Greenhouse environment intelligent prediction and regulation method based on multi-source sensing data fusion

Abstract

This invention discloses an intelligent prediction and control method for greenhouse environments based on multi-source sensor data fusion, comprising: collecting and preprocessing multi-source time-series data to generate a unified multi-source time-series dataset; establishing a causal topology and fusing the processed data to generate a causal latent variable sequence of equipment-environment-crop; constructing a response kernel structure based on WaveNet to generate short-term state prediction sequences for the environment and crops; constructing a stress time window and evaluating the stress degree to generate a set of candidate equipment control sequences; constructing a counterfactual simulation domain for the environment and crops and obtaining corresponding counterfactual simulation prediction results; and performing multi-objective optimization to select the optimal control strategy and drive various control devices to perform specific actions. This invention achieves accurate prediction and intelligent control of greenhouse environments through multi-source time-series data fusion, causal reasoning, and improved WaveNet predictive modeling.

Description

Technical Field

[0001] This invention relates to the field of multi-source sensor data fusion technology, and in particular to a method for intelligent prediction and control of greenhouse environment based on multi-source sensor data fusion. Background Technology

[0002] Currently, greenhouse environmental control systems generally rely on limited environmental sensor data such as temperature, humidity, light, and carbon dioxide levels. They adjust equipment like fans, sprinklers, heaters, and shading curtains using threshold control, simple logic control, or conventional PID control. In traditional solutions, environmental conditions are treated as independent variables, and the control process is based on instantaneous sensor readings, with little consideration for the dynamic impact of equipment actions on the environment. For example, starting a fan can cause a rapid drop in temperature within minutes, and a sprinkler action can cause a rapid increase in humidity instantaneously. High-frequency dynamics are difficult to accurately represent using conventional control methods and simple time-series models. Existing technologies typically rely on manual observation or simple image recognition for crop status monitoring, resulting in a weak correlation with environmental control and making it difficult to truly align control strategies with crop growth needs.

[0003] Existing research has attempted to use deep learning for environmental prediction, but most of these studies use single environmental sequences as input, lacking the ability to model the causal relationship between equipment actions and environmental responses. They also ignore the significant abruptness, lag, and volatility of environmental changes caused by equipment actions. Furthermore, related technologies have not established a systematic characterization of the physiological responses of crops to environmental changes. This means that control strategies often cannot assess the impact of regulatory behaviors on future crop stress levels in advance, leading to problems such as high energy consumption, over-regulation, or under-regulation. The insufficient accuracy of environmental prediction and the lack of a quantitative mechanism for crop responses cause existing intelligent greenhouse control systems to exhibit limited adaptability in dynamic and complex scenarios.

[0004] Current greenhouse control technologies generally lack the ability to perform counterfactual simulations and multi-objective evaluations of future control schemes. Traditional methods cannot extrapolate the environmental response and crop stress changes of different equipment combinations before executing control actions, nor can they find the optimal solution that balances crop safety, growth stability, and energy efficiency among multiple possible control paths. The lack of an integrated intelligent control framework that combines prediction and control makes existing greenhouse environmental control systems unable to meet the development needs of precision agriculture and intelligent agriculture.

[0005] Therefore, how to provide a method for intelligent prediction and control of greenhouse environment based on multi-source sensor data fusion is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose an intelligent prediction and control method for greenhouse environments based on multi-source sensor data fusion. This invention comprehensively utilizes multi-source time-series data acquisition technology, differentiable causal fusion technology, improved WaveNet deep time-series modeling technology, and dual-domain counterfactual strategy optimization technology. By constructing a causal relationship structure between equipment actions, environmental states, and crop states, it generates causal latent variables that accurately reflect environmental changes and crop physiological responses. Based on this, a WaveNet network with enhanced reaction kernels is used to capture the high-frequency dynamic response characteristics of the greenhouse environment. Furthermore, this invention constructs counterfactual simulation mechanisms in both the environmental and crop domains to extrapolate future environmental and crop stress changes for candidate control strategies. Based on a multi-objective trade-off principle, the optimal control strategy is selected, achieving comprehensive optimization of crop safety, growth stability, and control energy consumption. This invention possesses advantages such as high prediction accuracy, fast control response, high strategy intelligence, and adaptability to complex dynamic environments, significantly improving the level of refined management in greenhouse environment control.

[0007] The intelligent prediction and control method for greenhouse environment based on multi-source sensor data fusion according to embodiments of the present invention includes:

[0008] Collect multi-source time-series data from inside the greenhouse, perform preprocessing on the multi-source time-series data, and obtain a unified multi-source time-series dataset;

[0009] A causal topological structure is established based on a multi-source time-series dataset, a differentiable causal fusion structure is constructed, intervention operations are performed on equipment action data, and a causal latent variable sequence is generated.

[0010] By using the causal latent variable sequence and the pulse sequence of equipment action data as input, a reaction kernel modeling structure based on WaveNet network is constructed to generate environmental prediction sequences and crop short-term state prediction sequences for multiple future time steps.

[0011] A crop stress time window is constructed based on environmental prediction sequences and crop status prediction sequences. The environmental deviation and crop physiological changes are accumulated over time to obtain the future crop stress degree sequence. Candidate equipment control sequences are constructed based on the executable actions of greenhouse control equipment.

[0012] An environmental counterfactual simulation domain is constructed based on the environmental prediction sequence. Counterfactual environmental simulation is performed on the candidate equipment control sequence. A crop counterfactual simulation domain is constructed based on the future crop stress degree sequence. Counterfactual crop simulation is performed to obtain environmental counterfactual results and crop counterfactual results.

[0013] A multi-objective strategy optimization mechanism is constructed based on environmental counterfactual results and crop counterfactual results. Multi-objective optimization search is performed on multiple candidate equipment control sequences, and the optimal equipment control sequence is selected as the greenhouse environment control strategy to drive the fans, sprinkler devices, heaters, wet curtains and shade curtains to perform corresponding actions.

[0014] Optionally, the multi-source time-series data includes environmental sensing data, crop status data, and equipment action data. The environmental sensing data includes air temperature, air humidity, light intensity, carbon dioxide concentration, soil moisture, and soil temperature. The crop status data includes canopy image data, multispectral reflectance data, and crop leaf temperature data. The equipment action data includes the on / off status and power parameters of fans, sprinkler systems, heaters, wet curtains, and shade curtains.

[0015] Optionally, the preprocessing of the multi-source time-series data includes:

[0016] Perform timestamp alignment, sampling interval unification, and drift calibration on environmental sensor data;

[0017] Perform image denoising, feature extraction, and data standardization on crop status data;

[0018] Perform switch status encoding and power quantization processing on the equipment action data.

[0019] Optionally, generating the causal latent variable sequence includes:

[0020] Organize equipment action data into equipment action time series, organize environmental sensor data into environmental state time series, and organize crop state data into crop state time series.

[0021] Based on the prior causal relationship between equipment action timing, environmental state timing and crop state timing, a causal topology structure is established that points from equipment action to environmental state and from environmental state to crop state. The causal topology structure is divided into equipment-to-environment causal channel and environment-to-crop causal channel. Differentiable causal weight parameters are set for different sensing dimensions in each channel.

