Greenhouse production model construction method and device based on mechanism model mapping

Abstract

The invention provides a greenhouse production model construction method and device based on mechanism model mapping. The method comprises the following steps: acquiring initial crop state data, initial greenhouse environment state data, a greenhouse control input sequence and an outdoor environment state sequence of a greenhouse; inputting the initial crop state data, the initial greenhouse environment state data, a greenhouse control input sequence and an outdoor environment state sequence into a greenhouse crop production system mechanism model to obtain a crop state sequence and a greenhouse environment state sequence, large-scale acquisition of crop growth data and greenhouse environment data in the greenhouse is realized in a low-cost mode; a data set is constructed based on the crop state sequence, the greenhouse environment state sequence, the greenhouse control input sequence and the outdoor environment state sequence to train a neural network model, and a greenhouse crop production system data model for predicting the crop state and the greenhouse environment state at the next moment is obtained. Accurate construction of the greenhouse crop production system data model is realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of agricultural environment control, and particularly relates to a greenhouse production model construction method and device based on mechanism model mapping. BACKGROUND

[0002] The greenhouse production model constructed in the prior art includes a mechanism model and a data model driven by pure data. The mechanism model is relatively mature for glass greenhouses, lettuce and tomatoes, and the modeling of sunlight greenhouse environment has also developed to a certain extent. If economic benefit maximization is taken as a control target, the greenhouse-crop interaction model needs to be calibrated with production data. The data model has relatively more research on greenhouse environment prediction but relatively less research on crop growth modeling, and most of the data models do not consider the influence of the actuator operation on the greenhouse environment. When a reinforcement learning training environment is constructed with economic benefit maximization as a control target, crop data and control data need to be established in the data model, and the time sequence characteristics of the process data also need to be considered.

[0003] The mechanism model in the prior art relies on accurate parameters, but the actual production parameters are time-varying, which leads to the failure of the modeling method under extreme working conditions. The data model driven by pure data can capture nonlinear characteristics, but it is prone to overfitting due to the scarcity of crop growth data. At the same time, the long production cycle and low fault tolerance of the greenhouse result in high cost of data acquisition. How to acquire crop growth data and greenhouse environment data in the greenhouse at a low cost and on a large scale to accurately construct the greenhouse production model is an important issue to be solved in the industry. SUMMARY

[0004] The present application provides a greenhouse production model construction method and device based on mechanism model mapping, which realizes the accurate construction of the greenhouse production model by acquiring crop growth data and greenhouse environment data in the greenhouse at a low cost.

[0005] The present application provides a greenhouse production model construction method based on mechanism model mapping, which includes the following steps: Obtain initial crop state data, initial greenhouse environment state data, greenhouse control input sequences and outdoor environment state sequences in a preset time period in the greenhouse, wherein the greenhouse control input sequences and the outdoor environment state sequences are randomly generated based on upper and lower limit environmental constraints in the greenhouse; Input the initial crop state data, the initial greenhouse environment state data, the greenhouse control input sequences and the outdoor environment state sequences into a greenhouse crop production system mechanism model to obtain a crop state sequence and a greenhouse environment state sequence output by the greenhouse crop production system mechanism model in the preset time period, wherein the greenhouse crop production system mechanism model is used to represent a mathematical relationship model between the crop growth state, the greenhouse environment, the actuator operation and the external weather; inputting the crop state sequence, the greenhouse environment state sequence, the greenhouse control input sequence and the outdoor environment state sequence as an input of a training sample of a dataset, and inputting the crop state data and the greenhouse environment state data of the next time point of the preset time period as a label of the training sample; training a neural network model based on the dataset to obtain a greenhouse crop production system data model.

[0006] According to the greenhouse production model construction method based on mechanism model mapping provided by the application, the neural network model comprises a Transformer network and a bidirectional long short-term memory network. The Transformer network comprises an input layer, a position encoding layer and a Transformer encoder connected in sequence, and the bidirectional long short-term memory network comprises a time embedding layer, a bidirectional LSTM layer, an attention layer, a flattening layer and an output layer connected in sequence. The Transformer encoder of the Transformer network is connected to the time embedding layer of the bidirectional long short-term memory network According to the greenhouse production model construction method based on mechanism model mapping provided by the application, based on a training sample in the dataset, the neural network model is trained, comprising: inputting the training sample into the input layer of the Transformer network to obtain a word embedding sequence output by the input layer of the Transformer network; the position encoding layer of the Transformer network performs position encoding on the word embedding sequence to obtain a position encoding sequence; and the Transformer encoder performs feature extraction on the position encoding sequence to output an enhanced feature sequence containing global long-term dependency. inputting the enhanced feature sequence into the time embedding layer of the bidirectional long short-term memory network to obtain a time-enhanced embedding feature output by the time embedding layer; the bidirectional LSTM layer in the bidirectional long short-term memory network performs bidirectional LSTM processing on the embedding feature to obtain a bidirectional fused hidden state sequence; the attention layer in the bidirectional long short-term memory network performs attention weight calculation on the bidirectional fused hidden state sequence to obtain an attention weighted context vector; the flattening layer in the bidirectional long short-term memory network flattens the context vector into two-dimensional data to obtain a two-dimensional feature matrix; and the output layer in the bidirectional long short-term memory network performs nonlinear transformation through a full connection layer to obtain an output result. adjusting parameters of the neural network model based on the output result and the label of the training sample.

[0007] The application provides a greenhouse production model construction method based on a mechanism model mapping. A teacher-student dual model framework is constructed, a teacher model in the teacher-student dual model framework is the greenhouse crop production system data model, a student model in the teacher-student dual model framework is a light-weight greenhouse crop production system data model obtained by knowledge migration from the teacher model through a knowledge migration mechanism, and the teacher model and the student model both comprise a feature extraction layer, a time series modeling layer and a prediction layer. In the feature extraction layer, the output of the Transformer encoder of the student model is forced to align with the multi-head attention matrix and the relative position encoding feature distribution of the teacher model through a distillation mechanism, an MSE loss function is used as a loss function of the feature layer distillation, the difference between the output of the student model and the output of the teacher model in the feature layer is calculated, and the student model is guided to adjust parameters based on the difference; In the time series modeling layer, a bidirectional hidden state distillation mechanism is introduced, the forward and backward hidden states of the bidirectional LSTM layer of the teacher model are dynamically weighted and matched with the hidden states of the corresponding layer of the student model, a cosine similarity is used as an evaluation index, the difference between the student model and the teacher model in capturing time series dependence is evaluated, and the student model is guided to optimize the time series modeling process through a distillation loss function; In the prediction layer, a temperature scaling technology is used to soften the output distribution of the full connection layer of the teacher model, a soft target is generated, and the student model is guided to learn the feature mapping process at the decision boundary of the teacher model.

