Greenhouse intelligent control system based on machine learning

Abstract

The invention provides a greenhouse intelligent control system based on machine learning. The greenhouse intelligent control system comprises the steps of collecting real-time monitoring data of a multi-dimensional sensor in a greenhouse, splicing continuous sampling data according to a time step, constructing a time sequence state feature chain representing an environment factor time-varying rule, and inputting a deep neural network for resolving to obtain a first group of equipment control parameters; deducing a future environment state change rate according to the parameters based on an environment dynamic model, calculating an environment factor change vector included angle cosine value to represent a mutual exclusion influence value, converting the value into an attention mask, and attenuating a historical data fragment weight causing a mutual exclusion trend in a time sequence state feature chain; and inputting the feature chain after weight attenuation into the model, generating a second group of control parameters, extracting a high-energy-consumption execution mechanism action instruction in the parameters, predicting energy consumption and environment steady state disturbance gain, constructing a balance coefficient to correct the action duration of other execution mechanisms, and outputting a control instruction sequence.

Description

Technical Field

[0001] This application belongs to the field of intelligent control, and in particular relates to an intelligent control system for greenhouses based on machine learning. Background Technology

[0002] In greenhouse control systems, multidimensional sensors are deployed to monitor environmental parameters in real time, and PID control algorithms or basic machine learning models are used to adjust actuators. However, the greenhouse environment is a multivariable and time-lag-dependent system, with interaction mechanisms between various environmental factors. Existing conventional control methods lack the ability to perform in-depth time-series analysis on continuously sampled data, neglect the interactions between environmental factors, and struggle to detect the underlying evolution trends in historical data.

[0003] Existing neural network models, when generating control parameters, cannot identify and avoid the negative synergistic effect where the adjustment of one factor leads to the deterioration of other factors. For example, forced cooling may cause humidity fluctuations, and they lack the ability to perform weight decay and recalculation on historical data that leads to mutually exclusive trends. Current technologies ignore the energy consumption differences between high-energy-consuming and ordinary actuators and the degree of disturbance to environmental steady state. They lack a deduction step based on environmental dynamics models and cannot construct balance coefficients based on the disturbance gain of high-energy-consuming equipment to correct the action duration of other actuators. Such control strategies not only easily induce greenhouse environment oscillations and overshoots but also cause energy waste, making it difficult to achieve optimal energy efficiency while ensuring environmental steady state. Summary of the Invention

[0004] This invention proposes a machine learning-based intelligent control system for greenhouses to address the shortcomings of existing technologies, such as the lack of ability to weight and recalculate historical data leading to mutually exclusive trends, and the inability to construct balance coefficients based on the disturbance gains of high-energy-consuming equipment to correct the action duration of other actuators. The system includes: The calculation module is used to collect real-time monitoring data from multi-dimensional sensors in the greenhouse, splice the continuously sampled data according to the time step, and construct a time-series state feature chain representing the time-varying law of environmental factors; input the time-series state feature chain into a preset deep neural network to calculate the first set of equipment control parameters; The conversion module is used to use the first set of equipment control parameters to extrapolate the rate of change of environmental state within a future time window, calculate the cosine value of the angle between different environmental factor change vectors to represent the mutual exclusion influence value, convert the mutual exclusion influence value into an attention mask, and perform weight decay processing on the historical data segments in the time-series state feature chain that lead to the mutual exclusion trend. The generation module is used to input the time-series state feature chain after weight decay back into the deep neural network to recalculate and generate the second set of device control parameters. The output module is used to extract the action commands of the preset high-energy-consuming actuators from the second set of equipment control parameters, predict the energy consumption of the commands and the disturbance gain to the steady state of the environment, construct a balance coefficient using the disturbance gain, correct the action duration of the other actuators in the same set of parameters, and output the corrected control command sequence.