[0022] A differentiable causal fusion structure is constructed, and multiple differentiable causal processing units are connected in series in the device-to-environment causal channel. Each differentiable causal processing unit uses weighted summation, inner product and continuous differentiable nonlinear transformation as basic operations. Residual connection units and gated causal weight update units are set between adjacent differentiable causal processing units.

[0023] Intervention processing is performed on the timing of equipment actions. In the causal channel from equipment to environment, the equipment action data at the current time step is replaced with the intervention action representation. The timing of equipment actions and environmental state after intervention is processed through a differentiable causal fusion structure to obtain the environmental causal timing.

[0024] In the environment-to-crop causal pathway, using the environmental causal time series and the crop state time series as inputs, a causal fusion process is performed on each time step in chronological order. This causal fusion process includes:

[0025] The environmental causal impact is decomposed into immediate causal impact components and delayed causal impact components. Adaptive causal weight allocation is performed for different growth stages, and time-accumulated fusion is performed on each causal component to generate causal latent variables for corresponding time steps. The causal latent variables of all time steps are combined in chronological order to form a causal latent variable sequence.

[0026] Optionally, generating environmental prediction sequences and crop short-term state prediction sequences for multiple future time steps includes:

[0027] The causal latent variable sequence is expanded in chronological order to obtain the causal latent variable time series at each time step. The equipment action data is pulse encoded in chronological order to obtain the equipment action pulse time series that characterizes the switching state and power change of the equipment.

[0028] Based on the dilated causal convolution and residual skip connection framework of the WaveNet network, a response kernel modeling structure is constructed. This structure is divided into device response-guided dilated convolution units, crop feedback-gated residual units, and multi-scale response kernel aggregation units.

[0029] Device response-guided dilated convolutional units are embedded in the dilated causal convolutional channels at each level of WaveNet. In each layer of convolutional computation, the corresponding device response kernel is called in real time to implement action pulse guidance and response weighting on the convolutional output.

[0030] Crop feedback gated residual units are arranged on the residual paths of the WaveNet network;

[0031] Multi-scale reaction core aggregation units are arranged on the jump-connection output path of the WaveNet network, and corresponding reaction core parameter sets are set for the fan spray device, heater, wet curtain and sunshade curtain respectively;

[0032] In the device response-guided dilated convolutional unit, the causal latent variable time series is used as the main sequence input of the WaveNet network, and the device action pulse time series is used as the guiding sequence input of the WaveNet network. Dilated causal convolution processing is performed on the main sequence in the causal convolutional layers at each dilation scale. Based on the position and intensity of each device action pulse in the guiding sequence, response weighting processing is performed on the corresponding dilated causal convolution output to obtain the intermediate time series feature sequence.

[0033] In the crop feedback gated residual unit, the intermediate time series feature sequence and the crop state-related feature sequence at the corresponding time step are combined in the WaveNet network residual path to generate a gated control signal. The feature sequences after gated residual processing at each expansion scale are sent to the multi-scale reaction kernel aggregation unit. Time alignment processing and cross-scale aggregation processing are performed on the feature sequences corresponding to different expansion scales and different equipment types along the jump connection path of the WaveNet network to obtain high-dimensional reaction kernel feature time series.

[0034] Channel mapping and time step decoding are performed on the high-dimensional reaction kernel feature time series to obtain environmental prediction sequences and crop short-term state prediction sequences corresponding to multiple future time steps.

[0035] Optionally, obtaining the future crop stress degree sequence and constructing a candidate equipment control sequence based on the executable actions of the greenhouse control equipment includes:

[0036] Obtain environmental prediction sequences and crop short-term state prediction sequences, combine crop variety information and growth cycle information to determine the growth stage of the target crop, and preset the corresponding crop stress time window length and the safe range parameters of environmental variables for the growth stage.

[0037] Within the crop stress time window, the environmental variables at each time step in the environmental prediction sequence are compared with the corresponding safe range parameters to determine the environmental deviation index at each time step. The physiological change characteristics at each time step in the crop state prediction sequence are extracted and processed to obtain the crop physiological change index.

[0038] Within the crop stress time window, time-cumulative processing and stage-weighted processing are performed on the environmental deviation indicators and crop physiological change indicators at each time step to generate future crop stress values ​​corresponding to each time step. The future crop stress values ​​at each time step are combined in chronological order to form a future crop stress sequence.

[0039] Based on the future crop stress degree sequence and the set of executable actions and duration range of greenhouse control equipment such as fans, sprinklers, heaters, wet curtains and shade curtains, multiple candidate equipment control sequences are constructed according to the preset time step.

[0040] Optionally, obtaining environmental counterfactual results and crop counterfactual results includes:

[0041] The environmental prediction sequence, the future crop stress degree sequence, and the candidate equipment control sequence are obtained. The environmental prediction sequence is used as the environmental baseline trajectory, and the future crop stress degree sequence is used as the crop baseline stress trajectory.

[0042] An environmental counterfactual simulation domain is constructed based on the environmental prediction sequence. Each candidate equipment control sequence is aligned with the environmental baseline trajectory according to a unified time axis. Based on the equipment response characteristics, corresponding environmental response templates are set for fans, sprinkler devices, heaters, wet curtains and sunshades respectively. The corresponding environmental response templates are superimposed one by one on the time axis for each control action in the candidate equipment control sequence to generate environmental counterfactual trajectories under the action of different candidate equipment control sequences.

[0043] Each environmental counterfactual trajectory is compared with the environmental baseline trajectory, and the environmental deviation duration, environmental deviation magnitude, and environmental fluctuation intensity indices within the simulation time range are extracted and combined to form the environmental counterfactual results of the corresponding candidate device control sequence.

[0044] A crop counterfactual simulation domain is constructed based on the future crop stress degree sequence. In the crop counterfactual simulation domain, each environmental counterfactual trajectory is paired with the crop baseline stress trajectory on a unified time axis. Stage sensitivity coefficients of crops to changes in temperature, humidity, light and carbon dioxide are preset for each growth stage. Based on the stage sensitivity coefficients, the environmental changes in the environmental counterfactual trajectory are weighted to generate the corresponding crop counterfactual stress trajectory.

[0045] Each crop counterfactual stress trajectory is compared with the crop baseline stress trajectory. The maximum stress level, cumulative stress amount, and stress recovery rate indices within the simulation time range are extracted and combined to form the crop counterfactual results of the corresponding candidate device control sequence.

[0046] Optionally, the selection of the optimal equipment control sequence as the greenhouse environment control strategy, driving the fans, sprinkler system, heater, wet curtain, and shade curtain to perform corresponding actions, includes:

[0047] Obtain environmental counterfactual results and crop counterfactual results, establish a one-to-one correspondence with candidate equipment control sequences, and use them as environmental assessment information, crop assessment information and objects to be optimized, respectively. Extend the current actual control actions of the greenhouse to form a benchmark control sequence, and generate several heuristic sequences based on the action combination rules of the greenhouse control equipment.

