Methods and systems for predicting steady-state carbon dioxide concentration in photosynthetic leaf chambers

Abstract

This invention discloses a method and system for predicting steady-state carbon dioxide concentration in photosynthetic leaf chambers. The method includes: collecting dynamic time-series data of carbon dioxide concentration in the leaf chambers, including sequences of decreasing and increasing carbon dioxide concentrations and environmental data, and preprocessing the data; processing the data using a pre-trained bi-branch time-series coding network model to obtain a predicted steady-state concentration value and prediction uncertainty, and comparing it with a preset uncertainty threshold; determining the prediction reliability based on the comparison result, outputting the predicted value if reliable, and continuing to collect data and re-predict if unreliable. The system includes a data acquisition and preprocessing module, a prediction and comparison module, and a judgment and control module. This invention, through a bi-branch time-series coding network and uncertainty estimation mechanism, can predict steady-state values ​​in advance during dynamic changes in carbon dioxide concentration, significantly shortening measurement waiting time, improving measurement efficiency, and ensuring prediction accuracy and reliability.

Description

Technical Field

[0001] This invention relates to the field of plant physiological monitoring and gas concentration analysis technology, and in particular to a method and system for predicting the steady-state concentration of carbon dioxide in photosynthetic leaf chambers. Background Technology

[0002] Plant photosynthesis is a core process for crop growth, carbon cycling, and energy and matter conversion in ecosystems. Its essence is the absorption of carbon dioxide (CO2) and the synthesis of organic matter. Accurate monitoring of CO2 concentration within the leaf chamber is crucial for calculating the CO2 assimilation rate during photosynthesis, thus significantly contributing to understanding plant photosynthetic physiology and responses to environmental stress. Gas exchange methods, which measure the CO2 concentration difference between the inlet and outlet of the leaf chamber and combine this with gas flow rate to achieve quantitative analysis of gas flux, have become an important tool for measuring plant photosynthesis.

[0003] However, when using the gas exchange method, because the plant leaves are placed in closed leaf chambers, the gas environment inside the leaf chamber differs from the external environment. After each gas path switch or environmental disturbance, the CO2 concentration needs to undergo a dynamic adjustment process before reaching a stable state again. This process typically lasts for several minutes and is significantly limited by multiple factors, including gas hybrid dynamics, sensor response characteristics, and the plant's own physiological changes, severely restricting the efficiency of real-time monitoring. Therefore, how to accurately predict steady-state CO2 concentration based on dynamic data has become a key issue in improving the system's response speed and measurement efficiency.

[0004] Existing CO2 concentration prediction methods are mostly based on deep learning techniques, using fixed-length time series windows as input to predict future concentration values ​​from historical concentration data. While these methods achieve modeling of dynamic processes to some extent, the fixed window length often struggles to adapt to changes in different environmental conditions and dynamic stages, easily introducing redundant information or losing key features, resulting in limited prediction accuracy and low computational efficiency. Furthermore, existing methods generally lack quantitative assessments of the reliability of prediction results, failing to provide a basis for confidence in terminating data collection or adopting model outputs in practical applications.

[0005] In recent years, uncertainty estimation techniques have gained increasing attention in the field of time series forecasting. By quantifying the confidence level of the model output, it provides a reference for the reliability of the prediction results. Existing research attempts to introduce uncertainty perception mechanisms into dynamic system modeling to improve the model's adaptability and decision-making safety in changing environments. However, in CO2 concentration prediction tasks, how to integrate time series feature extraction and uncertainty estimation to achieve adaptive selection of the input window and reliability assessment of the prediction results remains a technical challenge that has not yet been fully resolved.

[0006] Therefore, in order to address the problems of rigid window selection and unclear reliability of prediction results in existing CO2 concentration prediction methods, there is an urgent need to develop an adaptive prediction method that can integrate dynamic feature perception and uncertainty assessment, so as to achieve rapid, accurate and reliable prediction of CO2 steady-state concentration under complex changing conditions. Summary of the Invention

[0007] Therefore, the technical problem to be solved by the present invention is to overcome the problems in the prior art where the steady-state waiting time is too long due to gas path switching in the single sensor photosynthesis measurement system, and the existing prediction methods cannot quantify the reliability of the prediction results, thus making it difficult to autonomously decide to terminate data acquisition in advance while ensuring accuracy.

[0008] To address the aforementioned technical problems, this invention provides a method for predicting the steady-state concentration of carbon dioxide in photosynthetic leaf chambers, comprising the following steps:

[0009] S1: Collect dynamic time-series data of carbon dioxide concentration in leaf chambers and environmental parameters during plant photosynthesis measurement. The dynamic carbon dioxide concentration time-series data includes concentration decrease sequence and increase sequence generated by periodic gas path switching, as well as environmental data. The dynamic time-series data is preprocessed.