[0005] This invention constructs a time-series state feature chain representing the time-varying patterns of environmental factors by collecting continuous sampling data from multi-dimensional sensors within a greenhouse and sequentially stitching them together. Using a model to extrapolate the environmental state within a future time window, the cosine of the angle between the change vectors of different factors is calculated to represent mutually exclusive influences. This is then converted into an attention mask to attenuate the weights of historical data segments that lead to mutual exclusion. This mechanism identifies and avoids conflict effects in the multivariate adjustment process. Compensated control parameters are generated by re-inputting the data into a neural network, ensuring the coordination of multiple devices. For high-energy-consuming actuators, the disturbance gain of these actuators to the steady-state environment is predicted, and a balance coefficient is constructed to correct the action duration of other actuators in the same group. This reduces the environmental impact of high-energy-consuming equipment actions and improves the operational stability and energy efficiency of the greenhouse control system. Attached Figure Description

[0006] Figure 1 A schematic diagram of the time-varying temporal state feature chain matrix of environmental factors; Figure 2 A schematic diagram of the first set of device control parameters output by the deep neural network; Figure 3 This is a schematic diagram illustrating the detection of vector mutual exclusion of environmental factor changes. Figure 4 This is a schematic diagram of the balance coefficient and duration correction based on the perturbation gain. Detailed Implementation

[0007] The features and exemplary embodiments of various aspects of this application will be described in detail below. To make the objectives, technical solutions, and advantages of this application clearer, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are only intended to explain this application and not to limit it. For those skilled in the art, this application can be implemented without some of these specific details. The following description of the embodiments is merely to provide a better understanding of this application by illustrating examples.

[0008] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising..." does not exclude the presence of additional identical elements in the process, method, article, or apparatus that includes said element.

[0009] In the first embodiment, the present invention proposes a machine learning-based intelligent control system for greenhouses, comprising: The calculation module is used to collect real-time monitoring data from multi-dimensional sensors in the greenhouse, splice the continuously sampled data according to the time step, and construct a time-series state feature chain representing the time-varying law of environmental factors; input the time-series state feature chain into a preset deep neural network to calculate the first set of equipment control parameters; By deploying a sensor network in different areas of the greenhouse, the data from air temperature sensors, air humidity sensors, light intensity sensors, soil moisture sensors, and carbon dioxide concentration sensors are collected synchronously. The sampling time step is set to five minutes. After denoising and smoothing the collected raw data, the max-min normalization method is used to map the data of each dimension to the interval between zero and one, forming a multi-dimensional feature vector at the current moment. Using a sliding window mechanism, the window length is set to twelve time steps, i.e., one hour. The multi-dimensional feature vectors of the current moment and the past eleven moments are constructed into a two-dimensional matrix containing time dimension and feature dimension, which represents the time-series state feature chain of the time-varying law of environmental factors.

[0010] A Long Short-Term Memory (LSTM) network is used as the pre-built deep neural network. This network includes an input layer, several hidden layers, and an output layer. The pre-constructed time-series state feature chain is fed into the input layer as input data. The hidden layers use forget gates, input gates, and output gates to extract features from the relationships in the time-series data. The network output layer outputs a vector with values ​​between 0 and 1 through a sigmoid activation function. Each dimension of this vector corresponds to the control action intensity or on / off state of a specific actuator in the greenhouse, such as the percentage of the roof window opening angle, the fan speed setting, and the shading net unfolding ratio, thereby calculating the first set of preliminary equipment control parameters.

[0011] In an optional embodiment, the real-time monitoring data collected from multi-dimensional sensors within the greenhouse is spliced ​​together according to a time step to construct a time-series state feature chain representing the time-varying patterns of environmental factors, including: The sensor array deployed inside the greenhouse reads the values ​​of air temperature, air humidity, light intensity, and soil moisture. The read value is subjected to max-min normalization to map the value to the [0,1] interval; Set the time step t in minutes and select the normalized data of the 60 time steps before the current time. The selected data are arranged in chronological order to form a matrix. The matrix is ​​the time-series state feature chain.