[0048] A multi-objective strategy optimization mechanism is constructed, and hierarchical constraints and priority rules are set. The first layer is crop safety as a hard constraint, the second layer is environmental stability as a soft constraint, and the third layer is the control sequence smoothness as an execution constraint. A dual-domain consistency verification unit is set to simultaneously verify environmental evaluation information and crop evaluation information.

[0049] Under the multi-objective strategy optimization mechanism, the baseline control sequence and several heuristic sequences are used as the initial set. For each candidate device control sequence, a neighborhood transformation guided by the reaction kernel sensitivity is performed to fine-tune the start time, action duration and power amplitude. After each transformation, the dual-domain consistency verification unit is called synchronously to complete the validity screening.

[0050] Dynamic weighted alternating sorting and elimination are performed on the candidate equipment control sequences that have passed the screening. First, any sequence that violates the first layer of constraints in crop evaluation information is eliminated. Then, the remaining sequences are subjected to non-dominated comparison and multiple rounds of alternating sorting based on environmental evaluation information and smoothness constraints, and the multi-objective compromise set is retained.

[0051] In the multi-objective compromise optimization set, the optimal equipment control sequence is determined in the order of minimum cumulative crop stress, minimum environmental deviation, and smoothness of control sequence, and the equipment is driven to perform regulation actions at the corresponding time step.

[0052] The beneficial effects of this invention are:

[0053] This invention introduces a differentiable causal fusion mechanism, enabling the explicit expression and quantification of the causal relationships among equipment actions, environmental states, and crop states. This fundamentally solves the shortcomings of existing technologies that rely solely on instantaneous environmental data and cannot identify the dynamic impact of equipment actions on the environment and their transmission to crop physiological responses. By constructing a causal latent variable sequence, this invention achieves a realistic and structured characterization of the greenhouse environment control process, allowing the control system to accurately grasp the generation mechanism of environmental changes based on data-driven principles, thus improving the overall reliability and stability of predictions.

[0054] This invention employs a reaction kernel construction structure based on the WaveNet network, which can effectively capture high-frequency dynamic environmental fluctuations caused by the actions of equipment such as fans, sprinklers, heaters, wet curtains, and shading curtains. This solves the problem that traditional prediction methods cannot accurately describe the rapid environmental changes after equipment operation. By introducing structures such as equipment response-guided dilated convolution, crop feedback-gated residuals, and multi-scale reaction kernel aggregation, this invention not only improves the temporal resolution and response sensitivity of environmental prediction but also realizes dynamic feedback of crop physiological characteristics to the prediction process. This allows the prediction results to simultaneously reflect the dual characteristics of environmental evolution and crop response, making them more closely aligned with the actual operating conditions of greenhouses.

[0055] This invention constructs environmental counterfactual simulation domains and crop counterfactual simulation domains, and designs a multi-objective strategy optimization mechanism. This allows different candidate control sequences to anticipate future environmental and crop stress changes before execution, achieving a unity of predictive regulation and adaptive strategy optimization. Under the premise of ensuring crop growth safety, it accurately selects the optimal regulation scheme that reduces environmental deviations, minimizes crop stress, and balances energy consumption and control smoothness, thereby effectively avoiding over-regulation or under-regulation. This invention improves the intelligence level of greenhouse environmental regulation, achieving integrated fusion of prediction, simulation, and strategy decision-making, resulting in more accurate prediction, more efficient regulation, and more stable crop growth. Attached Figure Description

[0056] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0057] Figure 1 The flowchart shows the intelligent prediction and control method for greenhouse environment based on multi-source sensor data fusion proposed in this invention.

[0058] Figure 2 This is a schematic diagram of the reaction kernel construction structure based on WaveNet network for the intelligent prediction and control method of greenhouse environment based on multi-source sensor data fusion proposed in this invention.

[0059] Figure 3 This is a schematic diagram of the environmental counterfactual simulation domain and the crop counterfactual simulation domain of the intelligent prediction and control method for greenhouse environment based on multi-source sensor data fusion proposed in this invention. Detailed Implementation

[0060] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0061] refer to Figure 1 , Figure 2 and Figure 3 A method for intelligent prediction and control of greenhouse environment based on multi-source sensor data fusion includes:

[0062] Collect multi-source time-series data from inside the greenhouse, perform preprocessing on the multi-source time-series data, and obtain a unified multi-source time-series dataset;

[0063] A causal topological structure is established based on a multi-source time-series dataset, a differentiable causal fusion structure is constructed, intervention operations are performed on equipment action data, and a causal latent variable sequence is generated.

[0064] By using the causal latent variable sequence and the pulse sequence of equipment action data as input, a reaction kernel modeling structure based on WaveNet network is constructed to generate environmental prediction sequences and crop short-term state prediction sequences for multiple future time steps.

[0065] A crop stress time window is constructed based on environmental prediction sequences and crop status prediction sequences. The environmental deviation and crop physiological changes are accumulated over time to obtain the future crop stress degree sequence. Candidate equipment control sequences are constructed based on the executable actions of greenhouse control equipment.

[0066] An environmental counterfactual simulation domain is constructed based on the environmental prediction sequence. Counterfactual environmental simulation is performed on the candidate equipment control sequence. A crop counterfactual simulation domain is constructed based on the future crop stress degree sequence. Counterfactual crop simulation is performed to obtain environmental counterfactual results and crop counterfactual results.

[0067] A multi-objective strategy optimization mechanism is constructed based on environmental counterfactual results and crop counterfactual results. Multi-objective optimization search is performed on multiple candidate equipment control sequences, and the optimal equipment control sequence is selected as the greenhouse environment control strategy to drive the fans, sprinkler devices, heaters, wet curtains and shade curtains to perform corresponding actions.

[0068] This invention configures a dedicated set of response kernel matrices for each of the following devices: fan, spray device, heater, wet curtain, and shading curtain. Each set has the same number of input and output channels in each dilated convolutional layer of WaveNet. The input channels are formed by merging causal latent variable channels and device impulse channels, and the output channels are kept consistent to allow for residual summation. The convolutional kernels use a shorter receptive length in shallow layers to capture fast responses, and increase with the dilation factor in deeper layers to cover long time lags. During operation, WaveNet automatically selects the kernels of the corresponding devices for parallel convolution based on the impulse channels. If multiple devices operate simultaneously at the same time step, the channel-level convolution results are linearly superimposed. All device kernels are updated independently during end-to-end training. Initially, random orthogonal or group regularization is used to avoid parameter interference, thereby enabling the network to learn the dedicated dynamic response modes of fan cooling, spray humidification, heating, wet curtain cooling and humidification, and shading.

[0069] In this embodiment, the multi-source time-series data includes environmental sensing data, crop status data, and equipment action data. The environmental sensing data includes air temperature, air humidity, light intensity, carbon dioxide concentration, soil moisture, and soil temperature. The crop status data includes canopy image data, multispectral reflectance data, and crop leaf temperature data. The equipment action data includes the on / off status and power parameters of fans, sprinkler systems, heaters, wet curtains, and shade curtains.

