Greenhouse crop production environment monitoring control method based on multiple angles and multiple strategies

Abstract

The invention belongs to the technical field of agricultural intellectualization, and provides a multi-angle and multi-strategy-based greenhouse crop production environment monitoring control method, which comprises the following steps: collecting environment and soil crop data to construct a matrix and pre-processing the matrix; extracting time sequence characteristics by adopting Res-BiLSTM, and constructing a yield prediction model in combination with the improved XGBoost; integrating multi-dimensional features to construct a feature matrix, completing standardization, and determining an optimal cluster number and a cluster affiliation label by adopting an improved K-means clustering algorithm; for each cluster, nonlinear features are extracted through improved SVR, and a temperature and humidity prediction model is constructed by combining with a GBDT model fitting residual error; the key parameters of the model are optimized by adopting Chebyshev chaotic mapping for the CEO algorithm; and a layered cooperative regulation and control mechanism of MPC global optimization and feedforward PID real-time correction is adopted. According to the control method, multi-source data acquisition, precise prediction model construction and a layered cooperative regulation mechanism are integrated, precise monitoring and dynamic regulation of the greenhouse environment are realized, and the crop yield and quality stability are improved.

Description

Technical Field

[0001] This invention relates to the interdisciplinary fields of agricultural facility intelligence, intelligent control, and machine learning, and specifically to a method for monitoring and controlling the greenhouse crop production environment based on multiple angles and strategies. Background Technology

[0002] Greenhouses have become a key element in improving yield and resource utilization efficiency in facility agriculture. Their production process requires precise and dynamic control of environmental parameters such as temperature, humidity, light, CO2, and soil water and fertilizer. However, traditional management models relying on manual experience suffer from low precision, slow response, and resource waste, making it difficult to meet the differentiated needs of crop growth stages and resulting in unstable yields and quality.

[0003] While the Internet of Things (IoT) and intelligent algorithms have facilitated the digital management of greenhouse environments, existing systems still have shortcomings: the integration of multi-source environmental, soil, and crop data is not high; prediction models are mostly based on single algorithms, making it difficult to account for the nonlinearity and dynamics of complex environments; and control strategies still mainly rely on fixed thresholds, making it difficult to handle multi-device coupling and external disturbances, and failing to fully consider crop physiological characteristics. Therefore, constructing an intelligent greenhouse system that integrates multi-dimensional sensing, high-precision prediction, and hierarchical collaborative control is of great significance for improving the stability and sustainability of facility agriculture. Summary of the Invention

[0004] In view of this, embodiments of the present invention provide a multi-angle, multi-strategy-based method for monitoring and controlling the production environment of greenhouse crops, in order to solve the problems of insufficient multi-source data fusion, single prediction model, and rigid control strategy in existing intelligent control methods.

[0005] This invention provides a multi-angle, multi-strategy method for monitoring and controlling the greenhouse crop production environment, comprising the following steps: S1: Data Acquisition and Preprocessing: Set up monitoring points to collect environmental parameters (temperature, humidity, light intensity, CO2 concentration) and construct an environmental parameter matrix; measure soil crop information (soil physicochemical characteristics, crop agronomic characteristics and growth indicators) and establish a soil crop information matrix; preprocess the environmental parameter matrix and the soil crop information matrix. S2: Constructing a yield prediction model: The environmental parameter matrix is ​​sliced ​​according to crop growth stages, and the temporally relevant features (plant height changes and leaf number within stages) in the soil and crop information matrix are integrated to construct an LSTM input. Res-BiLSTM is used to extract temporal features, and an improved XGBoost is combined to construct a Res-BiLSTM-IXGBoost model. The predicted values ​​are spliced ​​together and output as the final predicted yield through a fully connected nested layer. S3: Construct a temperature and humidity prediction model: Integrate multi-dimensional features to construct a feature matrix and complete standardization. Use an improved K-means clustering algorithm (introducing temperature and humidity weights) to determine the optimal number of clusters and cluster affiliation labels. For each cluster, extract nonlinear features by improving SVR (introducing residual gradient and attention weights), and combine the residuals with the GBDT model to construct a Ti-AnnSVR-GBDT model; to achieve temperature and humidity prediction. S4: Parameter Optimization: The CEO used the Chebyshev chaotic mapping algorithm to optimize key parameters (number of hidden layer neurons) of the Ti-AnnSVR-GBDT model using the chaotic evolution optimization algorithm. Learning rate Penalty coefficient ); S5: Layered collaborative regulation: Combine the temperature and humidity stress thresholds defined by the growth stage to construct an explicit risk function and train an integrated model to predict the comprehensive risk; the upper-level MPC sets soft constraints and multi-objective functions to continuously optimize the control sequence; the lower-level PID introduces feedforward terms to correct the deviation, decouples the device control quantity through the allocation matrix, and uses a linear superposition structure to fuse the control quantity to obtain the STriskMPC-feedforward-PID controller, which collaboratively outputs the final control quantity.