[0010] S2: The pre-trained dual-branch temporal coding network model is used to process the pre-processed dynamic time series data to obtain the steady-state concentration prediction value corresponding to the current sequence and the prediction uncertainty of the steady-state concentration prediction value; the prediction uncertainty is compared with the uncertainty threshold preset for the current sequence type.

[0011] S3: When the prediction uncertainty is lower than the preset uncertainty threshold for the current sequence type, the prediction result is determined to be reliable, and the steady-state concentration prediction value is output as the final prediction value; when the prediction uncertainty is higher than or equal to the preset uncertainty threshold for the current sequence type, the prediction result is determined to be unreliable, and subsequent time series data is collected, returning to step S2 for a new round of prediction and determination.

[0012] In one embodiment of the present invention, step S2, the method for processing preprocessed dynamic time-series data using a pre-trained dual-branch temporal coding network model, includes the following steps:

[0013] S2.1: Set up two parallel encoder branches, input the falling segment sequence in the dynamic time series data into the first encoder branch, and input the rising segment sequence into the second encoder branch, and process them independently respectively;

[0014] S2.2: In each encoder branch, the input sequence is sequentially processed by differential channel enhancement, local convolution feature extraction, and linear channel fusion, so that each branch outputs high-order temporal features that can characterize the dynamic characteristics of its corresponding sequence.

[0015] S2.3: The high-order temporal features output by the first encoder branch and the second encoder branch are concatenated along the feature dimension to obtain the fused global feature representation;

[0016] S2.4: Perform a fully connected mapping on the fused global feature representation to obtain the predicted steady-state concentration value of the falling segment, the prediction uncertainty of the falling segment, the predicted steady-state concentration value of the rising segment, and the prediction uncertainty of the rising segment.

[0017] In one embodiment of the present invention, in step S2.2, the method of sequentially performing differential channel enhancement, local convolution feature extraction, and linear channel fusion processing on the input sequence in each encoder branch is as follows:

[0018] Differential channel enhancement is performed on the input multivariate time series data to simultaneously extract the absolute state and the trend of change;

[0019] One-dimensional convolution and non-linear activation operations are performed on the enhanced original and differential sequences to extract cross-channel joint features;

[0020] The cross-channel joint features are fully connected and mapped to compress the temporal dimension and capture long-range dependencies, resulting in dimensionally compressed temporal features.

[0021] In one embodiment of the present invention, step S1 involves preprocessing the dynamic time-series data by normalizing the collected dynamic time-series data and linearly mapping each feature in the dynamic time-series data to the [0,1] interval. The normalization formula is as follows:

[0022] ,

[0023] in, These are the normalized eigenvalues. These are the original eigenvalues. and These are the minimum and maximum values ​​of this feature in dynamic time-series data, respectively.

[0024] In one embodiment of the present invention, step S1, the preprocessing of the dynamic time-series data further includes data augmentation, specifically as follows:

[0025] Randomly extract the descending segment sequence to obtain the descending segment input sequence;

[0026] Locate the position in the rising segment sequence that matches the carbon dioxide concentration value at the end of the falling segment input sequence, and randomly extract from that position to obtain the rising segment input sequence;

[0027] The rising segment input sequence is selectively sampled using a random mask. When the random mask indicates a first state, the rising segment input sequence is replaced with padding values ​​and does not participate in model training. When the random mask indicates a second state, the rising segment input sequence is retained for model training.

[0028] In one embodiment of the present invention, in step S2, the dual-branch temporal coding network model is trained using a loss function based on Gaussian likelihood. The expression is:

[0029] ,

[0030] in, This represents the true steady-state concentration value. The steady-state concentration value predicted by the two-branch temporal coding network model. The prediction uncertainty is the output of the two-branch temporal coding network model. This is the magnification factor.

[0031] In one embodiment of the present invention, in step S1, the dynamic time series data includes a variety of physical quantities related to the gas exchange process, including gas concentration parameters, environmental parameters, and flow parameters.

[0032] This invention also provides a system for predicting the steady-state concentration of carbon dioxide in photosynthetic leaf chambers, comprising the following modules:

[0033] The data acquisition and preprocessing module is used to acquire dynamic time-series data of carbon dioxide concentration in leaf chambers during plant photosynthesis measurement. The dynamic time-series data includes concentration decrease and increase sequences generated by periodic gas path switching, as well as environmental data. The dynamic time-series data is preprocessed.

[0034] The prediction and comparison module is used to process the pre-processed dynamic time series data using a pre-trained dual-branch temporal coding network model, obtain the steady-state concentration prediction value corresponding to the current sequence and the prediction uncertainty of the steady-state concentration prediction value, and compare the prediction uncertainty with the uncertainty threshold preset for the current sequence type.