[0070] In this embodiment, the preprocessing of multi-source time-series data includes:

[0071] Perform timestamp alignment, sampling interval unification, and drift calibration on environmental sensor data;

[0072] Perform image denoising, feature extraction, and data standardization on crop status data;

[0073] Perform switch status encoding and power quantization processing on the equipment action data.

[0074] In this embodiment, generating the causal latent variable sequence includes:

[0075] Organize equipment action data into equipment action time series, organize environmental sensor data into environmental state time series, and organize crop state data into crop state time series.

[0076] Based on the prior causal relationship between equipment action timing, environmental state timing, and crop state timing, a causal topology structure is established that points from equipment action to environmental state and from environmental state to crop state. The causal topology structure is divided into equipment-to-environment causal channels and environment-to-crop causal channels. Differentiable causal weight parameters are set for different sensing dimensions in each channel.

[0077] A differentiable causal fusion structure is constructed, and multiple differentiable causal processing units are connected in series in the device-to-environment causal channel. Each differentiable causal processing unit uses weighted summation, inner product and continuous differentiable nonlinear transformation as basic operations. Residual connection units and gated causal weight update units are set between adjacent differentiable causal processing units.

[0078] The differentiable causal fusion structure is presented as a vertically stacked causal channel stack. Each layer is a differentiable causal processing unit. Internally, the device action, environmental state and crop state are first linearly weighted and aggregated with trainable weights. Then, the interaction strength between variables is characterized by tensor inner product. Subsequently, nonlinear mapping is achieved through differentiable activation functions. The residuals between layers are kept directly connected, so that the information at the lower level and the semantics at the higher level flow synchronously. At the same time, a gated weight update mechanism is inserted to dynamically adjust the causal weight matrix according to the current time series data, so as to ensure that the causal direction and strength adapt to the changes in the environment and converge. The entire multi-layer structure finally outputs a latent variable representation containing the progressive causal relationship between the device, environment and crop.

[0079] Intervention processing is performed on the device action timing. In the device-to-environment causal channel, the device action data at the current time step is replaced with the intervention action representation. The interventioned device action timing and environmental state timing are processed through a differentiable causal fusion structure to obtain the environmental causal timing. Specifically, the processing of the interventioned device action timing and environmental state timing through the differentiable causal fusion structure involves:

[0080] After intervention, the timing of equipment actions is processed by causal weight expansion in the causal channel from equipment to environment. Based on the equipment influence weight in the differentiable causal structure, the changes in equipment actions at each time step are weighted and propagated to form a temporal representation of the impact of equipment actions on the environmental state.

[0081] The environmental state time series is processed in the device-to-environment causal channel, and the influence representation of the device action time series is fused with the environmental state time series step by step to generate an intermediate environmental representation sequence containing the device intervention effect.

[0082] The intermediate environmental representation sequence is subjected to continuous causal channel processing. The correlation update of environmental representation at different time steps is completed through the continuous differentiable transformation unit in the differentiable causal structure, and finally an environmental causal time series reflecting the results of equipment intervention is formed.

[0083] In the environment-to-crop causal pathway, using the environmental causal time series and the crop state time series as inputs, a causal fusion process is performed on each time step in chronological order. This causal fusion process includes:

[0084] The environmental causal impact is decomposed into immediate causal impact components and delayed causal impact components. Adaptive causal weight allocation processing is performed for different growth stages, and time-accumulated fusion processing is performed on each causal component to generate causal latent variables for corresponding time steps. The causal latent variables for all time steps are then combined in chronological order to form a causal latent variable sequence. Specifically, the adaptive causal weight allocation processing for different growth stages and the time-accumulated fusion processing on each causal component are as follows:

[0085] Based on the stage identifier of the current growth stage of the crop, stage weight values ​​are set for the immediate causal influence component and the delayed causal influence component, and the stage weight values ​​and the corresponding causal influence components are weighted step by step to form a stage-weighted causal component sequence.

[0086] The causal component sequence after stage weighting is cumulatively superimposed according to the time step order, and the causal effects of different time steps are continuously accumulated on the time axis to obtain a causal cumulative sequence that reflects the cross-time relationship between equipment action and environmental change.

[0087] The causal cumulative sequence is fused and combined to merge the features of the immediate causal component and the delayed causal component at the corresponding time step, generating causal latent variables to describe the causal state at the time step, and arranging them in chronological order to form a causal latent variable sequence.

[0088] In this embodiment, generating environmental prediction sequences and crop short-term state prediction sequences for multiple future time steps includes:

[0089] The causal latent variable sequence is expanded in chronological order to obtain the causal latent variable time series at each time step. The equipment action data is pulse encoded in chronological order to obtain the equipment action pulse time series characterizing the switching state and power change of the equipment.

[0090] Based on the dilated causal convolution and residual skip connection framework of the WaveNet network, a response kernel modeling structure is constructed. This structure is divided into device response-guided dilated convolution units, crop feedback-gated residual units, and multi-scale response kernel aggregation units.

[0091] Device response-guided dilated convolutional units are embedded in the dilated causal convolutional channels at each level of WaveNet. In each layer of convolutional computation, the corresponding device response kernel is called in real time to implement action pulse guidance and response weighting on the convolutional output.

[0092] Crop feedback gated residual units are arranged on the residual paths of the WaveNet network;

[0093] Multi-scale reaction core aggregation units are arranged on the jump-connection output path of the WaveNet network, and corresponding reaction core parameter sets are set for the fan spray device, heater, wet curtain and sunshade curtain respectively;

[0094] In the device response-guided dilated convolutional unit, the causal latent variable time series is used as the main sequence input to the WaveNet network, and the device action pulse time series is used as the guiding sequence input to the WaveNet network. Dilated causal convolution processing is performed on the main sequence in the causal convolutional layers at each dilation scale. Response-weighted processing is then performed on the corresponding dilated causal convolution outputs based on the position and intensity of each device action pulse in the guiding sequence to obtain the intermediate temporal feature sequence, where:

[0095] Dilated causal convolution is performed on the main sequence in the causal convolutional layers at each dilation scale, specifically as follows:

[0096] At each expansion scale, the causal latent variable time series is processed by a sliding window of causal direction on a time step basis to ensure that the convolution of the current time step only uses information from the historical time steps.

[0097] Based on the preset expansion interval, the corresponding time step features are selected as effective convolution inputs within the convolution window to form convolution domains with different receptive ranges.

[0098] The convolution results obtained at each dilation scale are arranged in chronological order to form the main sequence convolution feature output under the dilation scale;

[0099] Based on the position and intensity of each device action pulse in the guiding sequence, response weighting is performed on the corresponding dilated causal convolution output, specifically as follows:

[0100] Extract the action trigger position and action intensity from the device action pulse signal at each time step in the guidance sequence to form the guidance weight for the corresponding time step;

[0101] The guiding weights are weighted step-by-step with the output of the dilated causal convolution at the same time step to enhance the sensitivity of the convolution output to the moment of device action triggering.