[0008] The application provides a greenhouse production model construction method based on a mechanism model mapping. The greenhouse control input sequence comprises CO2 supplement amount data, heating amount data and ventilation amount data. The greenhouse environment state sequence comprises indoor CO2 concentration data, indoor temperature data and indoor humidity data. The crop state sequence comprises crop dry weight data.

[0009] The application provides a greenhouse production model construction method based on a mechanism model mapping. All training samples in the data set are subjected to normalization processing; The data set subjected to the normalization processing is divided into a training set and a test set; The neural network model is trained based on the training set, and a trained neural network model is obtained. Test the trained neural network model based on the test set, and obtain the greenhouse crop production system data model in a case where it is determined that the prediction accuracy of the trained neural network model is greater than a preset accuracy threshold.

[0010] The application also provides a greenhouse production model construction device based on mechanism model mapping, comprising the following modules. A data acquisition module is configured to acquire initial crop state data, initial greenhouse environment state data, a greenhouse control input sequence, and an outdoor environment state sequence within a preset time period in a greenhouse, wherein the greenhouse control input sequence and the outdoor environment state sequence are randomly generated based on upper and lower limit environmental constraints in the greenhouse. A mechanism mapping module is configured to input the initial crop state data, the initial greenhouse environment state data, the greenhouse control input sequence, and the outdoor environment state sequence into a greenhouse crop production system mechanism model to obtain a crop state sequence and a greenhouse environment state sequence output by the greenhouse crop production system mechanism model within the preset time period, wherein the greenhouse crop production system mechanism model is used to represent a mathematical relationship model between a crop growth state, a greenhouse environment, an actuator operation, and external weather. A data set construction module is configured to input the crop state sequence, the greenhouse environment state sequence, the greenhouse control input sequence, and the outdoor environment state sequence as an input of a training sample of a data set, and input crop state data and greenhouse environment state data at a next time point of the preset time period as a label of the training sample. A model construction module is configured to train a neural network model based on the data set to obtain a greenhouse crop production system data model.

[0011] The application also provides an electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, wherein the processor implements the above-mentioned greenhouse production model construction method based on mechanism model mapping when executing the program.

[0012] The application also provides a non-transitory computer-readable storage medium having a computer program stored thereon, wherein the computer program is executed by a processor to implement the above-mentioned greenhouse production model construction method based on mechanism model mapping.

[0013] The application also provides a computer program product comprising a computer program, wherein the computer program is executed by a processor to implement the above-mentioned greenhouse production model construction method based on mechanism model mapping.

[0014] The greenhouse production model construction method and device based on mechanism model mapping provided by the application realize large-scale acquisition of crop growth data and greenhouse environment data in a low-cost manner by inputting initial crop state data, initial greenhouse environment state data, a greenhouse control input sequence and an outdoor environment state sequence into a greenhouse crop production system mechanism model to obtain a crop state sequence and a greenhouse environment state sequence in a preset time period. BRIEF DESCRIPTION OF DRAWINGS

[0015] In order to more clearly illustrate the technical solutions in the application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments or the prior art description. Obviously, the drawings in the following description are some embodiments of the application, and other drawings can be obtained by those skilled in the art without creative labor.

[0016] Figure 1 is a flowchart of the greenhouse production model construction method based on mechanism model mapping provided by the application.

[0017] Figure 2 is a schematic diagram of the greenhouse crop interaction mechanism provided by the application.

[0018] Figure 3 is a structural schematic diagram of the neural network model provided by the application.

[0019] Figure 4 is a structural schematic diagram of the greenhouse production model construction device based on mechanism model mapping provided by the application.

[0020] Figure 5 is a structural schematic diagram of the electronic device provided by the application. DETAILED DESCRIPTION

[0021] In order to make the objects, technical solutions and advantages of the application clearer, the technical solutions in the application will be described clearly and completely below with reference to the drawings in the application. Obviously, the described embodiments are some embodiments of the application, not all embodiments. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor are within the protection scope of the application.

[0022] Figure 1 is a flowchart of the greenhouse production model construction method based on mechanism model mapping provided by the application, as shown in Figure 1 The method comprises the following steps. In step 110, initial crop state data, initial greenhouse environment state data, a greenhouse control input sequence and an outdoor environment state sequence in a preset time period in the greenhouse are acquired, the greenhouse control input sequence and the outdoor environment state sequence being randomly generated based on upper and lower limit environment constraints in the greenhouse. In step 120, the initial crop state data, the initial greenhouse environment state data, the greenhouse control input sequence and the outdoor environment state sequence are input into a greenhouse crop production system mechanism model to obtain a crop state sequence and a greenhouse environment state sequence output by the greenhouse crop production system mechanism model in the preset time period, the greenhouse crop production system mechanism model being used to represent a mathematical relationship model between a crop growth state, a greenhouse environment, an actuator operation and external weather. In step 130, the crop state sequence, the greenhouse environment state sequence, the greenhouse control input sequence and the outdoor environment state sequence are input as an input of a training sample of a data set, and crop state data and greenhouse environment state data at a next time point of the preset time period are taken as labels of the training sample. In step 140, a neural network model is trained based on the data set to obtain a greenhouse crop production system data model.

[0023] The technical scheme of the present application will be described in detail below with the computer executing the greenhouse production model construction method based on the mechanism model mapping provided by the present application as an example.

[0024] In step 110, initial crop state data, initial greenhouse environment state data, a greenhouse control input sequence and an outdoor environment state sequence in a preset time period in the greenhouse are acquired, the greenhouse control input sequence and the outdoor environment state sequence being randomly generated based on upper and lower limit environment constraints in the greenhouse.

[0025] Firstly, the time range, frequency and specific environment parameters required for data collection need to be determined. The preset time period can be a time period in which the crop is in the greenhouse growth environment. To + For example, the time range is generally several minutes or several hours, for example, 30 minutes.