[0035] The determination and control module is used to determine that the prediction result is reliable when the prediction uncertainty is lower than the preset uncertainty threshold for the current sequence type, and output the steady-state concentration prediction value as the final prediction value; when the prediction uncertainty is higher than or equal to the preset uncertainty threshold for the current sequence type, the prediction result is determined to be unreliable, and subsequent time series data is collected and returned to the prediction and comparison module for a new round of prediction and determination.

[0036] The present invention also provides an electronic device comprising a processor, a memory, and a bus system, wherein the processor and the memory are connected via the bus system, the memory is used to store instructions, and the processor is used to execute the instructions stored in the memory to realize the method for predicting the steady-state concentration of carbon dioxide in the photosynthetic leaf chamber.

[0037] The present invention also provides a computer storage medium storing a computer software product, the computer software product including several instructions for causing a computer device to execute the method for predicting the steady-state concentration of carbon dioxide in photosynthetic leaf chambers.

[0038] The technical solution of the present invention has the following advantages compared with the prior art:

[0039] This invention, by constructing a bi-branch temporal coding network and introducing an uncertainty estimation mechanism, enables the early prediction of steady-state CO2 concentration during dynamic changes, significantly shortening the waiting time required for traditional measurements and greatly improving measurement efficiency. The method innovatively uses uncertainty to quantify prediction reliability; data acquisition can be terminated early when the prediction result reaches a confidence threshold, effectively improving the intelligence level of the measurement process while ensuring prediction accuracy. The unique bi-branch structure can separately handle the dynamic characteristics of the rising and falling segments of concentration changes. Combined with an adaptive input mechanism, it enhances the model's ability to fit nonlinear processes and its adaptability under different environmental conditions. This system can be directly integrated into existing single-sensor measurement platforms, optimizing the measurement process without increasing hardware complexity, providing an efficient and reliable solution for monitoring plant photosynthesis. Attached Figure Description

[0040] To make the content of this invention easier to understand, the invention will be further described in detail below with reference to specific embodiments and accompanying drawings.

[0041] Figure 1 This is a schematic flowchart of the method for predicting steady-state carbon dioxide concentration in the leaf chambers of photosynthesis provided in this embodiment of the invention;

[0042] Figure 2 This is a schematic diagram of the internal gas path and working principle of the plant photosynthesis measurement system in an embodiment of the present invention;

[0043] Figure 3 This is a schematic diagram showing the distribution of net CO2 assimilation of different plants under three light conditions in an embodiment of the present invention;

[0044] Figure 4 This is a schematic diagram showing the time distribution of measurements for different plants under three light conditions in an embodiment of the present invention;

[0045] Figure 5 This is a schematic diagram of the network structure of the dual-branch temporal coding network model in an embodiment of the present invention;

[0046] Figure 6 This describes the distribution of residuals for the falling and rising segments of corn in different uncertainty intervals in Experiment 3.

[0047] Figure 7 This describes the distribution of the descending and ascending residuals of potatoes in different uncertainty intervals in Experiment 3.

[0048] Figure 8 This shows the distribution of the descending and ascending residuals of loquat in different uncertainty intervals in Experiment 3;

[0049] Figure 9 This describes the distribution of the descending and ascending residuals of the radish in different uncertainty intervals during Experiment 3.

[0050] Figure 10 This describes the distribution of the descending and ascending residuals of the bok choy in Experiment 3 within different uncertainty intervals.

[0051] Figure 11 This describes the distribution of the descending and ascending segments of the tree structure in Experiment 3 within different uncertainty intervals.

[0052] Figure 12 This describes the distribution of the descending and ascending residuals of lettuce in different uncertainty intervals in Experiment 3.

[0053] Figure 13 This describes the distribution of the descending and ascending residuals of the orange tree in different uncertainty intervals in Experiment 3.

[0054] Figure 14 These are the statistical results of prediction errors and input ratios under different plant and light conditions in Experiment 4;

[0055] Figure 15 This is a schematic diagram of the steady-state carbon dioxide concentration prediction system for photosynthetic leaf chambers provided in an embodiment of the present invention. Detailed Implementation

[0056] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, so that those skilled in the art can better understand and implement the present invention. However, the embodiments described are not intended to limit the present invention.

[0057] Example 1:

[0058] like Figure 1 As shown, the present invention provides a method for predicting the steady-state concentration of carbon dioxide in a photosynthetic leaf chamber, characterized by comprising the following steps:

[0059] S1: Collect dynamic time-series data of carbon dioxide concentration in the leaf chamber and environmental data during plant photosynthesis measurement. The dynamic carbon dioxide concentration time-series data includes a concentration decrease sequence and an increase sequence generated by periodic gas path switching. The environmental parameter data includes Ts representing the CO2 sensor temperature (°C), Tleaf representing the leaf chamber temperature (°C), H representing the air humidity (%), P representing the air pressure (Pa), and F representing the gas flow rate (%). The dynamic time-series data is preprocessed.