[0006] Technical effects: By integrating multi-source data acquisition, precise prediction model construction and hierarchical collaborative control mechanism, this technology solves problems such as insufficient multi-source data fusion, single prediction model and rigid control strategy in existing technologies, and achieves precise monitoring and dynamic control of greenhouse environment, thereby improving crop yield and quality stability.

[0007] According to a specific implementation of an embodiment of the present invention, the deployment of monitoring points in step S1 includes: setting up m monitoring points inside the greenhouse, with each point equipped with sensors for temperature, humidity, light intensity, and CO2 concentration to continuously collect environmental parameters; soil and crop information includes soil physicochemical characteristics (type, nitrogen, phosphorus, and potassium content), static agronomic characteristics of crops (variety, planting density, and growth stage), and dynamic growth indicators (plant height and number of leaves). Technical effect: Multi-dimensional collection of environmental and crop data provides comprehensive input for subsequent models, avoiding the limitations of a single data source.

[0008] According to a specific implementation of an embodiment of the present invention, the preprocessing of the environmental parameter matrix and the soil crop information matrix in step S1 includes: mean interpolation to fill missing values, criterion-based outlier removal, and Z-score standardization to unify dimensions. Technical effect: Eliminates data noise and dimensional differences, providing reliable input for subsequent models.

[0009] According to a specific implementation of an embodiment of the present invention, in step S2, after extracting temporal features using Res-BiLSTM, attention weights need to be calculated using MLP to generate a temporal feature vector, which is then element-wise multiplied with the static features. Finally, an improved XGBoost model is constructed to build the Res-BiLSTM-IXGBoost model, as follows: ; In the formula, It is a linear transformation operation; It is a feature concatenation operation; This is a preliminary production forecast; It is a temporal feature vector; It is the weight matrix of the nested layer; It is the bias term of the nested layer.

[0010] Technical benefits: The model structure effectively integrates the complex interaction between dynamic time series and static attributes, providing more accurate feature support for greenhouse crop yield prediction and environmental control.

[0011] According to a specific implementation of an embodiment of the present invention, in step S3, the improved K-means clustering algorithm introduces a temperature and humidity weight term to improve the loss function, and determines the optimal number of clusters K through K-means++ initialization and the elbow rule; the loss function is expressed as follows: ; In the formula, It is the first Feature vectors of each sample; It is the first The feature center vectors of each cluster; It is the first The sample to the first The square of the Euclidean distance between the centers of each cluster; It is the first The measured temperature value of each sample; It is the first The average temperature of all samples within a cluster; No. The relative humidity value of each sample; It is the first The mean humidity of all samples within a cluster; and These are separate weighted loss terms for temperature and humidity, respectively.

[0012] Technical effect: Introducing temperature and humidity weight terms improves the clustering loss function, enhances the accuracy of clustering results, and provides a reasonable basis for accurate intra-cluster modeling of temperature and humidity prediction models.

[0013] According to a specific implementation of an embodiment of the present invention, in step S3, the loss function of the improved SVR is improved by introducing attention weights; GBDT iteratively generates a decision tree to fit the prediction residuals of the improved SVR; the constructed Ti-AnnSVR-GBDT model is as follows: ; In the formula, , It is the first The Lagrange multiplier corresponding to each sample; It is the first Kernel function for each feature; It is the first of the samples to be predicted One eigenvalue; It is a cluster Inner The first sample One eigenvalue; It is the first The SVR model bias term corresponding to each cluster; It is the residual compensation value; It is a cluster Inner The actual temperature and humidity values ​​of each sample.

[0014] Technical effect: By improving SVR to extract nonlinear features and combining GBDT to fit residuals, the nonlinear correlation between temperature and humidity and environmental factors is accurately captured, significantly improving the accuracy of temperature and humidity prediction.

[0015] According to a specific implementation of an embodiment of the present invention, in step S4, the key parameter includes the number of hidden layer neurons. Learning rate Penalty coefficient The Chebyshev chaotic mapping optimization includes: generating a chaotic sequence matrix through Chebyshev chaotic mapping and linearly mapping it to the parameter interval to obtain an initial matrix; iteratively updating the parameter matrix through chaotic perturbation and elite guidance mechanisms to output the globally optimal parameter combination. Technical effects: Optimizes key model parameters, enhancing the generalization ability and adaptability to complex scenarios of the Ti-AnnSVR-GBDT model.