[0012] Specifically, multiple sets of sensor nodes are deployed in the center and surrounding areas of the greenhouse, with a sampling frequency of once per minute to read real-time air temperature (e.g., 25℃), air humidity (e.g., 60%), light intensity (e.g., 1500 lux), and soil moisture (e.g., 40%). The raw data is processed using the max-min normalization formula. Assuming the temperature measurement range is -10 to 50℃, 25℃ is mapped to approximately 0.58. All environmental factor values ​​are converted into dimensionless floating-point numbers between 0 and 1, eliminating the interference of different sensor dimensions on the model weights.

[0013] A sliding time window technique was used to construct a data matrix. Starting from the current moment, data was traced back 60 minutes, and normalized data rows for each of the 60 time steps were extracted. Each row contained the four environmental features mentioned above. The data was then stacked in chronological order to form a 60×4 two-dimensional matrix X. This matrix fully preserves the fluctuation trend of greenhouse environmental factors over the past hour, reflecting the trajectory of environmental parameters over time. Figure 1 .

[0014] In an optional embodiment, the step of inputting the time-series state feature chain into a preset deep neural network to calculate the first set of device control parameters includes: Construct a deep neural network consisting of two long short-term memory network layers and one fully connected layer; The temporal state feature chain X is input into the long short-term memory network layer to extract the time-dimensional feature vector. ; The feature vector is input into the fully connected layer, and the output vector y is calculated using the Sigmoid activation function. Where W is the weight matrix and b is the bias vector. For the Sigmoid function, such that ; The elements in the output vector y are mapped to the rated operating range of each actuator to calculate the operation time of the ventilation fan, the unfolding ratio of the shading net, the start-up time of the irrigation pump, and the start-up time of the heating boiler, thus obtaining the first set of equipment control parameters.

[0015] The deep neural network structure comprises an input layer, two stacked long short-term memory (LSM) network layers, and a fully connected layer. The first LSM layer has 128 neurons to detect long-term features, the second layer has 64 neurons to extract short-term features, and the fully connected layer has 4 output nodes and uses the sigmoid activation function. The training set uses environmental monitoring data from the past year in a greenhouse history database as input samples, and the optimal equipment control commands verified by agricultural experts at the corresponding time point are used as target labels. During training, the mean squared error loss function is used to calculate the deviation between the model output and the target label, and the Adam optimizer is used to iteratively update the weight matrix W and the bias vector b through backpropagation until the model converges. The network input is a 60×4 temporal state feature chain matrix X, and the output is a control parameter vector y with values ​​ranging from 0 to 1. In application, the feature matrix is ​​input into the network to obtain the output vector. The vector elements are then mapped to physical control quantities. For example, an output value of 0.8 corresponds to a ventilation fan, so an 8-minute operating time is calculated; an output value of 0.5 corresponds to a shading net, so a 50% deployment ratio is calculated. This forms the first set of equipment control parameters. Figure 2 .

[0016] The conversion module is used to use the first set of equipment control parameters to extrapolate the rate of change of environmental state within a future time window, calculate the cosine value of the angle between different environmental factor change vectors to represent the mutual exclusion influence value, convert the mutual exclusion influence value into an attention mask, and perform weight decay processing on the historical data segments in the time-series state feature chain that lead to the mutual exclusion trend. The environmental dynamics model is constructed based on the greenhouse energy conservation equation and the mass conservation equation, including the heat balance equation and the humid air state equation. Substituting the current environmental state data and the first set of equipment control parameters into the model, the model simulates and extrapolates the rate of temperature change and the rate of humidity change in the next fifteen minutes. The temperature change trend and the humidity change trend are respectively constructed as change vectors on a two-dimensional plane. The cosine value of the angle between the temperature change vector and the humidity change vector is calculated using the vector dot product formula. When the cosine value of the angle is negative, it indicates that there is a conflict in the direction of change between the two. For example, cooling operation leads to an abnormal increase in humidity. The absolute value of the cosine value of the angle represents the mutual exclusion influence value between different environmental factors.