[0026] Optionally, the outdoor environment state data can include solar radiation intensity, CO2 concentration, temperature and humidity, the greenhouse control input data can include CO2 supplement amount, heating amount and ventilation amount, the greenhouse environment state data can include indoor CO2 concentration, temperature and humidity, and the crop state sequence includes crop dry weight data.

[0027] In step 120, the initial crop state data, initial greenhouse environment state data, greenhouse control input sequence and outdoor environment state sequence are input into a greenhouse crop production system mechanism model to obtain a crop state sequence and a greenhouse environment state sequence output by the greenhouse crop production system mechanism model within a preset time period, and the greenhouse crop production system mechanism model is used to represent a mathematical relationship model between a crop growth state, a greenhouse environment, an actuator operation and external weather.

[0028] A greenhouse crop production system mechanism model is constructed in advance, and the greenhouse crop production system mechanism model is used to represent a mathematical relationship model between a crop growth state, a greenhouse environment, an actuator operation and external weather. A construction process diagram of a greenhouse crop interaction mechanism can be as shown in Figure 2 The greenhouse crop interaction mechanism provided by the present application is shown in the schematic diagram. System dynamics analysis is performed on the greenhouse crop production system. The effects of outdoor environment state data (solar radiation intensity data , outdoor CO2 concentration data , outdoor temperature data , outdoor humidity data and greenhouse control input data (CO2 supplement amount data , heating amount data , ventilation amount data on greenhouse environment state data (indoor CO2 concentration data , indoor temperature data , indoor humidity data and crop state data (crop dry weight data ) are considered comprehensively, and the greenhouse crop production system mechanism model is constructed as shown in formulas (1)-(4): (1) (2) (3) (4) wherein, is a yield factor, is a canopy photosynthesis rate, is a dry matter respiration rate, is a CO2 exchange amount through ventilation, is a heat capacity of greenhouse gas, is an energy exchange amount through ventilation and penetration of the covering, is a solar radiation heat load, is a volume of greenhouse gas relative humidity, is a canopy transpiration rate, is a water vapor exchange amount through ventilation.

[0029] After the mechanism model of the greenhouse crop production system is constructed, a large number of greenhouse control input sequences and outdoor environment state sequences for subsequent model training processes are randomly generated within the upper and lower environmental constraints in the greenhouse. After the data for subsequent model training processes are generated based on the mechanism model, actual data can be further mixed to achieve mixed enhancement of the data.

[0030] It should be noted that, in terms of data models, there are relatively more studies on greenhouse environment prediction, but relatively fewer studies on crop growth modeling, and the implementation process of related methods does not consider the influence of actuator operation on the greenhouse environment. When constructing a reinforcement learning training environment with the control target of maximizing economic benefits, crop state data and greenhouse control inputs need to be established in the data model, and the time sequence characteristics of process data need to be considered.

[0031] The constructed mechanism model of the greenhouse crop production system relies on accurate parameters, but the actual production parameters are strongly time-varying, which leads to the failure of the modeling method under extreme working conditions; pure data-driven models can capture nonlinear characteristics, but are prone to overfitting due to the scarcity of crop growth data. Therefore, the amount of parameter data in the actual production process is small. Based on the constructed mechanism model of the greenhouse crop production system, a large amount of data can be generated to fill the sample gap based on the generated virtual data.

[0032] The mechanism model of the greenhouse crop production system is the basis for the mechanical explanation of control decisions, and it can also be used as a training environment for the model. However, considering the powerful nonlinear fitting ability, adaptive ability, and lightweight simulation calculation advantages of data models, when the actual production data is small, a large amount of data can be generated by the mechanism model of the greenhouse crop production system for feature mapping by the data model.

[0033] In step 130, the crop state sequence, the greenhouse environment state sequence, the greenhouse control input sequence, and the outdoor environment state sequence are used as the input of a training sample of a data set, and the crop state data and the greenhouse environment state data of the next time of the preset time period are used as the label of the training sample.

[0034] A plurality of training samples are constructed based on the obtained data, and a data set for the neural network model training process is constructed based on the plurality of training samples.

[0035] In step 140, the neural network model is trained based on the data set to obtain a data model of the greenhouse crop production system.

[0036] Specifically, after constructing the data set, the data in the data set is normalized, and is divided into a training set, a test set and a validation set according to 8:1:1. The deep learning hyperparameters are optimized, and a greenhouse crop production system data model mapped by a mechanism model is established. Considering the time delay problem in the greenhouse control system, a deep neural network data model is constructed by using time series data. In the training process, the initial greenhouse environment state data and the initial crop state data generated at the moment are taken as the initial state, the greenhouse control input sequence and the outdoor environment state sequence at the moment are input into the neural network model, and the crop state data and the greenhouse environment state data at the moment are simulated and solved. + +1 to + +1 moment. The neural network model is constructed with the crop state sequence, the greenhouse environment state sequence, the greenhouse control input sequence and the outdoor environment state sequence as input, and the crop state data and the greenhouse environment state data at the moment as output. to + + +1 moment.

[0037] It can be understood that the crop state sequence and the greenhouse environment state sequence are generated based on the greenhouse crop production system mechanism model, which can represent the mathematical relationship between the crop growth state, the greenhouse environment, the actuator operation and the external weather. This method makes the generated data set more consistent with the actual crop growth rule in the case of insufficient actual data, and improves the authenticity and reliability of the data.

[0038] The greenhouse production model construction method and device based on the mechanism model mapping provided by the application realize large-scale acquisition of crop growth data and greenhouse environment data in the greenhouse in a low-cost manner by inputting the initial crop state data, the initial greenhouse environment state data, the greenhouse control input sequence and the outdoor environment state sequence into the greenhouse crop production system mechanism model to obtain the crop state sequence and the greenhouse environment state sequence in a preset time period. The data set constructed based on the data acquired in a large scale based on the greenhouse crop production system mechanism model is used to train the neural network model to obtain the greenhouse crop production system data model, and the accurate construction of the greenhouse crop production system data model is realized.

