameters at the first sample time, the growth parameters at the second sample time are predicted.

6. The method according to claim 5, characterized in that The method of predicting the growth parameters at the second sample time based on the indoor parameters, water and fertilizer parameters, and growth parameters at the first sample time includes: Processing the indoor parameters, water and fertilizer parameters, and growth parameters at the first sample moment based on the graph attention mechanism to obtain a first feature representation, and processing the indoor parameters, water and fertilizer parameters, and growth parameters at the first sample moment based on the multi-head attention mechanism to obtain a second feature representation; Fusing the first feature representation and the second feature representation to obtain a fused feature representation; Autoregressive decoding is performed based on the fused feature representation to obtain the growth parameter at the second sample moment.

7. The method according to claim 1, characterized in that The step of acquiring the farming activity action at the second sample moment comprises: Obtaining pest and disease parameters and growth parameters at the first sample moment; Based on the pest and disease parameters and growth parameters at the first sample moment, the farming activities at the second sample moment are predicted.

8. The method according to claim 1, characterized in that The step of obtaining the production parameter at the second sample moment includes: Obtaining growth parameters and yield parameters at the first sample moment; Based on the growth parameters and yield parameters at the first sample moment, the yield parameters at the second sample moment are predicted.

9. The method according to any one of claims 1 to 8, characterized in that: The step of acquiring the sample growth data comprises: Selecting the current sample moment as the first sample moment, selecting the sample prediction parameter predicted at a sample moment before the current sample moment as the sample current parameter at the current sample moment, and simulating and generating the outdoor parameter, pest and disease parameter, and greenhouse control action at the current sample moment; Based on the sample current parameters, outdoor parameters, pest and disease parameters and greenhouse control actions at the current sample moment, the sample prediction parameters at the second sample moment are predicted; wherein the sample prediction parameters include the indoor parameters, the water and fertilizer parameters, the growth parameters, the yield parameters and the farming activity actions; The second sample moment is used as a new current sample moment, and the step of selecting the current sample moment as the first sample moment is returned to iterate until an iteration end condition is met.

10. The method according to any one of claims 1 to 8, characterized in that: The reinforcement learning training aims to maximize the model reward, and the model reward includes the difference between the planting income and the planting cost of the target greenhouse crop, and the planting income is obtained based on the yield parameter, and the planting cost is obtained based on the cost consumed in executing the action parameter.

11. The method according to any one of claims 1 to 8, characterized in that: The outdoor parameters include: at least one of outdoor temperature, outdoor humidity, solar radiation, wind speed and direction, rainfall, total radiation, and photosynthetically active radiation; And / or, the indoor parameters include: at least one of indoor temperature, indoor humidity, carbon dioxide concentration, vent status, shade curtain status, heating pipe temperature, carbon dioxide generator supplement value, photosynthetically active radiation, and supplementary light radiation value; And / or, the water and fertilizer parameters include: at least one of: irrigation water pH value, irrigation water EC value, irrigation water ion concentration, real-time irrigation status, and irrigation water flow rate; And / or, the growth parameters include: at least one of stem diameter, planting density, growth stage, leaf area index, fruit number, and fruit size; And / or, the pest and disease parameters include: at least one of the degree of disease occurrence, the type of disease, the degree of pest occurrence, and the type of pest; And / or, the greenhouse control action includes: at least one of an environment control action and a water and fertilizer regulation action; And / or, the farming activities include: at least one of thinning flowers and fruits, removing old leaves, removing vines, harvesting, and pest and disease management.

12. A greenhouse operation decision-making device, characterized in that: include: An acquisition module, used to acquire a greenhouse operation decision model obtained through reinforcement learning training based on sample growth data of target greenhouse crops; wherein the sample growth data includes state parameters of the target greenhouse crops at a second sample time after adopting action parameters under state parameters at a first sample time, the state parameters include outdoor parameters, indoor parameters, water and fertilizer parameters, growth parameters, pest and disease parameters, and yield parameters, and the action parameters include greenhouse control actions and farming activity actions; The decision module is used to process the state parameters of the target greenhouse crop at the current moment based on the greenhouse operation decision model to obtain the action parameters to be executed on the target greenhouse crop at the current moment.

13. An electronic device, characterized in that: It at least comprises a memory and a processor coupled to each other, wherein the memory at least stores program instructions, and the processor is used to execute the program instructions to implement the greenhouse operation decision method according to any one of claims 1 to 11.

14. A computer-readable storage medium, characterized in that: Program instructions that can be executed by a processor are stored, and the program instructions are used to implement the greenhouse operation decision-making method described in any one of claims 1 to 11.

Images