[0102] The weighted convolutional outputs at each time step are integrated according to the dilation scale to form an intermediate temporal feature sequence containing device response features;

[0103] In the crop feedback gated residual unit, the intermediate time series feature sequence and the crop state-related feature sequence at the corresponding time step are combined in the WaveNet network residual path to generate a gated control signal. The feature sequences after gated residual processing at each expansion scale are sent to the multi-scale reaction kernel aggregation unit. Time alignment processing and cross-scale aggregation processing are performed on the feature sequences corresponding to different expansion scales and different equipment types along the jump connection path of the WaveNet network to obtain high-dimensional reaction kernel feature time series.

[0104] Channel mapping and time-step decoding are performed on the high-dimensional response kernel feature time series to obtain environmental prediction sequences and crop short-term state prediction sequences corresponding to multiple future time steps. Specifically, the channel mapping and time-step decoding processes on the high-dimensional response kernel feature time series are as follows:

[0105] The high-dimensional reaction kernel feature time series is linearly mapped according to the channel dimension, and the features of different convolution channels are mapped to the preset dimensions corresponding to the environment prediction channel and the crop state prediction channel.

[0106] The features mapped by the channel are decoded in the time dimension. The features of consecutive time steps are input into the multi-time step decoding unit in sequence, and the predicted feature representations of each future time step are generated by time step alignment.

[0107] The predicted features of each future time step are combined by category to form the future environment prediction sequence and the future crop short-term state prediction sequence.

[0108] In this embodiment, obtaining the future crop stress degree sequence and constructing a candidate equipment control sequence based on the executable actions of the greenhouse control equipment includes:

[0109] The environmental prediction sequence and crop short-term state prediction sequence are obtained. Combined with crop variety information and growth cycle information, the growth stage of the target crop is determined. The corresponding crop stress time window length and safe range parameters of environmental variables are preset for the growth stage. The environmental variables include air temperature, air humidity, light intensity, carbon dioxide concentration, soil temperature and soil moisture.

[0110] Within the crop stress time window, environmental variables at each time step in the environmental prediction sequence are compared with their corresponding safe range parameters to determine the environmental deviation index at each time step. Physiological change characteristics at each time step in the crop state prediction sequence are extracted to obtain crop physiological change indices. Specifically, determining the environmental deviation index at each time step involves:

[0111] Within the crop stress time window, the air temperature, air humidity, light intensity, carbon dioxide concentration, soil temperature and soil humidity at each time step are compared with the corresponding upper and lower limits of the safe range to determine whether each environmental variable at the time step exceeds the safe range.

[0112] At each time step when an environmental variable exceeds the safe range, the absolute deviation is calculated for the extent of the excess, and the environmental variable deviation value at that time step is obtained.

[0113] The deviation values ​​of each environmental variable are combined according to the time step to form the environmental deviation index of the time step;

[0114] Within the crop stress time window, time-cumulative processing and stage-weighted processing are performed on environmental deviation indicators and crop physiological change indicators at each time step to generate future crop stress values ​​corresponding to each time step. These future crop stress values ​​are then combined in chronological order to form a future crop stress sequence. Specifically, the time-cumulative processing and stage-weighted processing of environmental deviation indicators and crop physiological change indicators at each time step are as follows:

[0115] Within the crop stress time window, the environmental deviation indicators and crop physiological change indicators at each time step are cumulatively superimposed in chronological order, so that the cumulative deviation of the previous time step and the deviation of the next time step are continuously synthesized to form a time cumulative deviation sequence.

[0116] Based on the growth stage of the target crop, the corresponding stage weights are selected from the stage weight table, and the deviations of each time step in the cumulative deviation sequence are weighted according to the stage weights to form a stage-weighted deviation sequence.

[0117] The stage-weighted deviation sequence is processed by feature combination at each time step. The environmental deviation component and the crop physiological change component at each time step are combined to generate the future crop stress value at the corresponding time step, and arranged in chronological order to form the future crop stress sequence.

[0118] Based on the future crop stress degree sequence and the set of executable actions and duration range of greenhouse control equipment such as fans, sprinklers, heaters, wet curtains and shade curtains, multiple candidate equipment control sequences are constructed according to a preset time step, where the preset time step is a fixed time interval set according to the rate of change of the greenhouse environment, with a value of 5 minutes.

[0119] In this embodiment, obtaining environmental counterfactual results and crop counterfactual results includes:

[0120] The environmental prediction sequence, the future crop stress degree sequence, and the candidate equipment control sequence are obtained. The environmental prediction sequence is used as the environmental baseline trajectory, and the future crop stress degree sequence is used as the crop baseline stress trajectory.

[0121] An environmental counterfactual simulation domain is constructed based on environmental prediction sequences. Each candidate device control sequence is aligned with the environmental baseline trajectory along a unified time axis. Based on device response characteristics, corresponding environmental response templates are set for fans, sprinkler systems, heaters, wet curtains, and sunshades. The corresponding environmental response templates are then superimposed one by one on the time axis for each control action in the candidate device control sequences, generating environmental counterfactual trajectories under different candidate device control sequences. Specifically, the construction of the environmental counterfactual simulation domain based on environmental prediction sequences involves:

[0122] A continuous baseline environmental trajectory is established for the environmental prediction sequence according to time steps, and the environmental prediction values ​​of all future time steps are arranged to form a baseline time axis of the counterfactual domain.

[0123] An independent environmental variable channel is established for each environmental variable on the baseline time axis. The predicted values ​​of air temperature, air humidity, light intensity, carbon dioxide concentration, soil temperature and soil moisture are mapped to the corresponding channels to form a multi-channel environmental baseline matrix.

[0124] By setting a uniform time step resolution and simulation duration on a multi-channel environmental baseline matrix, an environmental counterfactual simulation domain with continuous time structure, capable of accepting the influence of control sequences and performing sequence inference is formed.

[0125] Each environmental counterfactual trajectory is compared with the environmental baseline trajectory, and the environmental deviation duration, environmental deviation magnitude, and environmental fluctuation intensity indices within the simulation time range are extracted and combined to form the environmental counterfactual results of the corresponding candidate device control sequence.

[0126] A crop counterfactual simulation domain is constructed based on the future crop stress degree sequence. In this domain, each environmental counterfactual trajectory is paired with the crop baseline stress trajectory on a unified time axis. Stage sensitivity coefficients for changes in temperature, humidity, light, and carbon dioxide are preset for each growth stage. Environmental changes in the environmental counterfactual trajectory are weighted according to these stage sensitivity coefficients to generate the corresponding crop counterfactual stress trajectory. Specifically, the construction of the crop counterfactual simulation domain based on the future crop stress degree sequence is as follows:

[0127] Establish a continuous stress baseline time axis by arranging the future crop stress degree sequence in chronological order, and use the time axis as the simulation reference framework for the crop counterfactual simulation domain.

[0128] In the simulation reference framework, independent stress channels are established for temperature stress, humidity stress, light stress and carbon dioxide stress respectively. The corresponding stress components in the future crop stress degree sequence are mapped to each channel according to time steps to form a multi-channel crop benchmark stress matrix.