[0039] ​​​​In one embodiment, the neural network model comprises a Transformer network and a bidirectional long short-term memory network; the Transformer network comprises an input layer, a position encoding layer and a Transformer encoder connected in sequence, and the bidirectional long short-term memory network comprises a time embedding layer, a bidirectional LSTM layer, an attention layer, a flattening layer and an output layer connected in sequence; the Transformer encoder of the Transformer network is connected to the time embedding layer of the bidirectional long short-term memory network.

[0040] The constructed neural network model can be as follows Figure 3 The neural network model provided by the application is shown in the structural diagram. The neural network model can be a Transformer-BiLSTM hybrid model, which is a hybrid deep learning model combining a Transformer network and a bidirectional long short-term memory network (BiLSTM), and is designed to integrate the global feature extraction capability of the Transformer and the time sequence dependence modeling advantage of the BiLSTM, and is suitable for processing long-range dependence and local feature interaction in time sequence data.

[0041] The model input layer is time sequence data with a size of 1×11, and the output layer is a greenhouse-crop state estimation neuron with a size of 1×4. The Transformer-BiLSTM model is used for time sequence feature mapping. A suitable sampling interval is selected through calculation experiments from aspects of training difficulty and prediction accuracy , and a greenhouse crop production system data model is constructed.

[0042] In one embodiment, based on one training sample in the data set, the neural network model is trained, including: inputting the training sample into the input layer of the Transformer network, obtaining the word embedding sequence output by the input layer of the Transformer network; the position encoding layer of the Transformer network performs position encoding on the word embedding sequence to obtain a position encoding sequence; the Transformer encoder extracts features from the position encoding sequence and outputs an enhanced feature sequence containing global long-term dependencies; input the enhanced feature sequence into the time embedding layer of the bidirectional long short-term memory network, and obtain the time-enhanced embedding feature output by the time embedding layer; the bidirectional LSTM layer in the bidirectional long short-term memory network performs bidirectional LSTM processing on the embedding feature to obtain a bidirectional fused hidden state sequence; the attention layer in the bidirectional long short-term memory network performs attention weight calculation on the bidirectional fused hidden state sequence to obtain an attention weighted context vector; the flattening layer in the bidirectional long short-term memory network flattens the context vector into two-dimensional data to obtain a two-dimensional feature matrix; the output layer in the bidirectional long short-term memory network performs nonlinear transformation through a fully connected layer to obtain an output result; based on the output result and the label of the training sample, the parameters of the neural network model are adjusted.

[0043] First, the neural network model receives input training samples and uses learnable position encoding to preliminarily model time information. Then, an improved Transformer network is used as the main feature extractor, which contains a multi-head attention mechanism to capture global long-term dependencies in the input sequence through self-attention. In addition, relative position encoding and hybrid sparse attention mechanism are introduced to enhance the ability to capture local time features and reduce computational complexity when processing long sequences.

[0044] Each Transformer network also includes a feedforward neural network, layer normalization, and residual connections to improve the model's expressive power and training stability. The output of the Transformer network is passed through a time embedding layer, which encodes information from different sampling intervals. The enhanced feature sequence is then input into a BiLSTM network to further explore the bidirectional sequential dependencies between local time sequences through forward and backward LSTM units. The output of the BiLSTM network is passed through an attention layer to dynamically adjust the importance of different time sequence features to generate a context vector. Finally, the features are flattened through a Flatten layer and mapped to the final output result through a fully connected layer. In terms of model optimization, various regularization techniques are used, including layer normalization, dropout, and weight decay, to prevent overfitting and enhance generalization ability.

[0045] During the model training process, the AdamW optimizer, cosine annealing learning rate adjustment strategy with warm restart, and early stopping strategy can be used to further improve the training efficiency and convergence speed. In addition, the data augmentation technique of adding random noise to the input features can improve the robustness of the model in a noisy environment.

[0046] Based on the obtained output results and the labels of the training samples, the parameters of the neural network model are adjusted to realize the training process of the model.

[0047] The sample processing flow implemented based on each layer in the model includes: inputting the training sample into the input layer of the Transformer network. The main task of the input layer is to perform word embedding processing on the input data. Word embedding is a technique that maps words or subwords to a low-dimensional vector space, which can capture the semantic and grammatical relationships between words. Through word embedding, the input layer converts each word or subword into a fixed-dimensional vector, and these vectors are combined to form a word embedding sequence.

[0048] After obtaining the word embedding sequence, the position encoding layer of the Transformer network begins to play a role. The position encoding layer uses a specific mathematical function to generate a position vector for each position in the word embedding sequence. This position vector is added to the corresponding word vector to obtain an encoded vector containing position information. All position encoded vectors are combined to form a position encoding sequence.

[0049] The position encoding sequence is input into the Transformer encoder for feature extraction. The Transformer encoder is composed of multiple self-attention layers and feedforward neural network layers.

[0050] The enhanced feature sequence output by the Transformer network is input into the time embedding layer of the BiLSTM network. The main role of the time embedding layer is to further process the enhanced feature sequence to better adapt to the time series processing requirements of the BiLSTM network.

[0051] The bidirectional LSTM layer in the BiLSTM network performs bidirectional LSTM processing on the time-enhanced embedded features. LSTM (Long Short-Term Memory Network) is a recurrent neural network specifically designed for processing time series data, which solves the gradient vanishing and gradient explosion problems in traditional RNN by introducing a gating mechanism.

[0052] The bidirectional LSTM layer is composed of two directional LSTM units, one processing sequence data from front to back, and the other processing sequence data from back to front. In this way, the hidden state of each time step contains not only the information before the current time step, but also the information after the current time step, so as to more comprehensively capture the time series dependencies in the sequence.

[0053] After the bidirectional LSTM processing, a bidirectional fused hidden state sequence is obtained. Each hidden state vector in this sequence contains information of the current time step and its previous and subsequent time steps, which is a deep representation of the time series data.

[0054] The attention layer in the BiLSTM network performs attention weight calculation on the bidirectional fused hidden state sequence. The attention layer calculates the similarity between each hidden state vector and other hidden state vectors to obtain the attention weight of each hidden state vector. Then, the hidden state vectors are weighted and summed according to these attention weights to obtain an attention-weighted context vector. After obtaining the attention-weighted context vector, the flattening layer in the BiLSTM network flattens it into two-dimensional data to obtain a two-dimensional feature matrix.