[0017] A threshold for the mutual exclusion influence value is set. When the calculated mutual exclusion influence value exceeds this threshold, the gradient analysis method is used to backtrack the deep neural network to identify the time step segment in the time-series state feature chain that contributes the most to the current mutual exclusion output. An attention mask matrix with the same dimension as the time-series state feature chain is generated. In the mask matrix, the weight coefficients of the time step positions that lead to the mutual exclusion trend are set to low coefficient values ​​between 0.1 and 0.5, while the other positions are kept at 1. The attention mask matrix is ​​then multiplied element-wise with the original time-series state feature chain to reduce the influence of historical data that leads to control conflicts, thus completing the weight decay process.

[0018] In an optional embodiment, the environmental dynamics model is constructed by: first, establishing a core set of equations describing the dynamics of the greenhouse system based on the two major physical laws of energy conservation and mass conservation. The core of the model comprises two main equations: one is the heat balance equation, in the form of… The first method correlates the rate of change of air temperature inside the greenhouse with the dynamic balance of various heat inputs and heat losses, where equipment control parameters are directly substituted into the corresponding Q term as control terms; the second method is the humid air state equation, in the form of... The model correlates the rate of change of atmospheric water vapor density with various water vapor sources and loss processes, with control parameters serving as key inputs. It uses currently measured temperature, humidity, and other state data as initial conditions for the differential equations, and combines equipment control commands with the energy and mass change rates of the environment through preset physical parameters. This constructs a system solvable numerically and capable of predicting the evolution of the environmental state under given control actions.

[0019] In an optional embodiment, the step of using the first set of device control parameters to extrapolate the rate of change of environmental state within a future time window includes: Retrieve the preset set of differential equations for greenhouse energy and mass balance: Where X(t) is the environmental state vector, U(t) is the control input containing the first set of equipment control parameters, and D(t) is the external disturbance; The fourth-order Runge-Kutta method was used to numerically solve the system of differential equations, yielding the temperature prediction sequence for the next N time steps. Humidity prediction series ; The state change rate sequence of environmental factors is calculated using the first-order difference formula.

[0020] Specifically, the system retrieves a set of differential equations for greenhouse energy and mass balance based on thermodynamic laws. This physical model quantitatively represents the relationship between the enthalpy of the air and water vapor content inside the greenhouse, and external solar radiation and equipment operation. The first set of equipment control parameters generated by the neural network is used as the control input U(t). Combined with the current environmental state X(t) and external weather forecast data D(t), a fourth-order Runge-Kutta numerical integration algorithm is used for iterative solution. With a prediction step size of 1 minute, the system extrapolates the state 30 minutes into the future, thereby obtaining the temperature prediction sequence for a future period under the influence of these control parameters. Humidity prediction series .

[0021] For example, the first set of equipment control parameters is used as the control input vector U(t), combined with the measured environmental state vector X(t) = [temperature, humidity] at the current moment and the known external disturbance D(t), to input the preset greenhouse energy and mass balance differential equations. The fourth-order Runge-Kutta method is used for numerical solution, with the specific steps being: setting the integration step size... It is 1 minute, starting from the current moment. Begin iteration; in each iteration, calculate the four slopes in sequence: Based on the current state ; Based on the estimated state ; Based on another predicted state ; Based on the estimated state According to the formula Update the state vector X; repeat this process N=30 times to obtain the temperature prediction sequence for each minute of the next 30 minutes. ] and humidity prediction value series .

[0022] Based on this, the first-order difference calculation method is applied to the predicted sequence, that is, the predicted value of the previous time step is subtracted from the predicted value of the next time step and divided by the time interval. For example, if the predicted temperature is 25.0°C at the 5th minute and 25.5°C at the 6th minute, then the temperature change rate at that moment is 0.5°C per minute. The environmental predictions are then converted into a sequence of environmental factor state change rates.