[0016] According to a specific implementation of an embodiment of the present invention, step S5 specifically comprises: S5.1: Define the temperature stress threshold based on the physiological characteristics of the growth stage. Humidity stress threshold Construct an explicit risk function based on the threshold: ; In the formula, It is the high / low temperature stress weighting coefficient; It is the high / low humidity stress weighting coefficient; It is the ReLU function; / It is the high / low temperature stress threshold corresponding to growth stage s; This refers to the high / low humidity stress threshold corresponding to growth stage s; A learning enhancement module is introduced to train an ensemble model using historical data on environment, soil, and growth status. Predicting the future Step-by-step comprehensive risks: ; In the formula, It is the Ti-AnnSVR-GBDT model; It is an identifier for the type of risk in the model; It is an index for future time steps; These are predictions from the Ti-AnnSVR-GBDT model; This refers to the rate of change of temperature and humidity; S5.2: Employs MPC as the upper-level controller, using linearized state equations to describe temperature and humidity dynamics, balancing system complexity with real-time control performance. ; In the formula, Temperature and humidity status; This is the output of the device; External disturbance; A is the state transition matrix; B is the control input matrix; Define the comfort zone based on the stress threshold. , Introducing slack variables , , , Transform into soft constraints: ; Design a multi-objective MPC objective function: (18); In the formula, These are the weighting coefficients of the constraint relaxation term; These are the weighting coefficients for explicit risk items; It is the weighting coefficient of the implicit risk item; It is the future The predicted state vector at time step; It is the future Predicted temperature and humidity levels inside the greenhouse at any given time; This is an ideal reference value; It is the future The amount of control at any given moment; Rolling optimization, with state equations, soft constraints, and control range as conditions, solves for the future. Step-optimal control sequence; S5.3: Define temperature error: ; In the formula, This is the MPC benchmark value; These are measured values; The integral term uses an anti-saturation strategy to update and feedback the control quantity. ; In the formula, It is the saturation function of the error; These are measured values; Incorporating feedforward terms based on Ti-AnnVR-GBDT predictions: ; S5.4: Upper-layer MPC per Optimize once per PID step, outputting a baseline control quantity. The lower-level PID each Corrects every second, responding to real-time disturbances; By allocating the matrix , Decoupling control quantity: ; in, , ; It refers to the number of devices; The global optimization characteristics of the upper-level model predictive control (MPC) and the local disturbance suppression capability of the lower-level PID controller are combined using a linear superposition structure to fuse the control inputs. .

[0017] Technical effects: It enables the coordination of upper-level MPC global optimization and lower-level PID real-time response, solves the problem of multi-device coupling, improves the accuracy of environmental control and anti-interference ability, and ensures the stability of the crop growth environment. Attached Figure Description

[0018] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the accompanying drawings used in the description of the specific embodiments or the prior art will be briefly introduced below.

[0019] Figure 1 A flowchart of a multi-angle, multi-strategy greenhouse crop production environment monitoring and control method according to an embodiment of the present invention is shown; Figure 2A flowchart of data acquisition in an embodiment of the present invention is shown; Figure 3 A flowchart of product prediction in an embodiment of the present invention is shown; Figure 4 A flowchart of temperature and humidity prediction in an embodiment of the present invention is shown; Figure 5 A flowchart of parameter optimization in an embodiment of the present invention is shown; Figure 6 A flowchart of the joint controller in an embodiment of the present invention is shown; Figure 7 A flowchart illustrating the controller fusion improvement in an embodiment of the present invention is shown. Detailed Implementation

[0020] The present invention will be further illustrated below with reference to specific embodiments. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. After reading the present invention, any modifications of the present invention in various equivalent forms by those skilled in the art will fall within the scope defined by the appended claims.

[0021] It should be noted that, unless otherwise stated, the technical or scientific terms used in this application should have the ordinary meaning as understood by those skilled in the art to which this invention pertains.

[0022] Figure 1 A flowchart illustrating the steps of a multi-angle, multi-strategy greenhouse crop production environment monitoring and control method provided in this embodiment of the invention is shown below. Figure 1 The method includes the following steps: S1: Data Acquisition and Preprocessing: Environmental parameters (temperature, humidity, light intensity, CO2 concentration) are collected by setting up monitoring points to construct an environmental parameter matrix; soil crop information (soil physicochemical characteristics, crop agronomic characteristics, and growth indicators) is measured to establish a soil crop information matrix; the environmental parameter matrix and soil crop information matrix are preprocessed. Specifically, such as... Figure 2 As shown: S1.1: A multi-region, multi-sensor time-series sampling scheme is adopted, with m monitoring points deployed inside the facility greenhouse. Each point is equipped with a temperature and humidity sensor, a light sensor, and a CO2 concentration sensor. Data is continuously collected at a fixed time step for a duration of [duration missing]. Environmental data, total time steps The environmental parameters include four categories: temperature (T), humidity (H), light intensity (L), and CO2 concentration (C). A time-series environmental parameter matrix was ultimately constructed. This method enables the collection of environmental factors across multiple regions in a spatiotemporal and multidimensional manner. Through multi-sample, multi-modal information collection, p crop planting samples were selected. Soil physicochemical characteristics, including soil type, nitrogen, phosphorus, and potassium content, were measured using a soil analyzer. Combining manual recording with instrumental testing, static agronomic characteristics such as crop variety, planting density, and growth stage, as well as dynamic growth indicators such as plant height and leaf number, were obtained, totaling nine categories of characteristics. A soil-crop information matrix was then constructed. This enables cross-modal fusion acquisition of multi-attribute information of soil and crops.