[0055] Finally, the output layer in the BiLSTM network performs nonlinear transformation on the two-dimensional feature matrix through a fully connected layer. The fully connected layer multiplies each element in the two-dimensional feature matrix with a set of weights and adds a bias term, and then performs nonlinear transformation through an activation function (such as Sigmoid, Softmax, etc.) to obtain the final output result.

[0056] In one embodiment, after obtaining the greenhouse crop production system data model, further comprising: constructing a teacher-student dual model framework, the teacher model in the teacher-student dual model framework is the greenhouse crop production system data model, the student model in the teacher-student dual model framework is a lightweight greenhouse crop production system data model obtained by knowledge transfer mechanism from the teacher model, and the teacher model and the student model both contain feature extraction layer, time series modeling layer and prediction layer; in the feature extraction layer, the output of the Transformer encoder of the student model is forced to align with the multi-head attention matrix and the relative position encoding feature distribution of the teacher model through the distillation mechanism, and the mean square error loss function is used as the loss function of the feature layer distillation to calculate the difference between the student model and the teacher model in the feature layer output, and guide the student model to adjust the parameters; in the time series modeling layer, the forward and backward hidden states of the teacher model bidirectional LSTM layer are dynamically weighted and matched with the hidden states of the corresponding layer of the student model through the bidirectional hidden state distillation mechanism, and the cosine similarity is used as the evaluation index to evaluate the difference between the student model and the teacher model in capturing time series dependence, and the distillation loss function is used to guide the student model to optimize the time series modeling process; in the prediction layer, the temperature scaling technology is used to soften the output distribution of the teacher model fully connected layer to generate a soft target, and guide the student model to learn the feature mapping process at the decision boundary of the teacher model.

[0057] Specifically, a teacher-student dual model framework is constructed: a teacher model and a student model are designed, wherein the teacher model is a greenhouse crop production system data model, and the student model is a lightweight target model; the teacher model and the student model each comprise a feature extraction layer, a time series modeling layer, and a prediction layer.

[0058] Feature layer distillation aligns spatio-temporal feature representation: in the feature extraction layer, the output of the Transformer encoder of the student model is forced to align with the multi-head attention matrix and the relative position encoding feature distribution of the teacher model through a distillation mechanism; a mean square error loss function is used as the loss function of feature layer distillation, the difference between the outputs of the student model and the teacher model in the feature layer is calculated, and the student model is guided to adjust the parameters.

[0059] Bidirectional hidden state distillation in the time series modeling layer: in the time series modeling layer, a bidirectional hidden state distillation mechanism is introduced to dynamically weight and match the forward and backward hidden states of the BiLSTM layer of the teacher model with the hidden states of the corresponding layer of the student model; cosine similarity is used as an evaluation index to evaluate the difference between the student model and the teacher model in capturing time series dependency, and a distillation loss function is used to guide the student model to optimize the time series modeling capability.

[0060] Soft target distillation in the prediction layer: in the prediction layer, the output distribution of the teacher model full connection layer is softened by temperature scaling technology to generate a soft target; the student model is guided to learn the fine feature mapping at the decision boundary of the teacher model, and the cross-entropy loss function or the KL divergence loss function can be used to calculate the loss of soft target distillation, and the student model parameters are updated.

[0061] Further, a hierarchical adaptive weight adjustment mechanism can be designed to dynamically balance the contribution proportion of the feature layer, the time series layer, and the prediction layer distillation loss in the training stage; through a weight allocation strategy, different training stages focus on optimizing the distillation target of different levels.

[0062] In the training process, a progressive distillation strategy is adopted, which focuses on low-level feature alignment in the early stage, increases the weight of feature layer distillation loss or reduces the weight of other layer loss, and guides the student model to learn basic features first; gradually increase the high-level semantic supervision in the later stage, and adjust the weight proportion to make the student model closer to the teacher model in high-level semantics.

[0063] The student model inherits the regularization configuration and data enhancement strategy of the teacher model, including but not limited to Dropout, weight decay, data perturbation, etc.; through inheritance and enhancement strategy, the generalization ability and robustness of the student model are improved.

[0064] Through the iterative training process, the student model parameters are constantly updated, so that the student model gradually approaches the teacher model in feature extraction, time series modeling, and prediction ability, and finally obtains a lightweight model with optimized performance.

[0065] In one embodiment, the outdoor environment state sequence includes outdoor solar radiation intensity data, outdoor CO2 concentration data, outdoor temperature data, and outdoor humidity data; the greenhouse control input sequence includes CO2 supplement amount data, heating amount data, and ventilation amount data; the greenhouse environment state sequence includes indoor CO2 concentration data, indoor temperature data, and indoor humidity data; and the crop state sequence includes crop dry weight data.

[0066] It can be understood that the outdoor environment state sequence, the greenhouse control input sequence, the greenhouse environment state sequence, and the crop state sequence can realize a multi-dimensional data coverage process in the greenhouse crop production system data model construction. These data not only comprehensively reflect the environmental conditions for crop growth, but also accurately capture the key factors affecting crop growth. By integrating these data, the accuracy and reliability of the greenhouse crop production system data model construction can be improved.

[0067] In one embodiment, the neural network model is trained based on the data set to obtain a greenhouse crop production system data model, including: normalizing all training samples in the data set; dividing the normalized data set into a training set and a test set; training the neural network model based on the training set to obtain a trained neural network model; testing the trained neural network model based on the test set, and obtaining the greenhouse crop production system data model in a case where the prediction accuracy of the trained neural network model is greater than a preset accuracy threshold.

[0068] The training process of the neural network model is a systematic and step-by-step process, aiming to obtain an accurate prediction model through data processing, model training and testing.

[0069] The normalization of data is to eliminate the dimensional differences between different features, so that the model training is more stable and the convergence is accelerated. All training samples in the data set are normalized to obtain a normalized data set.

[0070] The data set is divided into a training set and a test set for model training and performance evaluation. Specifically, it can be usually divided according to the proportion of 80% training set and 20% test set, or adjusted according to the data volume and task requirements.

[0071] The neural network model is trained based on the training set, the performance of the trained neural network model on the test set is evaluated to ensure its generalization ability. The trained neural network model is tested based on the test set to determine the prediction accuracy. The prediction accuracy is compared with the preset accuracy threshold. If the prediction accuracy is greater than the preset accuracy threshold, it is considered that the model training is successful, and the pre-trained model is determined as the greenhouse crop production system data model. If the prediction accuracy does not reach the threshold, the model structure, hyperparameters or retraining need to be adjusted.