[0023] In an optional embodiment, calculating the cosine of the angle between the vectors of different environmental factor changes to represent the mutually exclusive influence value includes: Construct a temperature change vector consisting of the rate of temperature change within the prediction time window. and the humidity change vector composed of the humidity change rate ; Calculate the cosine similarity of the angle between the temperature change vector and the humidity change vector; Determine mutual exclusivity and calculate mutual exclusivity impact value If the cosine of the included angle is less than 0, then ;otherwise .

[0024] The temperature change rate values ​​obtained above for the next 30 time steps are used to form a temperature change vector. At the same time, the corresponding humidity change rate values ​​are used to form a humidity change vector. The two high-dimensional vectors represent the evolution directions of the temperature and humidity fields inside the greenhouse in phase space under the current control strategy. The cosine of the angle between the two vectors is calculated using the vector dot product formula.

[0025] The sign of the cosine of the angle between environmental factors is used to determine whether there is mutual exclusivity. If the calculation result is negative, such as -0.8, the changes in temperature and humidity show an obtuse angle or even an inverse relationship, meaning that an improvement in one factor may lead to a deterioration in another. In this case, the absolute value of 0.8 is taken as the mutual exclusivity value. If the calculation result is non-negative, it indicates that the changing trends of the two are coordinated or unrelated, and there is no control conflict. In this case, the mutual exclusion influence value is set to 0. Figure 3 .

[0026] In an optional embodiment, the step of converting the mutual exclusion influence value into an attention mask and performing weight decay processing on the historical data segments in the temporal state feature chain that lead to the mutual exclusion trend includes: Using mutually exclusive influence values Constructing attention weight factors ; Generate a mask matrix M with the same dimensions as the time-series state feature chain. For the columns corresponding to the temperature and humidity feature dimensions in the matrix, set the element values ​​to... The element values ​​for the other dimensions are set to 1; Perform a Hadamard product operation between the mask matrix and the temporal state feature chain to obtain the feature matrix after weight decay. : .

[0027] Specifically, based on the mutually exclusive influence values ​​calculated above... The attention weighting factor is calculated using a formula. For example, when the mutual exclusion influence value is 0.8, the calculated weighting factor is... A value of 0.2 indicates a severe control conflict, necessitating a reduction in the focus on related features. A 60×4 mask matrix M is constructed. In this matrix, all elements in the columns corresponding to the two conflicting environmental factors (temperature and humidity) are uniformly assigned a value of 0.2, while the elements in the columns corresponding to the non-conflicting factors (light and soil moisture) are kept to a value of 1.

[0028] Perform the Hadamard product operation between the mask matrix and the original temporal state feature chain X, i.e., perform element-by-element multiplication. Scale and attenuate the temperature and humidity feature values ​​from the historical data by a factor of 0.2, while keeping other features unchanged, to generate the corrected feature matrix. During model iteration or re-input, it can suppress the influence of historical state features that lead to conflicts in control objectives on decision-making, and guide the model to focus on other feature information that is conducive to balance.

[0029] The generation module is used to input the time-series state feature chain after weight decay back into the deep neural network to recalculate and generate the second set of device control parameters. The modified feature chain, after attention masking, is re-input into the Long Short-Term Memory (LSTM) network. After suppressing the interference of historical features that cause conflicts, the network performs forward propagation calculations, and the output layer regenerates a new set of control vectors. While maintaining the main control objective, this set of vectors adjusts the equipment parameters that cause conflicts. For example, while maintaining the cooling intensity, it fine-tunes the opening amplitude of the wet curtain to compensate for humidity changes, thereby generating a second set of equipment control parameters.

[0030] The output module is used to extract the action commands of the preset high-energy-consuming actuators from the second set of equipment control parameters, predict the energy consumption of the commands and the disturbance gain to the steady state of the environment, construct a balance coefficient using the disturbance gain, correct the action duration of the other actuators in the same set of parameters, and output the corrected control command sequence.