[0023] S1.2: First, missing values ​​in matrices E and F are filled using time series mean interpolation and the mean of samples of the same class, respectively; secondly, through... The criteria identify and remove outliers; finally, Z-score standardization is performed on all features to achieve dimensional unification and distribution optimization of multi-source data.

[0024] S2: Constructing a yield prediction model: The environmental parameter matrix is ​​sliced ​​according to crop growth stages, and time-series-dependent features (plant height changes and leaf number within stages) from the soil and crop information matrices are integrated to construct an LSTM input. Res-BiLSTM is used to extract time-series features, and attention weights are calculated using MLP to generate a time-series feature vector. This vector is then element-wise multiplied with static features. An improved XGBoost model is then constructed using Res-BiLSTM-IXGBoost. After optimization, the input model is optimized, and the concatenated predicted values ​​are passed through a fully connected nested layer to output the final predicted yield. Specifically, as shown... Figure 3 As shown: S2.1: Slice the environmental parameter matrix E according to crop growth stages, and simultaneously fuse the time-series-specific features (plant height changes and leaf number within stages) in the soil and crop information matrix F with the environmental features in E to construct the LSTM input format. .in, p represents the number of samples (i.e., the number of samples p in the soil and crop information matrix F). The time step for a single growth stage. This represents the number of time-series features after fusion.

[0025] S2.2: Output the temporal feature matrix through a bidirectional long short-term memory network (Res-BiLSTM) with integrated residual connections: (1); In the formula, It is the input tensor of the LSTM model, which is constructed by slicing the environmental parameter matrix according to the growth stage and fusing it with the temporal features of soil and crop. These are the weight matrix and bias vector of the Res-BiLSTM network; Then, time-step attention weights are calculated using a multilayer perceptron (MLP): (2); In the formula, These are the weight matrix and bias vector of the attention layer; It is a bidirectional LSTM in the first... The hidden state at each time step; It is in a hidden state A multilayer perceptron as input; Finally, the time-series feature vector is output through weighted aggregation. To achieve multi-scale fusion of key temporal information of environment and soil-crop; S2.3: Extracting static features from the soil and crop information matrix F (M represents the number of dimensions of static features, including soil type, soil nitrogen content, soil phosphorus content, soil potassium content, crop variety, planting density, and growth stage) and is combined with Element-wise multiplication is performed to obtain stage-temporal interaction features, which are then fused with static attributes and dynamic temporal information. .

[0026] An improvement to the traditional XGBoost objective function is made by incorporating two core factors: soil type and crop variety, and introducing dynamic weights related to growth stages. The objective function can then be expressed as: (3); In the formula, The loss function; No. The actual output value of each sample; It is an improvement to XGBoost for the first Preliminary yield forecasts for one sample; It is a regular term; Will Input the improved XGBoost and optimize iteratively using the gradient boosting algorithm. Output preliminary production forecast values ,Will and The input is a fully connected nested layer, the output is the final predicted yield value, and the fusion model Res-BiLSTM-IXGBoost is obtained: (4); In the formula, It is a linear transformation operation; It is a feature concatenation operation; It is the weight matrix of the nested layer; It is the bias term of the nested layer.