[0072] The mechanism model mapping-based greenhouse production model construction device provided by the present application is described below, and the mechanism model mapping-based greenhouse production model construction device described below can be correspondingly referred to the mechanism model mapping-based greenhouse production model construction method described above.

[0073] As shown in Figure 4 The device comprises: A data acquisition module 410 is configured to acquire initial crop state data, initial greenhouse environment state data, a greenhouse control input sequence and an outdoor environment state sequence in a predetermined time period in a greenhouse, wherein the greenhouse control input sequence and the outdoor environment state sequence are randomly generated based on upper and lower limit environmental constraints in the greenhouse. A mechanism mapping module 420 is configured to input the initial crop state data, the initial greenhouse environment state data, the greenhouse control input sequence and the outdoor environment state sequence into a greenhouse crop production system mechanism model to obtain a crop state sequence and a greenhouse environment state sequence output by the greenhouse crop production system mechanism model in the predetermined time period, wherein the greenhouse crop production system mechanism model is used to represent a mathematical relationship model between a crop growth state, a greenhouse environment, an actuator operation and external weather. A data set construction module 430 is configured to input the crop state sequence, the greenhouse environment state sequence, the greenhouse control input sequence and the outdoor environment state sequence as an input of a training sample of a data set, and input crop state data and greenhouse environment state data at a next time point of the predetermined time period as a label of the training sample. A model construction module 440 is configured to train a neural network model based on the data set to obtain a greenhouse crop production system data model.

[0074] The greenhouse production model construction device based on a mechanism model mapping provided by the application realizes large-scale acquisition of crop growth data and greenhouse environment data in a low-cost manner by inputting initial crop state data, initial greenhouse environment state data, a greenhouse control input sequence and an outdoor environment state sequence into a greenhouse crop production system mechanism model to obtain a crop state sequence and a greenhouse environment state sequence in a preset time period.

[0075] In one embodiment, the model construction module 440 is specifically configured to: determine that the neural network model comprises a Transformer network and a bidirectional long short-term memory network; the Transformer network comprises an input layer, a position encoding layer and a Transformer encoder connected in sequence, and the bidirectional long short-term memory network comprises a time embedding layer, a bidirectional LSTM layer, an attention layer, a flattening layer and an output layer connected in sequence; the Transformer encoder of the Transformer network is connected to the time embedding layer of the bidirectional long short-term memory network.

[0076] In one embodiment, the data set construction module 430 is specifically configured to: train the neural network model based on one training sample in the data set, comprising: input the training sample into the input layer of the Transformer network to obtain a word embedding sequence output by the input layer of the Transformer network; the position encoding layer of the Transformer network performs position encoding on the word embedding sequence to obtain a position encoding sequence; and the Transformer encoder performs feature extraction on the position encoding sequence to output an enhanced feature sequence containing global long-term dependency relationships; input the enhanced feature sequence into the time embedding layer of the bidirectional long short-term memory network to obtain a time-enhanced embedding feature output by the time embedding layer; the bidirectional LSTM layer in the bidirectional long short-term memory network performs bidirectional LSTM processing on the embedding feature to obtain a bidirectional fused hidden state sequence; the attention layer in the bidirectional long short-term memory network performs attention weight calculation on the bidirectional fused hidden state sequence to obtain an attention weighted context vector; the flattening layer in the bidirectional long short-term memory network flattens the context vector into a two-dimensional data to obtain a two-dimensional feature matrix; and the output layer in the bidirectional long short-term memory network performs nonlinear transformation through a fully connected layer to obtain an output result; Based on the output result and the label of the training sample, the parameters of the neural network model are adjusted.

[0077] In one embodiment, the model construction module 440 is specifically configured to: After obtaining the greenhouse crop production system data model, further comprising: A teacher-student dual model framework is constructed, the teacher model in the teacher-student dual model framework is the greenhouse crop production system data model, the student model in the teacher-student dual model framework is a lightweight greenhouse crop production system data model obtained by knowledge transfer mechanism from the teacher model, and the teacher model and the student model both contain a feature extraction layer, a time series modeling layer and a prediction layer; In the feature extraction layer, the output of the Transformer encoder of the student model is forced to align with the multi-head attention matrix and the relative position encoding feature distribution of the teacher model through the distillation mechanism, and the mean square error loss function is used as the loss function of the feature layer distillation to calculate the difference between the output of the student model and the teacher model in the feature layer, and guide the student model to adjust the parameters; In the time series modeling layer, a bidirectional hidden state distillation mechanism is introduced to dynamically weight and match the forward and backward hidden states of the bidirectional LSTM layer of the teacher model with the hidden states of the corresponding layer of the student model, and the cosine similarity is used as a measure index to evaluate the difference between the student model and the teacher model in capturing time series dependence, and the distillation loss function is used to guide the student model to optimize the time series modeling process; In the prediction layer, the temperature scaling technology is used to soften the output distribution of the full connection layer of the teacher model to generate a soft target and guide the student model to learn the feature mapping process at the decision boundary of the teacher model.

[0078] In one embodiment, the data acquisition module 410 is specifically configured to: The outdoor environment state sequence includes outdoor solar radiation intensity data, outdoor CO2 concentration data, outdoor temperature data and outdoor humidity data; The greenhouse control input sequence includes CO2 supplement amount data, heating amount data and ventilation amount data; The greenhouse environment state sequence includes indoor CO2 concentration data, indoor temperature data and indoor humidity data; The crop state sequence includes crop dry weight data.

[0079] In one embodiment, the model construction module 440 is further specifically configured to: Based on the data set, a neural network model is trained to obtain a greenhouse crop production system data model, including: All training samples in the data set are normalized. Divide the normalized dataset into training set and test set; Based on the training set, the neural network model is trained to obtain a trained neural network model; The trained neural network model is tested based on the test set, and when it is determined that the prediction accuracy of the trained neural network model is greater than a preset accuracy threshold, the greenhouse crop production system data model is obtained.