[0031] The system identifies commands from the second set of equipment control parameters related to electric heaters, high-power wet curtain pumps, and preset high-energy-consuming actuators. Based on rated power and command running time, it calculates the estimated energy consumption. Simultaneously, it predicts the contribution of the high-energy-consuming equipment's instantaneous action to the rate of change in ambient temperature and humidity using an environmental dynamics model. This contribution is defined as the disturbance gain. The disturbance gain is mapped to a balance coefficient between 0 and 1 using an inverse proportional function or an exponential decay function; that is, the larger the disturbance, the smaller the coefficient. Other actuators in the same parameter group, such as internal circulation fans and supplementary lighting, are selected, and their original action duration is multiplied by the square root or logarithm of this balance coefficient to reduce the environmental impact caused by the high-energy-consuming equipment. The corrected action commands for each device are then packaged and sent to the underlying PLC controller for execution.

[0032] In an optional embodiment, the step of extracting the action commands of preset high-energy-consuming actuators from the second set of device control parameters, and predicting the energy consumption of the commands and the disturbance gain to the steady state of the environment, includes: Identify the start-up duration command for the corresponding heating boiler in the second set of equipment control parameters. ; According to the rated power of the heating boiler Calculate energy consumption prediction values ; Calculate using greenhouse heat load model The absolute overshoot of indoor temperature relative to the set target caused by turning on the heating boiler for a given duration. ; Calculate the disturbance gain Defined as the product of normalized energy consumption and normalized temperature overshoot: in, This represents the maximum allowable energy consumption value within a preset time window. This is the preset temperature tolerance constant.

[0033] The optimized second set of equipment control parameters were analyzed to identify and extract the start-up duration command for high-energy-consuming equipment in the heating boiler. For example, the instruction might be to turn on for 30 minutes. Based on the rated power of the heating boiler, such as 50 kilowatts, the estimated energy consumption for this operating cycle can be calculated. The value is 25 kWh. Simultaneously, a greenhouse heat load physical model is invoked to simulate the indoor temperature trajectory after the boiler operates at full power for 30 minutes under current environmental conditions. This trajectory is then compared with a preset optimal target temperature curve to determine the maximum absolute overshoot of the temperature deviation from the target value. For example, an overshoot of 3°C.

[0034] Calculate the combined disturbance gain Assess the impact of this high-energy-consuming action on the system. Divide the predicted energy consumption of 25 kWh by the set maximum allowable energy consumption value, such as 100 kWh, to obtain an energy consumption ratio of 0.25; divide the temperature overshoot of 3°C by the allowable deviation constant, such as 1°C, to obtain an overshoot ratio of 3. Multiplying the two yields a disturbance gain of 0.75. This index reflects the degree of disruption to the steady state of the greenhouse thermal environment caused by the high-energy-consuming equipment's action while consuming energy; a higher value indicates a more severe disturbance to the equilibrium caused by the action.

[0035] In an optional embodiment, the step of constructing a balance coefficient using the perturbation gain to correct the action duration of other actuators in the same parameter group includes: The perturbation gain Substitute into the exponential decay function to calculate the equilibrium coefficient : Where k is a preset sensitivity constant; Select the original operating time of the corresponding ventilation fan from the same group of parameters. ; The original motion duration is corrected using the aforementioned balance coefficient to obtain the corrected motion duration. : Will Update to the control command sequence.

[0036] Specifically, using the calculated disturbance gain A damping adjustment mechanism is constructed. The sensitivity constant k is set to 0.5, and the aforementioned disturbance gain of 0.75 is substituted into the exponential damping function to calculate an equilibrium coefficient of approximately 0.687. This coefficient represents the degree of conservative control required to maintain steady state in the presence of a strong disturbance source.

[0037] The coefficients are applied to correct related equipment, such as ventilation fans, within the same group of control parameters. Assume the original calculated operating time of the ventilation fan. The original duration was 10 minutes. By multiplying this by a balance factor of 0.687, the corrected motion duration was obtained. Approximately 6.87 minutes. The corrected value is updated to the control command sequence to ensure heating requirements are met while suppressing the operating time of ventilation equipment, thus avoiding energy waste and drastic fluctuations in environmental parameters caused by simultaneous high-intensity operation of heating and ventilation. Figure 4 .