[0129] A unified time step resolution, simulation duration range, and stage-sensitive parameter position markers for stress inference are set on the multi-channel crop baseline stress matrix, so that the matrix has a temporal structure that can perform subsequent counterfactual inferences, thus forming a crop counterfactual simulation domain for crop stress inference.

[0130] Each crop counterfactual stress trajectory is compared with the crop baseline stress trajectory. The maximum stress level, cumulative stress amount, and stress recovery rate indices within the simulation time range are extracted and combined to form the crop counterfactual results of the corresponding candidate device control sequence.

[0131] In this embodiment, selecting the optimal equipment control sequence as the greenhouse environment control strategy to drive the fan, sprinkler system, heater, wet curtain, and shade curtain to perform corresponding actions includes:

[0132] Obtain environmental counterfactual results and crop counterfactual results, establish a one-to-one correspondence with candidate equipment control sequences, and use them as environmental assessment information, crop assessment information, and the object to be optimized, respectively. Extend the current actual control actions of the greenhouse to form a baseline control sequence. Generate several heuristic sequences based on the action combination rules of the greenhouse control equipment. Specifically, the generation of several heuristic sequences based on the action combination rules of the greenhouse control equipment is as follows:

[0133] Based on the action types of fans, sprinkler devices, heaters, wet curtains and sunshades, the executable actions of each device are classified according to opening, closing, power level, opening and closing range and operating mode. Actions that can be executed in parallel within the same time step are arranged according to their combinable relationships to form a basic action combination list.

[0134] In the list of basic action combinations, the execution time of each combination is arranged according to the preset time step, and the action combinations are sequentially distributed in multiple consecutive time steps to generate temporal combinations with different action orders and action intervals, thereby forming a temporal heuristic sequence.

[0135] A multi-objective strategy optimization mechanism is constructed, and hierarchical constraints and priority rules are set. The first layer is crop safety as a hard constraint, the second layer is environmental stability as a soft constraint, and the third layer is the control sequence smoothness as an execution constraint. A dual-domain consistency verification unit is set to simultaneously verify environmental evaluation information and crop evaluation information.

[0136] Under the multi-objective strategy optimization mechanism, a baseline control sequence and several heuristic sequences are used as the initial set. For each candidate device control sequence, a reaction kernel sensitivity-guided neighborhood transformation is performed, with limited fine-tuning of the start-up time, action duration, and power amplitude. After each transformation, a dual-domain consistency verification unit is synchronously invoked to complete the validity screening. Specifically, the reaction kernel sensitivity-guided neighborhood transformation for each candidate device control sequence is as follows:

[0137] Based on the sensitivity of the reaction nucleus to temperature, humidity, light and carbon dioxide variables, the start time of each device action in the candidate device control sequence is slightly shifted forward or backward to form neighborhood variant sequences with different trigger times.

[0138] Based on the sensitivity change trend corresponding to the action intensity of each device in the reaction kernel characteristics, the action duration in the candidate device control sequence is subject to restricted addition and subtraction processing, so that the action duration forms extended or shortened neighborhood variants within the allowable range;

[0139] By combining the sensitive positions of the action amplitude in the reaction kernel features to the degree of response to environmental variables, the power level or opening and closing amplitude in the candidate device control sequence is fine-tuned to generate a power-adjustable neighborhood variant sequence without exceeding the device action limit.

[0140] Dynamic weighted alternating sorting and elimination are performed on the candidate equipment control sequences that have passed the screening. First, any sequence that violates the first layer of constraints in crop evaluation information is eliminated. Then, the remaining sequences are subjected to non-dominated comparison and multiple rounds of alternating sorting based on environmental evaluation information and smoothness constraints, and the multi-objective compromise set is retained.

[0141] In the multi-objective compromise optimization set, the optimal equipment control sequence is determined in the order of minimum cumulative crop stress, minimum environmental deviation, and smoothness of control sequence, and the equipment is driven to perform regulation actions at the corresponding time step.

[0142] Example 1:

[0143] To verify the feasibility of this invention in practice, it was applied to a modern multi-span intelligent greenhouse with an area of ​​approximately 520 square meters. The greenhouse cultivated cherry tomatoes and was equipped with conventional environmental control equipment such as fans, sprinklers, heaters, wet curtains, and shade curtains. Sensors for air temperature, air humidity, light intensity, carbon dioxide concentration, soil temperature, and soil moisture were installed inside the greenhouse. Multispectral cameras and leaf temperature infrared sensors were also deployed to collect short-term crop status data. During the high-temperature period in July 2025, the greenhouse repeatedly experienced rapid environmental fluctuations. For example, drastic changes in light intensity at midday caused fluctuations in canopy temperature, and rapid increases in humidity after the sprinkler system was activated led to increased disease risk. Traditional threshold control methods could not respond to these dynamic changes in a timely manner, easily causing mild heat stress or excessive humidity accumulation in the crops. To address these shortcomings, this invention provides an intelligent predictive control method based on multi-source time-series data fusion, which was validated through 20 consecutive days of actual operation in the greenhouse.

[0144] In the application of this invention, firstly, multi-source sensing devices within the greenhouse continuously collect data on air temperature, humidity, light intensity, carbon dioxide, soil temperature and humidity, as well as multispectral images and leaf temperature. This data undergoes time synchronization, noise reduction, and feature generation to construct a unified multi-source time-series dataset. Subsequently, based on a differentiable causal fusion structure, the impact of equipment actions on the environment and the influence of environmental changes on short-term crop physiological responses are causally fused. For example, after a fan starts with a pulse input, the causal fusion structure can identify the rate difference in temperature drop caused by the fan action, no longer treating the equipment action as a common feature, but rather modeling its causal impact as an intervention. This allows the environmental state to accurately reflect the trajectory of the equipment's action in predictions.

[0145] The causal latent variable sequence, processed by differentiable causal fusion, is input into a response kernel construction structure based on the WaveNet network. This structure includes equipment response-guided dilated convolution, crop feedback-gated residuals, and multi-scale response kernel aggregation, enabling accurate capture of high-frequency dynamics within the greenhouse caused by equipment actions. In actual operation, waveforms showing rapid humidity increases after the sprinkler system is activated, slow temperature increases during heater reheating, and step-like changes in light intensity after partial closure of the shading curtain can all be modeled in the response kernel structure using convolutions at different scales. Through this structure, this invention can generate environmental prediction sequences and short-term crop state prediction sequences within the range of 20 minutes to 1 hour, giving the control system a forward-looking understanding of future environmental evolution.

[0146] This invention provides a comprehensive assessment of future environmental changes and crop stress. Based on predictions for different time periods, a crop stress time window is constructed. Accumulated analysis is performed on future temperature deviations, humidity deviations, leaf temperature changes, and multispectral reflectance characteristics to obtain a future crop stress degree sequence. Simultaneously, based on the executable actions of fans, sprinklers, heaters, evaporative cooling pads, and shading curtains, candidate device control sequences are constructed.