[0027] S3: Constructing a Temperature and Humidity Prediction Model: Integrating multi-dimensional features to construct a feature matrix and standardizing it, an improved K-means clustering algorithm (introducing temperature and humidity weights) is used to determine the optimal number of clusters and their affiliation labels; for each cluster, nonlinear features are extracted through an improved SVR (introducing residual gradients and attention weights), and the residuals are fitted using a GBDT model to construct a Ti-AnnSVR-GBDT model for temperature and humidity prediction. Specifically, as shown... Figure 4 As shown: S3.1: Temperature, as a key driver of crop physiological metabolism and environmental stress response, requires precise prediction for intelligent environmental regulation. Through accurate temperature prediction, the system can proactively trigger environmental control strategies such as ventilation, heating / cooling, effectively mitigating abiotic stresses like high-temperature heat damage and low-temperature freezing damage. This maintains a suitable temperature range for crop growth within the greenhouse, ensuring the normal operation of physiological processes such as photosynthetic efficiency and enzyme activity. Ultimately, this provides environmental protection for crop yield formation and quality optimization, and is a key technology for achieving precision production and efficient resource utilization in greenhouse crops. First, it integrates basic environmental characteristics (measured temperature T, relative humidity H) and coupled interactive characteristics (…). ) and time-series dynamic characteristics (temperature change rate), humidity fluctuation standard deviation ( ), construct the feature matrix Furthermore, dimensional differences are eliminated through standardization. Based on this, a temperature and humidity weight term is introduced to improve the K-means algorithm, and the loss function is defined as: (5); In the formula, It is the first Feature vectors of each sample; It is the first The feature center vectors of each cluster; It is the first The sample to the first The square of the Euclidean distance between the centers of each cluster; It is the first The measured temperature value of each sample; It is the first The average temperature of all samples within a cluster; No. The relative humidity value of each sample; It is the first The mean humidity of all samples within a cluster; and These are separate weighted loss terms for temperature and humidity, respectively. The optimal number of clusters K is determined through K-means++ initialization and the elbow rule, and the cluster assignment labels of the samples are finally output. .

[0028] S3.2: Although traditional SVR can capture nonlinear relationships through kernel functions, it has shortcomings in handling residual optimization, especially for each cluster. First, nonlinear feature extraction is achieved by improving SVR. GBDT residual gradient is introduced to improve the SVR loss function. To enhance SVR's focus on key features, an attention weight term is introduced into the improved SVR loss function to dynamically weight the residual gradient contributions of different features. (6); In the formula, It is the first The number of samples in each cluster. , No. Loss slack variables for each sample; It is the dimension of the input features; These are the coefficients that balance the SVR loss term and the GBDT residual gradient term; It is the first Attention weights for each feature; No. The first sample GBDT residual gradients corresponding to each feature; It is the first in SVR The weight components of each feature; No. The first sample RBF kernel mapping function for each feature; It is the bias term of the SVR model; It is the regularization term of the SVR model.

[0029] By solving for Lagrange duality, an improved SVR prediction formula is obtained, which enhances the extraction of complex nonlinear correlation features between temperature and humidity. (7); In the formula, , It is the first The Lagrange multiplier corresponding to each sample; It is the first Kernel function for each feature; It is the first of the samples to be predicted One eigenvalue; It is a cluster Inner The first sample One eigenvalue; It is the first The SVR model bias term corresponding to each cluster.

[0030] Subsequently, the residuals of the improved SVR prediction were calculated. ( It is a cluster Inner (the true temperature and humidity values ​​of each sample), and the residual Input features The residuals are concatenated and trained to fit the GBDT model. GBDT generates decision trees iteratively. Minimize residual loss and output residual compensation value: (8); Finally, the preliminary predictions of the improved SVR are superimposed with the residual compensation of GBDT to obtain the temperature and humidity prediction model Ti-AnnSVR-GBDT, with intra-cluster fusion prediction: (9).

[0031] S4: Parameter Optimization: The CEO used the Chebyshev chaotic mapping algorithm to optimize key parameters (number of hidden layer neurons) of the Ti-AnnSVR-GBDT model using the chaotic evolution optimization algorithm. Learning rate Penalty coefficient This leads to the final prediction model, ICEO-Ti-AnnSVR-GBDT; that is, a sequence matrix is ​​generated using the Chebyshev chaotic mapping, mapped to the parameter interval to obtain the initial matrix, and iteratively updated through chaotic perturbation (adding factor matrices) and elite guidance (extracting elite parameters), with population selection iteratively outputting the globally optimal parameters. Specifically, as shown... Figure 5 As shown: S4.1: To improve the prediction accuracy and generalization ability of the Ti-AnnSVR-GBDT model in complex scenarios, it is necessary to adjust its key parameters (number of hidden layer neurons). Learning rate Penalty coefficient Optimize accordingly. , , The chaotic sequence matrix is ​​generated based on the Chebyshev chaotic mapping recursive formula with three parameters: (10); In the formula, , Population size; (correspond , , ); The chaotic sequence matrix is ​​linearly mapped to the actual optimization range of each parameter to obtain the initial parameter matrix. The mapping formula is: (11); In the formula, This is the lower bound of the parameter; This is the upper bound of the parameter; S4.2: Based on the initial parameter matrix described above, the parameter matrix is ​​iteratively updated using the Chebyshev chaotic perturbation and elite-guided mechanism. This updates the temporary parameter matrix output by the LFGP stage. Introducing the Chebyshev chaotic perturbation factor matrix Combined with dynamic disturbance intensity The perturbated parameter matrix is ​​then updated. (12); Extracting the elite parameter vector for each generation (Optimal parameter solution), generate the Chebyshev chaotic direction vector matrix. Combined with elite guidance weight Update the parameter matrix of the WRPU stage: (13); Ultimately, through population selection mechanisms, this is achieved. , , The parameter matrix is ​​iteratively optimized until the iteration terminates, at which point the globally optimal parameter combination is output.