[0080] Figure 5 An example of a physical structure diagram of an electronic device is shown below. Figure 5 As shown, the electronic device may include: a processor 510, a communications interface 520, a memory 530, and a communications bus 540, wherein the processor 510, the communications interface 520, and the memory 530 communicate with each other via the communications bus 540. The processor 510 may call logic instructions in the memory 530 to execute a greenhouse production model construction method based on mechanism model mapping, the method comprising: obtaining initial crop state data, initial greenhouse environment state data, a greenhouse control input sequence, and an outdoor environment state sequence within a preset time period in the greenhouse, wherein the greenhouse control input sequence and the outdoor environment state sequence are randomly generated based on upper and lower environmental constraints in the greenhouse; Inputting the initial crop state data, initial greenhouse environment state data, greenhouse control input sequence, and outdoor environment state sequence into a greenhouse crop production system mechanism model to obtain a crop state sequence and a greenhouse environment state sequence output by the greenhouse crop production system mechanism model within a preset time period, wherein the greenhouse crop production system mechanism model is used to represent a mathematical relationship model among crop growth state, greenhouse environment, actuator operation, and external weather; Using the crop state sequence, the greenhouse environment state sequence, the greenhouse control input sequence, and the outdoor environment state sequence as inputs of a training sample of a data set, and using the crop state data and the greenhouse environment state data at the next moment of the preset time period as labels of the training sample; A neural network model is trained based on the data set to obtain a greenhouse crop production system data model.

[0081] Further, the logic instructions in the memory 530 described above can be implemented in the form of software functional units and sold or used as independent products, and can be stored in a computer readable storage medium. Based on such understanding, the technical solutions of the present application essentially or the parts that make contributions to the prior art or parts of the technical solutions can be embodied in the form of a software product. The computer software product is stored in a storage medium, and includes several instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present application. The aforementioned storage medium includes: a U disk, a mobile hard disk, a read-only memory (ROM, Read-Only Memory), a random access memory (RAM, Random Access Memory), a magnetic disk or an optical disk, and various media that can store program codes.

[0082] In another aspect, the present application also provides a computer program product, which comprises a computer program, the computer program can be stored on a non-transitory computer readable storage medium, and the computer program can be executed by a processor to enable a computer to execute the mechanism model mapping based greenhouse production model construction method provided by the above-mentioned methods. The method comprises: obtaining initial crop state data, initial greenhouse environment state data, a greenhouse control input sequence and an outdoor environment state sequence in a predetermined time period in a greenhouse, wherein the greenhouse control input sequence and the outdoor environment state sequence are randomly generated based on upper and lower limit environmental constraints in the greenhouse; inputting the initial crop state data, the initial greenhouse environment state data, the greenhouse control input sequence and the outdoor environment state sequence into a greenhouse crop production system mechanism model to obtain a crop state sequence and a greenhouse environment state sequence output by the greenhouse crop production system mechanism model in the predetermined time period, wherein the greenhouse crop production system mechanism model is used to represent a mathematical relationship model between a crop growth state, a greenhouse environment, an actuator operation and external weather; inputting the crop state sequence, the greenhouse environment state sequence, the greenhouse control input sequence and the outdoor environment state sequence as an input of a training sample of a data set, and inputting crop state data and greenhouse environment state data at the next time of the predetermined time period as labels of the training sample; training a neural network model based on the data set to obtain a greenhouse crop production system data model.

[0083] In yet another aspect, the present application also provides a non-transitory computer readable storage medium having stored thereon a computer program, which, when executed by a processor, implements a mechanism model mapping based greenhouse production model construction method provided by the above method, the method comprising: obtaining initial crop state data, initial greenhouse environment state data, a greenhouse control input sequence and an outdoor environment state sequence in a predetermined time period in a greenhouse, the greenhouse control input sequence and the outdoor environment state sequence being randomly generated based on upper and lower limit environmental constraints in the greenhouse; inputting the initial crop state data, the initial greenhouse environment state data, the greenhouse control input sequence and the outdoor environment state sequence into a greenhouse crop production system mechanism model to obtain a crop state sequence and a greenhouse environment state sequence output by the greenhouse crop production system mechanism model in the predetermined time period, the greenhouse crop production system mechanism model being used to represent a mathematical relationship model between a crop growth state, a greenhouse environment, an actuator operation and external weather; inputting the crop state sequence, the greenhouse environment state sequence, the greenhouse control input sequence and the outdoor environment state sequence as an input of a training sample of a data set, and inputting crop state data and greenhouse environment state data at a next time point of the predetermined time period as labels of the training sample; training a neural network model based on the data set to obtain a greenhouse crop production system data model.

[0084] The device embodiments described above are merely illustrative, wherein the units illustrated as separate components can or can not be physically separated, and the components illustrated as units can or can not be physical units, i.e., can be located in one place or distributed on multiple network units. Part or all of the modules can be selected to achieve the purpose of the embodiment scheme according to actual needs. Those skilled in the art can understand and implement without creative labor.

[0085] From the above description of the embodiments, those skilled in the art can clearly understand that the embodiments can be realized by means of software plus necessary general hardware platforms, and of course can also be realized by hardware. Based on such understanding, the above technical solutions can be embodied in the form of a software product, which can be stored in a computer readable storage medium, such as a ROM / RAM, a magnetic disk, an optical disk, etc., and includes a plurality of instructions to make a computer device (which can be a personal computer, a server, or a network device, etc.) execute the methods described in each embodiment or some parts of the embodiments.

[0086] It should be pointed out finally that the above embodiments are only used to illustrate the technical solutions of the present application, but not to limit the same; and although the present application has been described in detail with reference to the foregoing embodiments, it should be appreciated by those skilled in the art that the technical solutions recorded in the foregoing embodiments can be modified, or some technical features thereof can be replaced equivalently; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A greenhouse production model construction method based on mechanism model mapping, characterized in that: include: Acquiring initial crop state data, initial greenhouse environmental state data, a greenhouse control input sequence, and an outdoor environmental state sequence within a preset time period in the greenhouse, wherein the greenhouse control input sequence and the outdoor environmental state sequence are randomly generated based on upper and lower environmental constraints in the greenhouse; Inputting the initial crop state data, initial greenhouse environment state data, greenhouse control input sequence, and outdoor environment state sequence into a greenhouse crop production system mechanism model to obtain a crop state sequence and a greenhouse environment state sequence output by the greenhouse crop production system mechanism model within a preset time period, wherein the greenhouse crop production system mechanism model is used to represent a mathematical relationship model among crop growth state, greenhouse environment, actuator operation, and external weather; Using the crop state sequence, the greenhouse environment state sequence, the greenhouse control input sequence, and the outdoor environment state sequence as inputs of a training sample of a data set, and using the crop state data and the greenhouse environment state data at the next moment of the preset time period as labels of the training sample; A neural network model is trained based on the data set to obtain a greenhouse crop production system data model.