[0147] Furthermore, this invention constructs counterfactual simulation mechanisms in both the environmental and crop domains, performing environmental and crop counterfactual inferences on different candidate device control sequences. Environmental counterfactual simulation can predict future environmental deviations without actually executing device actions. For example, if temperatures surge on July 15th, the system predicts that a 2-minute delay in fan startup would cause the greenhouse temperature to exceed the safe range by 2.1°C within the next 15 minutes, thus automatically eliminating this option. Crop counterfactual simulation infers future crop temperature changes, multispectral characteristics, and leaf temperature trends to determine the impact of different device strategies on crop stress levels. Through dual-domain counterfactual simulation, this invention can avoid device action schemes that could potentially damage crop physiology.

[0148] The final equipment actions are selected by a multi-objective strategy optimization mechanism. This mechanism considers factors such as crop safety, environmental stability, and the smoothness of equipment actions, and optimizes the control sequence of candidate equipment through a neighborhood search guided by reaction kernel sensitivity. In actual operation, this invention can select a regulation strategy that can suppress rapid environmental fluctuations, reduce energy consumption, and avoid crop stress. For example, when sunlight intensifies at noon on sunny days, this invention uses a combination strategy of partially closing the shading curtains and operating low-speed fans to stabilize the temperature between 27.5℃ and 29.1℃, rather than continuously starting fans and sprinkler systems as in traditional control methods, which leads to excessively high humidity.

[0149] Table 1 Comparison of the effects of greenhouse environmental control

[0150]

[0151] The data in Table 1 clearly demonstrates that this invention significantly outperforms traditional control methods in terms of environmental prediction accuracy. The temperature prediction error decreased from 1.2℃ to 0.5℃, and the humidity prediction error decreased from 5.8%RH to 2.3%RH. This indicates that the causal fusion of the microstructure and the WaveNet-based reaction kernel modeling can more accurately characterize the high-frequency dynamic changes caused by equipment actions, enabling the system to identify future environmental trends in advance. This forward-looking predictive capability effectively avoids the problem of excessive environmental fluctuations caused by delayed response in traditional systems.

[0152] In terms of crop physiological state, this invention also demonstrates significant advantages. Under traditional control methods, the highest crop leaf temperature reaches 34.6℃, while this invention controls it at 32.8℃ and reduces the crop stress index from 0.42 to 0.27. This indicates that the system can assess potential stress risks through crop counterfactual simulation before implementing control actions, thereby selecting a more favorable equipment combination. By combining crop short-term state prediction sequences with stress time windows, this invention can suppress crop heat stress and effectively improve crop growth stability under high temperature and rapid light change conditions.

[0153] This invention also demonstrates outstanding performance in energy utilization and equipment operation optimization. The number of start-ups and shutdowns of the fan and sprinkler system decreased by approximately 39% and 41%, respectively, indicating that through dual-domain counterfactual inference and multi-objective strategy optimization, the system can reduce unnecessary frequent operations and achieve a smoother, more energy-efficient control mode. Total energy consumption decreased from 24.7 kWh / day to 21.3 kWh / day, saving approximately 14% of energy. This invention significantly outperforms traditional systems in terms of prediction accuracy, crop physiological protection, and equipment operating efficiency, fully validating its application value and feasibility in actual greenhouse environments.

[0154] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for intelligent prediction and control of greenhouse environment based on multi-source sensor data fusion, characterized in that, include: Collect multi-source time-series data from inside the greenhouse, perform preprocessing on the multi-source time-series data, and obtain a unified multi-source time-series dataset; A causal topological structure is established based on a multi-source time-series dataset, a differentiable causal fusion structure is constructed, intervention operations are performed on equipment action data, and a causal latent variable sequence is generated. By using the causal latent variable sequence and the pulse sequence of equipment action data as input, a reaction kernel modeling structure based on WaveNet network is constructed to generate environmental prediction sequences and crop short-term state prediction sequences for multiple future time steps. A crop stress time window is constructed based on environmental prediction sequences and crop status prediction sequences. The environmental deviation and crop physiological changes are accumulated over time to obtain the future crop stress degree sequence. Candidate equipment control sequences are constructed based on the executable actions of greenhouse control equipment. An environmental counterfactual simulation domain is constructed based on the environmental prediction sequence. Counterfactual environmental simulation is performed on the candidate equipment control sequence. A crop counterfactual simulation domain is constructed based on the future crop stress degree sequence. Counterfactual crop simulation is performed to obtain environmental counterfactual results and crop counterfactual results. A multi-objective strategy optimization mechanism is constructed based on environmental counterfactual results and crop counterfactual results. Multi-objective optimization search is performed on multiple candidate equipment control sequences, and the optimal equipment control sequence is selected as the greenhouse environment control strategy to drive the fans, sprinkler devices, heaters, wet curtains and shade curtains to perform corresponding actions. The generation of the causal latent variable sequence includes: Organize equipment action data into equipment action time series, organize environmental sensor data into environmental state time series, and organize crop state data into crop state time series. Based on the prior causal relationship between equipment action timing, environmental state timing and crop state timing, a causal topology structure is established that points from equipment action to environmental state and from environmental state to crop state. The causal topology structure is divided into equipment-to-environment causal channel and environment-to-crop causal channel. Differentiable causal weight parameters are set for different sensing dimensions in each channel. A differentiable causal fusion structure is constructed, and multiple differentiable causal processing units are connected in series in the device-to-environment causal channel. Each differentiable causal processing unit uses weighted summation, inner product and continuous differentiable nonlinear transformation as basic operations. Residual connection units and gated causal weight update units are set between adjacent differentiable causal processing units. Intervention processing is performed on the timing of equipment actions. In the causal channel from equipment to environment, the equipment action data at the current time step is replaced with the intervention action representation. The timing of equipment actions and environmental state after intervention is processed through a differentiable causal fusion structure to obtain the environmental causal timing. In the environment-to-crop causal pathway, using the environmental causal time series and the crop state time series as inputs, a causal fusion process is performed on each time step in chronological order. This causal fusion process includes: The environmental causal impact is decomposed into immediate causal impact components and delayed causal impact components. Adaptive causal weight allocation processing is performed for different growth stages, and time accumulation fusion processing is performed on each causal component to generate causal latent variables for corresponding time steps. The causal latent variables of all time steps are combined in time order to form a causal latent variable sequence. The generation of environmental prediction sequences and crop short-term state prediction sequences for multiple future time steps includes: The causal latent variable sequence is expanded in chronological order to obtain the causal latent variable time series at each time step. The equipment action data is pulse encoded in chronological order to obtain the equipment action pulse time series characterizing the switching state and power change of the equipment. Based on the dilated causal convolution and residual skip connection framework of the WaveNet network, a response kernel modeling structure is constructed. This structure is divided into device response-guided dilated convolution units, crop feedback-gated residual units, and multi-scale response kernel aggregation units. Device response-guided dilated convolutional units are embedded in the dilated causal convolutional channels at each level of WaveNet. In each layer of convolutional computation, the corresponding device response kernel is called in real time to implement action pulse guidance and response weighting on the convolutional output. Crop feedback gated residual units are arranged on the residual paths of the WaveNet network; Multi-scale reaction core aggregation units are arranged on the jump-connection output path of the WaveNet network, and corresponding reaction core parameter sets are set for the fan spray device, heater, wet curtain and sunshade curtain respectively; In the device response-guided dilated convolutional unit, the causal latent variable time series is used as the main sequence input of the WaveNet network, and the device action pulse time series is used as the guiding sequence input of the WaveNet network. Dilated causal convolution processing is performed on the main sequence in the causal convolutional layers at each dilation scale. Based on the position and intensity of each device action pulse in the guiding sequence, response weighting processing is performed on the corresponding dilated causal convolution output to obtain the intermediate time series feature sequence. In the crop feedback gated residual unit, the intermediate time series feature sequence and the crop state-related feature sequence at the corresponding time step are combined in the WaveNet network residual path to generate a gated control signal. The feature sequences after gated residual processing at each expansion scale are sent to the multi-scale reaction kernel aggregation unit. Time alignment processing and cross-scale aggregation processing are performed on the feature sequences corresponding to different expansion scales and different equipment types along the jump connection path of the WaveNet network to obtain high-dimensional reaction kernel feature time series. Channel mapping and time step decoding are performed on the high-dimensional reaction kernel feature time series to obtain environmental prediction sequences and crop short-term state prediction sequences corresponding to multiple future time steps; The obtained environmental counterfactual results and crop counterfactual results include: The environmental prediction sequence, the future crop stress degree sequence, and the candidate equipment control sequence are obtained. The environmental prediction sequence is used as the environmental baseline trajectory, and the future crop stress degree sequence is used as the crop baseline stress trajectory. An environmental counterfactual simulation domain is constructed based on the environmental prediction sequence. Each candidate equipment control sequence is aligned with the environmental baseline trajectory according to a unified time axis. Based on the equipment response characteristics, corresponding environmental response templates are set for fans, sprinkler devices, heaters, wet curtains and sunshades respectively. The corresponding environmental response templates are superimposed one by one on the time axis for each control action in the candidate equipment control sequence to generate environmental counterfactual trajectories under the action of different candidate equipment control sequences. Each environmental counterfactual trajectory is compared with the environmental baseline trajectory, and the environmental deviation duration, environmental deviation magnitude, and environmental fluctuation intensity indices within the simulation time range are extracted and combined to form the environmental counterfactual results of the corresponding candidate device control sequence. A crop counterfactual simulation domain is constructed based on the future crop stress degree sequence. In the crop counterfactual simulation domain, each environmental counterfactual trajectory is paired with the crop baseline stress trajectory on a unified time axis. Stage sensitivity coefficients of crops to changes in temperature, humidity, light and carbon dioxide are preset for each growth stage. Based on the stage sensitivity coefficients, the environmental changes in the environmental counterfactual trajectory are weighted to generate the corresponding crop counterfactual stress trajectory. Each crop counterfactual stress trajectory is compared with the crop baseline stress trajectory. The maximum stress level, cumulative stress amount, and stress recovery rate indices within the simulation time range are extracted and combined to form the crop counterfactual results of the corresponding candidate device control sequence.