[0032] S5: Hierarchical Collaborative Regulation: Combining the defined temperature and humidity stress thresholds with the growth stage, an explicit risk function is constructed, and an ensemble model is trained to predict the overall risk. The upper-level MPC sets soft constraints and a multi-objective function, and continuously optimizes the control sequence. The lower-level PID introduces feedforward terms to correct deviations, decouples the device control input through an allocation matrix, and fuses the control inputs using a linear superposition structure to obtain the STriskMPC-feedforward-PID controller, which collaboratively outputs the final control input. Specifically, such as... Figure 6 , 7 As shown: S5.1: The adaptability of organisms to temperature and humidity varies significantly at different growth stages, and using a fixed threshold can easily lead to misjudgment of risk. Therefore, a temperature stress threshold is defined based on the physiological characteristics of growth stage s. Humidity stress threshold An explicit risk function is constructed based on a threshold to quantify the nonlinear effects of temperature and humidity exceeding the limits. (14); In the formula, It is the high / low temperature stress weighting coefficient; It is the high / low humidity stress weighting coefficient; It is the ReLU function; It is the high / low temperature stress threshold corresponding to growth stage s; This refers to the high / low humidity stress threshold corresponding to growth stage s; Explicit risk functions only reflect immediate stress; potential risks such as diseases and yield losses require quantification based on multiple factors. A learning reinforcement module is introduced to train an ensemble model using historical data on environment, soil, and growth status. Predicting the future Comprehensive risks: (15); In the formula, It is the Ti-AnnSVR-GBDT model; It is an identifier for the type of risk in the model; It is an index for future time steps; These are predictions from the Ti-AnnSVR-GBDT model; This refers to the rate of change of temperature and humidity; S5.2: Employs MPC as the upper-level controller, using linearized state equations to describe temperature and humidity dynamics, balancing system complexity with real-time control performance. (16); In the formula, Temperature and humidity status; This refers to the equipment output (heating power, ventilation volume); For external disturbances (sudden temperature changes, rainfall), A is the state transition matrix; B is the control input matrix. Define the comfort zone based on the stress threshold. , To avoid hard constraints leading to an unsolvable optimization problem, slack variables are introduced. , , , By transforming it into a soft constraint, the relaxation variables can be weighted using the objective function to limit the degree of temperature and humidity deviation, thus balancing control stability and accuracy. (17); Design a multi-objective MPC objective function that combines tracking accuracy, energy consumption, constraint relaxation, and stress risk: (18); In the formula, These are the weighting coefficients of the constraint relaxation term; These are the weighting coefficients for explicit risk items; It is the weighting coefficient of the implicit risk item; It is the future The predicted state vector at time step; It is the future Predicted temperature and humidity levels inside the greenhouse at all times; This is an ideal reference value; It is the future The amount of control at any given moment; Rolling optimization, with state equations, soft constraints, and control range as conditions, solves for the future. Step-optimal control sequence, only execute the current one The state is updated and the optimization is repeated after each disturbance to ensure dynamic adaptability.

[0033] S5.3: MPC optimization has a long cycle and struggles to handle sudden disturbances. Therefore, feedforward is introduced into the standard PID controller as the lower-level controller to quickly correct deviations. Taking temperature as an example, and humidity similarly, the temperature error is defined as follows: (19); In the formula, This is the MPC benchmark value; These are measured values; The integral term uses an anti-saturation strategy to update and feedback the control quantity. (20); In the formula, It is the saturation function of the error; These are measured values; Incorporating a feedforward term based on Ti-AnnVR-GBDT predictions to compensate for environmental change trends and the total PID output (feedback and feedforward terms superimposed): (twenty one); S5.4: Upper-layer MPC per Optimize once per PID step, outputting a baseline control quantity. The lower-level PID each The system corrects errors every second, responding to real-time disturbances. In facility environmental control systems, actuators commonly exhibit multi-variable coupling characteristics. For example, the airflow exchange process in a ventilation unit simultaneously affects temperature and humidity state variables, forming an input-output cross-coupling relationship. Direct control can easily lead to interference between state variables, reducing control accuracy. To address this issue, an allocation matrix is ​​used... , Decoupling control quantity: (twenty two); in, , ; The number of devices ( (Includes at least heating and ventilation equipment); Matrix elements are assigned values ​​only in the effective channels of the device-state. Heating units only operate under temperature conditions. The row vector corresponding to the central heating channel is a unit vector. The corresponding row vector is the zero vector; the ventilation unit simultaneously operates under temperature and humidity conditions. , All row vectors corresponding to the ventilation ducts are non-zero vectors. Through linear mapping of the allocation matrix, decoupling of control quantities between coupled channels can be achieved, ensuring the independent control accuracy of multi-state quantities. The synergy between the global optimization characteristics of the upper-level model predictive control (MPC) and the local disturbance suppression capability of the lower-level PID controller is achieved by using a linear superposition structure to fuse the control quantities. (twenty three).