[0038] It should be clarified that this application is not limited to the specific configurations and processes described above and shown in the figures. For the sake of brevity, detailed descriptions of known methods are omitted here. In the above embodiments, several specific steps are described and shown as examples. However, the method process of this application is not limited to the specific steps described and shown. Those skilled in the art can make various changes, modifications, and additions, or change the order of steps, after understanding the spirit of this application.

[0039] The functional modules shown in the above-described block diagram can be implemented as hardware, software, firmware, or a combination thereof. When implemented in hardware, they can be, for example, electronic circuits, application-specific integrated circuits (ASICs), appropriate firmware, plug-ins, function cards, etc. When implemented in software, the elements of this application are programs or code segments used to perform the required tasks. Programs or code segments can be stored on a machine-readable medium or transmitted over a transmission medium or communication link via data signals carried on a carrier wave. "Machine-readable medium" can include any medium capable of storing or transmitting information. Examples of machine-readable media include electronic circuits, semiconductor memory devices, ROM, flash memory, erasable ROM (EROM), floppy disks, CD-ROMs, optical disks, hard disks, fiber optic media, radio frequency (RF) links, etc. Code segments can be downloaded via computer networks such as the Internet, intranets, etc.

[0040] It should also be noted that the exemplary embodiments mentioned in this application describe methods or systems based on a series of steps or apparatus. However, this application is not limited to the order of the above steps; that is, the steps can be performed in the order mentioned in the embodiments, or in a different order, or several steps can be performed simultaneously.

[0041] The aspects of this application have been described above with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It should be understood that each block in the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing apparatus to produce a machine such that these instructions, executable via the processor of the computer or other programmable data processing apparatus, enable the implementation of the functions / actions specified in one or more blocks of the flowchart illustrations and / or block diagrams. Such a processor can be, but is not limited to, a general-purpose processor, a special-purpose processor, a special application processor, or a field-programmable logic circuit. It is also understood that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can also be implemented by dedicated hardware performing the specified functions or actions, or can be implemented by a combination of dedicated hardware and computer instructions.

[0042] The above description is merely a specific implementation of this application. Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, modules, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here. It should be understood that the protection scope of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and these modifications or substitutions should all be covered within the protection scope of this application.

Claims

1. A machine learning-based intelligent control system for greenhouses, characterized in that, Includes the following modules: The calculation module is used to collect real-time monitoring data from multi-dimensional sensors in the greenhouse, and splice the continuously sampled data according to the time step to construct a time-series state feature chain that represents the time-varying law of environmental factors. The time-series state feature chain is input into a preset deep neural network to calculate the first set of device control parameters; The conversion module is used to use the first set of equipment control parameters to extrapolate the rate of change of environmental state within a future time window, and to calculate the cosine value of the angle between the change vectors of different environmental factors to represent the mutually exclusive influence value. The mutual exclusion influence value is converted into an attention mask, and the weight decay processing is applied to the historical data segments in the temporal state feature chain that lead to the mutual exclusion trend. The generation module is used to input the time-series state feature chain after weight decay back into the deep neural network to recalculate and generate the second set of device control parameters. The output module is used to extract the action commands of the preset high-energy-consuming actuators from the second set of equipment control parameters, predict the energy consumption of the commands and the disturbance gain to the steady state of the environment, construct a balance coefficient using the disturbance gain, correct the action duration of the other actuators in the same set of parameters, and output the corrected control command sequence.