2. The greenhouse production model construction method based on mechanism model mapping according to claim 1 is characterized in that: The neural network model includes a Transformer network and a bidirectional long short-term memory network; The Transformer network includes an input layer, a position encoding layer, and a Transformer encoder connected in sequence, and the bidirectional long short-term memory network includes a time embedding layer, a bidirectional LSTM layer, an attention layer, a flattening layer, and an output layer connected in sequence; The Transformer encoder of the Transformer network is connected to the time embedding layer of the bidirectional long short-term memory network.

3. The greenhouse production model construction method based on mechanism model mapping according to claim 2 is characterized in that: Training a neural network model based on a training sample in the data set includes: Inputting the training sample into the input layer of the Transformer network to obtain a word embedding sequence output by the input layer of the Transformer network; the position encoding layer of the Transformer network position-encodes the word embedding sequence to obtain a position encoding sequence; the Transformer encoder extracts features from the position encoding sequence and outputs an enhanced feature sequence containing global long-term dependencies; The enhanced feature sequence is input into the time embedding layer of the bidirectional long short-term memory network to obtain the time-series enhanced embedding features output by the time embedding layer; the bidirectional LSTM layer in the bidirectional long short-term memory network performs bidirectional LSTM processing on the embedding features to obtain a bidirectional fused hidden state sequence; the attention layer in the bidirectional long short-term memory network performs attention weight calculation on the bidirectional fused hidden state sequence to obtain an attention-weighted context vector; the flattening layer in the bidirectional long short-term memory network flattens the context vector into two-dimensional data to obtain a two-dimensional feature matrix; the output layer in the bidirectional long short-term memory network performs nonlinear transformation through a fully connected layer to obtain an output result; Based on the output results and the labels of the training samples, the parameters of the neural network model are adjusted.

4. The greenhouse production model construction method based on mechanism model mapping according to claim 2 is characterized in that: After obtaining the greenhouse crop production system data model, the method further includes: Constructing a teacher-student dual-model framework, wherein the teacher model in the teacher-student dual-model framework is the greenhouse crop production system data model, and the student model in the teacher-student dual-model framework is a lightweight greenhouse crop production system data model obtained by migrating knowledge from the teacher model through a knowledge transfer mechanism, and both the teacher model and the student model include a feature extraction layer, a time series modeling layer, and a prediction layer; At the feature extraction layer, a distillation mechanism is used to force the Transformer encoder output of the student model to align with the multi-head attention matrix and relative position encoding feature distribution of the teacher model. The mean squared error loss function is used as the loss function for feature layer distillation. The difference between the feature layer outputs of the student model and the teacher model is calculated to guide the adjustment of the student model parameters. In the time series modeling layer, a bidirectional hidden state distillation mechanism is introduced to dynamically weighted match the forward and backward hidden states of the bidirectional LSTM layer of the teacher model with the hidden states of the corresponding layers of the student model. Cosine similarity is used as a measurement metric to evaluate the difference between the student model and the teacher model in capturing time series dependencies. The distillation loss function is used to guide the student model to optimize the time series modeling process. In the prediction layer, the output distribution of the fully connected layer of the teacher model is softened by the temperature scaling technique to generate soft targets and guide the student model to learn the feature mapping process at the decision boundary of the teacher model.

5. The greenhouse production model construction method based on mechanism model mapping according to claim 1 is characterized in that: The outdoor environment state sequence includes outdoor solar radiation intensity data, outdoor CO2 concentration data, outdoor temperature data and outdoor humidity data; The greenhouse control input sequence includes CO2 replenishment data, heating data and ventilation data; The greenhouse environment state sequence includes indoor CO2 concentration data, indoor temperature data and indoor humidity data; The crop status sequence includes crop dry weight data.

6. The greenhouse production model construction method based on mechanism model mapping according to claim 1 is characterized in that: The neural network model is trained based on the data set to obtain a greenhouse crop production system data model, including: Normalizing all training samples in the data set; Divide the normalized dataset into training set and test set; Based on the training set, the neural network model is trained to obtain a trained neural network model; The trained neural network model is tested based on the test set, and when it is determined that the prediction accuracy of the trained neural network model is greater than a preset accuracy threshold, the greenhouse crop production system data model is obtained.

7. A greenhouse production model construction device based on mechanism model mapping, characterized in that: include: a data acquisition module for acquiring initial crop state data, initial greenhouse environmental state data, a greenhouse control input sequence, and an outdoor environmental state sequence within a preset time period in the greenhouse, wherein the greenhouse control input sequence and the outdoor environmental state sequence are randomly generated based on upper and lower environmental constraints in the greenhouse; a mechanism mapping module for inputting the initial crop state data, initial greenhouse environment state data, greenhouse control input sequence, and outdoor environment state sequence into a greenhouse crop production system mechanism model to obtain a crop state sequence and a greenhouse environment state sequence output by the greenhouse crop production system mechanism model within a preset time period, wherein the greenhouse crop production system mechanism model is used to represent a mathematical relationship model among crop growth status, greenhouse environment, actuator operation, and external weather; a data set construction module, configured to use the crop state sequence, the greenhouse environment state sequence, the greenhouse control input sequence, and the outdoor environment state sequence as inputs of a training sample of a data set, and use the crop state data and greenhouse environment state data at the next moment of the preset time period as labels of the training sample; The model building module is used to train the neural network model based on the data set to obtain a greenhouse crop production system data model.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that: When the processor executes the computer program, the method for constructing a greenhouse production model based on mechanism model mapping as described in any one of claims 1 to 6 is implemented.

9. 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 method for constructing a greenhouse production model based on mechanism model mapping as described in any one of claims 1 to 6 is implemented.

10. A computer program product comprising a computer program, characterized in that When the computer program is executed by a processor, the method for constructing a greenhouse production model based on mechanism model mapping as described in any one of claims 1 to 6 is implemented.

Images