[0034] It should be noted that the arrangement of each step in a flow layout is merely one embodiment of the present invention, and other arrangements may also be used, which are not limited in the present invention.

[0035] The embodiments of the present invention have the following technical effects: 1. Improve data utilization and prediction accuracy: The system integrates environmental, soil and crop time-series data collected from multiple regions and multiple sensors. After preprocessing, it combines Res-BiLSTM to extract time-series features, improves XGBoost to build a yield prediction model, and improves K-means clustering + Ti-AnnSVR-GBDT model to predict temperature and humidity. The system also optimizes model parameters, which greatly improves prediction accuracy and generalization ability, and solves the problem of traditional single algorithms being unable to adapt to complex environments.

[0036] 2. Achieve precise dynamic environmental control: Define temperature and humidity stress thresholds and risk functions based on crop growth stage characteristics, and adopt a hierarchical control mechanism of upper-level MPC global optimization + lower-level feedforward PID real-time correction to decouple equipment control quantities, avoid misjudgment of fixed thresholds, solve the problem of low control accuracy caused by multi-device coupling and disturbance, and accurately maintain a suitable growth environment. 3. Ensure stable crop yield and quality: Rely on temperature and humidity forecasts to trigger control strategies in advance, avoid abiotic stresses, and ensure normal crop physiological processes; combine yield forecasts to guide production decisions, take into account the differentiated needs of different growth stages, solve the problem of unstable yield and quality under artificial management, and improve production efficiency. 4. Enhance system anti-interference and adaptability: The control process introduces feedforward PID to correct sudden disturbances, MPC rolling optimization to adapt to environmental changes, and improved model design to adapt to different crop and soil planting scenarios, breaking the limitation of traditional single-scenario adaptation and improving the risk resistance of facility agriculture production.

[0037] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for monitoring and controlling the greenhouse crop production environment based on multiple angles and strategies, characterized in that, Includes the following steps: S1: Data Acquisition and Preprocessing: Set up monitoring points to collect environmental parameters and construct an environmental parameter matrix; measure soil and crop information to establish a soil and crop information matrix; preprocess the environmental parameter matrix and the soil and crop information matrix; S2: Construct a yield prediction model: Slice the environmental parameter matrix according to the crop growth stage, and integrate the time-series-determinable features in the soil and crop information matrix to construct an LSTM input. Use Res-BiLSTM to extract time-series features, and combine with the improved XGBoost to construct a Res-BiLSTM-IXGBoost model. S3: Construct a temperature and humidity prediction model: Integrate multi-dimensional features to construct a feature matrix and complete standardization. Use an improved K-means clustering algorithm to determine the optimal number of clusters and cluster affiliation labels. For each cluster, extract nonlinear features by improving SVR and combine the residuals with the GBDT model to construct a Ti-AnnSVR-GBDT model. S4: Parameter optimization: The CEO algorithm is modified by Chebyshev chaotic mapping to obtain the ICEO algorithm. The ICEO algorithm optimizes the key parameters of the Ti-AnnSVR-GBDT model. S5: Hierarchical collaborative regulation: Combine the growth stage to define temperature and humidity stress thresholds, construct an explicit risk function, train an ensemble model to predict comprehensive risk; the upper-level MPC sets soft constraints and multi-objective functions, and continuously optimizes the control sequence; The lower-level PID introduces a feedforward term to correct the deviation, decouples the device control quantity through the allocation matrix, and uses a linear superposition structure to fuse the control quantity to obtain the STriskMPC-feedforward-PID controller, which then outputs the final control quantity in a coordinated manner.

2. The method for monitoring and controlling the greenhouse crop production environment according to claim 1, characterized in that: The monitoring point deployment in step S1 includes: setting up m monitoring points in the greenhouse, with each point equipped with temperature, humidity, light, and CO2 concentration sensors to continuously collect environmental parameters; soil and crop information includes soil physicochemical characteristics, static agronomic characteristics of crops, and dynamic growth indicators.

3. The method for monitoring and controlling the greenhouse crop production environment according to claim 1, characterized in that: The preprocessing of the environmental parameter matrix and soil crop information matrix in step S1 includes: mean interpolation to fill missing values, criterion to remove outliers, and Z-score standardization to unify dimensions.