2. The system according to claim 1, characterized in that, The real-time monitoring data collected from multi-dimensional sensors within the greenhouse is spliced ​​together according to a time step to construct a time-series state feature chain representing the time-varying patterns of environmental factors, including: The sensor array deployed inside the greenhouse reads the values ​​of air temperature, air humidity, light intensity, and soil moisture. The read value is subjected to max-min normalization to map the value to the [0,1] interval; Set the time step t in minutes and select the normalized data of the 60 time steps before the current time. The selected data are arranged in chronological order to form a matrix. The matrix is ​​the time-series state feature chain.

3. The system according to claim 1, characterized in that, The step of inputting the time-series state feature chain into a preset deep neural network to calculate the first set of device control parameters includes: Construct a deep neural network consisting of two long short-term memory network layers and one fully connected layer; The temporal state feature chain X is input into the long short-term memory network layer to extract the time-dimensional feature vector. ; The feature vector is input into the fully connected layer, and the output vector y is calculated using the Sigmoid activation function. Where W is the weight matrix and b is the bias vector. For the Sigmoid function, such that ; The elements in the output vector y are mapped to the rated operating range of each actuator to calculate the operation time of the ventilation fan, the unfolding ratio of the shading net, the start-up time of the irrigation pump, and the start-up time of the heating boiler, thus obtaining the first set of equipment control parameters.

4. The system according to claim 1, characterized in that, The step of using the first set of equipment control parameters to extrapolate the rate of change of environmental state within a future time window includes: Retrieve the preset set of differential equations for greenhouse energy and mass balance: Where X(t) is the environmental state vector, U(t) is the control input containing the first set of equipment control parameters, and D(t) is the external disturbance; The fourth-order Runge-Kutta method was used to numerically solve the system of differential equations, yielding the temperature prediction sequence for the next N time steps. Humidity prediction series ; The state change rate sequence of environmental factors is calculated using the first-order difference formula.

5. The system according to claim 1, characterized in that, The calculation of the cosine value of the angle between the vectors of change of different environmental factors represents the mutually exclusive influence value, including: Construct a temperature change vector consisting of the rate of temperature change within the prediction time window. and the humidity change vector composed of the humidity change rate ; Calculate the cosine similarity of the angle between the temperature change vector and the humidity change vector; Determine mutual exclusivity and calculate mutual exclusivity impact value If the cosine of the included angle is less than 0, then ;otherwise .

6. The system according to claim 5, characterized in that, The step of converting the mutual exclusion influence value into an attention mask and performing weight decay processing on the historical data segments in the temporal state feature chain that lead to the mutual exclusion trend includes: Using mutually exclusive influence values Constructing attention weight factors ; Generate a mask matrix M with the same dimensions as the time-series state feature chain. For the columns corresponding to the temperature and humidity feature dimensions in the matrix, set the element values ​​to... The element values ​​for the other dimensions are set to 1; Perform a Hadamard product operation between the mask matrix and the temporal state feature chain to obtain the feature matrix after weight decay. .

7. The system according to claim 6, characterized in that, The step of extracting the action commands of the preset high-energy-consuming actuators from the second set of equipment control parameters, and predicting the energy consumption of the commands and the disturbance gain to the steady state of the environment, includes: Identify the start-up duration command for the corresponding heating boiler in the second set of equipment control parameters. ; According to the rated power of the heating boiler Calculate energy consumption prediction values ; Calculate using greenhouse heat load model The absolute overshoot of indoor temperature relative to the set target caused by turning on the heating boiler for a given duration. . ; Calculate the disturbance gain Defined as the product of normalized energy consumption and normalized temperature overshoot: in, This represents the maximum allowable energy consumption value within a preset time window. This is the preset temperature tolerance constant.

8. The system according to claim 1, characterized in that, The step of constructing a balance coefficient using the disturbance gain to correct the action time of other actuators in the same parameter group includes: The perturbation gain Substitute into the exponential decay function to calculate the equilibrium coefficient : Where k is a preset sensitivity constant; Select the original operating time of the corresponding ventilation fan from the same group of parameters. ; The original motion duration is corrected using the aforementioned balance coefficient to obtain the corrected motion duration. : Will Update to the control command sequence.

Images