2. The intelligent prediction and control method for greenhouse environment based on multi-source sensor data fusion according to claim 1, characterized in that, The multi-source time-series data includes environmental sensor data, crop status data, and equipment action data. The environmental sensor data includes air temperature, air humidity, light intensity, carbon dioxide concentration, soil moisture, and soil temperature. The crop status data includes canopy image data, multispectral reflectance data, and crop leaf temperature data. The equipment action data includes the on / off status and power parameters of fans, sprinkler systems, heaters, wet curtains, and shade curtains.

3. The intelligent prediction and control method for greenhouse environment based on multi-source sensor data fusion according to claim 1, characterized in that, The preprocessing of multi-source time-series data includes: Perform timestamp alignment, sampling interval unification, and drift calibration on environmental sensor data; Perform image denoising, feature extraction, and data standardization on crop status data; Perform switch status encoding and power quantization processing on the equipment action data.

4. The intelligent prediction and control method for greenhouse environment based on multi-source sensor data fusion according to claim 1, characterized in that, The process involves obtaining a future crop stress sequence and constructing a candidate equipment control sequence based on the executable actions of the greenhouse control equipment, including: Obtain environmental prediction sequences and crop short-term state prediction sequences, combine crop variety information and growth cycle information to determine the growth stage of the target crop, and preset the corresponding crop stress time window length and the safe range parameters of environmental variables for the growth stage. Within the crop stress time window, the environmental variables at each time step in the environmental prediction sequence are compared with the corresponding safe range parameters to determine the environmental deviation index at each time step. The physiological change characteristics at each time step in the crop state prediction sequence are extracted and processed to obtain the crop physiological change index. Within the crop stress time window, time-cumulative processing and stage-weighted processing are performed on the environmental deviation indicators and crop physiological change indicators at each time step to generate future crop stress values ​​corresponding to each time step. The future crop stress values ​​at each time step are combined in chronological order to form a future crop stress sequence. Based on the future crop stress degree sequence and the set of executable actions and duration range of greenhouse control equipment such as fans, sprinklers, heaters, wet curtains and shade curtains, multiple candidate equipment control sequences are constructed according to the preset time step.

5. The intelligent prediction and control method for greenhouse environment based on multi-source sensor data fusion according to claim 1, characterized in that, The selection of the optimal equipment control sequence as the greenhouse environment control strategy drives the fans, sprinkler system, heater, wet curtain, and shade curtain to perform corresponding actions, including: Obtain environmental counterfactual results and crop counterfactual results, establish a one-to-one correspondence with candidate equipment control sequences, and use them as environmental assessment information, crop assessment information and objects to be optimized, respectively. Extend the current actual control actions of the greenhouse to form a benchmark control sequence, and generate several heuristic sequences based on the action combination rules of the greenhouse control equipment. A multi-objective strategy optimization mechanism is constructed, and hierarchical constraints and priority rules are set. The first layer is crop safety as a hard constraint, the second layer is environmental stability as a soft constraint, and the third layer is the control sequence smoothness as an execution constraint. A dual-domain consistency verification unit is set to simultaneously verify environmental evaluation information and crop evaluation information. Under the multi-objective strategy optimization mechanism, the baseline control sequence and several heuristic sequences are used as the initial set. For each candidate device control sequence, a neighborhood transformation guided by the reaction kernel sensitivity is performed to fine-tune the start time, action duration and power amplitude. After each transformation, the dual-domain consistency verification unit is called synchronously to complete the validity screening. Dynamic weighted alternating sorting and elimination are performed on the candidate equipment control sequences that have passed the screening. First, any sequence that violates the first layer of constraints in crop evaluation information is eliminated. Then, the remaining sequences are subjected to non-dominated comparison and multiple rounds of alternating sorting based on environmental evaluation information and smoothness constraints, and the multi-objective compromise set is retained. In the multi-objective compromise optimization set, the optimal equipment control sequence is determined in the order of minimum cumulative crop stress, minimum environmental deviation, and smoothness of control sequence, and the equipment is driven to perform regulation actions at the corresponding time step.

Images