4. The method for monitoring and controlling the greenhouse crop production environment according to claim 1, characterized in that: In step S2, after extracting temporal features using Res-BiLSTM, attention weights need to be calculated using MLP to generate a temporal feature vector. This vector is then element-wise multiplied with the static features, and finally, an improved XGBoost model is constructed to build the Res-BiLSTM-IXGBoost model, as follows: ； In the formula, It is a linear transformation operation; It is a feature concatenation operation; This is a preliminary production forecast; It is a temporal feature vector; It is the weight matrix of the nested layer; It is the bias term of the nested layer.

5. The method for monitoring and controlling the greenhouse crop production environment according to claim 1, characterized in that: In step S3, the improved K-means clustering algorithm introduces a temperature and humidity weight term to improve the loss function, and determines the optimal number of clusters K through K-means++ initialization and the elbow rule; the loss function is expressed as follows: ； In the formula, It is the first Feature vectors of each sample; It is the first The feature center vectors of each cluster; It is the first The sample to the first The square of the Euclidean distance between the centers of each cluster; It is the first The measured temperature values ​​of each sample; It is the first The average temperature of all samples within a cluster; No. The relative humidity value of each sample; It is the first The mean humidity of all samples within a cluster; and These are separate weighted loss terms for temperature and humidity, respectively.

6. The method for monitoring and controlling the greenhouse crop production environment according to claim 5, characterized in that: In step S3, the loss function of the improved SVR incorporates attention weights; the GBDT model improves the prediction residuals of the SVR by iteratively generating a decision tree; the constructed Ti-AnnSVR-GBDT model is as follows: ； In the formula, , It is the first The Lagrange multiplier corresponding to each sample; It is the first Kernel function for each feature; It is the first of the samples to be predicted One eigenvalue; It is a cluster Inner The first sample One eigenvalue; It is the first The SVR model bias term corresponding to each cluster; It is the residual compensation value; It is a cluster Inner The actual temperature and humidity values ​​of each sample.

7. The method for monitoring and controlling the greenhouse crop production environment according to claim 1, characterized in that: In step S4, the key parameters include the number of hidden layer neurons. Learning rate Penalty coefficient ; The Chebyshev chaotic mapping optimization includes: generating a chaotic sequence matrix through Chebyshev chaotic mapping and linearly mapping it to an initial matrix in the parameter range; iteratively updating the parameter matrix through chaotic perturbation and elite guidance mechanism; and outputting the globally optimal parameter combination.

8. The method for monitoring and controlling the greenhouse crop production environment according to claim 1, characterized in that: Step S5 specifically involves: S5.1: Define the temperature stress threshold based on the physiological characteristics of the growth stage. Humidity stress threshold Construct an explicit risk function based on the threshold: ； In the formula, It is the high / low temperature stress weighting coefficient; It is the high / low humidity stress weighting coefficient; It is the ReLU function; It is the high / low temperature stress threshold corresponding to growth stage s; This refers to the high / low humidity stress threshold corresponding to growth stage s; A learning enhancement module is introduced to train an ensemble model using historical data on environment, soil, and growth status. Predicting the future Step-by-step comprehensive risks: ； In the formula, It is the Ti-AnnSVR-GBDT model; It is an identifier for the type of risk in the model; It is an index for future time steps; These are predictions from the Ti-AnnSVR-GBDT model; This refers to the rate of change of temperature and humidity; S5.2: Employs MPC as the upper-level controller, using linearized state equations to describe temperature and humidity dynamics, balancing system complexity with real-time control performance. ； In the formula, Temperature and humidity status; This is the output of the device; External disturbances; A is the state transition matrix; B is the control input matrix. Define the comfort zone based on the stress threshold. , Introducing slack variables , , , Transform into soft constraints: ； Design a multi-objective MPC objective function: (18)； In the formula, These are the weighting coefficients of the constraint relaxation term; These are the weighting coefficients for explicit risk items; It is the weighting coefficient of the implicit risk item; It is the future The predicted state vector at time step; It is the future Predicted temperature and humidity levels inside the greenhouse at any given time; This is an ideal reference value; It is the future The amount of control at any given moment; Rolling optimization, with state equations, soft constraints, and control range as conditions, solves for the future. Step-optimal control sequence; S5.3: Define temperature error: ； In the formula, This is the MPC benchmark value; These are measured values; The integral term uses an anti-saturation strategy to update and feedback the control quantity. ； In the formula, It is the saturation function of the error; These are measured values; Incorporating feedforward terms based on Ti-AnnVR-GBDT predictions: ； S5.4: Upper-layer MPC per Optimize once per PID step, outputting a baseline control quantity. The lower-level PID each Corrects every second, responding to real-time disturbances; By allocating the matrix , Decoupling control quantity: ； in, , ; It refers to the number of devices; The global optimization characteristics of the upper-level model predictive control (MPC) and the local disturbance suppression capability of the lower-level PID controller are combined using a linear superposition structure to fuse the control inputs. 。

Images