[0060] S2: The pre-trained dual-branch temporal coding network model is used to process the pre-processed dynamic time series data to obtain the steady-state concentration prediction value corresponding to the current sequence and the prediction uncertainty of the steady-state concentration prediction value; the prediction uncertainty is compared with the uncertainty threshold preset for the current sequence type.

[0061] S3: When the prediction uncertainty is lower than the preset uncertainty threshold for the current sequence type, the prediction result is determined to be reliable, and the steady-state concentration prediction value is output as the final prediction value; when the prediction uncertainty is higher than or equal to the preset uncertainty threshold for the current sequence type, the prediction result is determined to be unreliable, and subsequent time series data is collected, returning to step S2 for a new round of prediction and determination.

[0062] This invention achieves accurate prediction of steady-state CO2 concentration even before the concentration has fully stabilized by dynamically acquiring time-series data and utilizing a bi-branch time-series coding network for intelligent analysis. By introducing prediction uncertainty as a criterion, the invention automatically outputs a reliable prediction result when the uncertainty is below a preset threshold; otherwise, it continues data acquisition and re-predicting. This mechanism effectively overcomes the problems of long waiting times and low efficiency associated with traditional methods, significantly improving measurement efficiency while ensuring the accuracy and reliability of prediction results. It provides a more intelligent and efficient monitoring method for plant photosynthesis research.

[0063] Specifically, in step S1, dynamic time-series data of carbon dioxide concentration in leaf chambers and environmental data are collected during the plant photosynthesis measurement process. In this embodiment, eight representative plants—corn, potato, radish, bok choy, lettuce, loquat, orange tree, and paper mulberry—are selected as experimental samples. To comprehensively examine the effects of different light conditions on plant photosynthesis, three strict light intensity gradients are set: low light intensity condition is light intensity below 500 μmol·m⁻¹. -2 ·s -1 Under medium light intensity conditions, the range is between 500 and 1000 μmol·m -2 ·s -1 The range is between 1000 μmol·m⁻¹, while high light intensity conditions require a light intensity greater than 1000 μmol·m⁻¹. -2 ·s -1 .

[0064] like Figure 2 As shown, in this embodiment, a plant photosynthesis measurement system based on a gas exchange method with a single-sensor structure is constructed. The system adopts a modular integrated design, with the main unit serving as the central control unit to coordinate the collaborative work of all components. The system's gas path structure includes two independent air pumps: Air pump 1 serves as the main air supply unit, responsible for providing a stable air source to the leaf chambers; Air pump 2 serves as the auxiliary air supply unit, responsible for providing an environmental reference airflow to the sensor. The gas is first pretreated through a buffer filter tube to remove impurities and particles, and then passes through a drying tube to reduce the gas humidity, creating ideal conditions for subsequent accurate measurements.

[0065] The core detection component of the system employs a high-precision non-dispersive infrared (NDIR) carbon dioxide sensor to monitor changes in CO2 concentration in the gas in real time. To comprehensively record environmental parameters, the system integrates a multi-sensor array: a humidity sensor monitors relative humidity, a constant humidity and pressure sensor simultaneously acquires pressure data, and a gas flow sensor accurately measures gas flow rate. The plant samples being tested are placed in a dedicated leaf chamber, which is equipped with an LED light source above it, allowing for precise adjustment of light intensity according to experimental needs.

[0066] During system operation, a precise three-way solenoid valve controls the switching of gas pathways, enabling alternating measurements of the reference airflow and the leaf chamber airflow. The specific workflow is as follows: When the solenoid valve switches to leaf chamber airflow measurement mode, gas flows through the leaf chamber containing the plant sample. The leaves consume CO2 through photosynthesis, causing the CO2 concentration detected by the sensor to continuously decrease, forming a decreasing sequence. When the solenoid valve switches to reference airflow measurement mode, the gas bypasses the leaf chamber and directly enters the detection unit, causing the CO2 concentration recorded by the sensor to show an increasing trend, forming an increasing sequence. This complete measurement cycle accurately records the typical dynamic response process of gas exchange: "decreasing segment - steady state segment - increasing segment - steady state segment".

[0067] During the measurement process, the system simultaneously records multiple physical parameters, including CO2 concentration (ppm), CO2 sensor temperature (°C), leaf chamber temperature (°C), air humidity (%), air pressure (Pa), and gas flow rate (ml·min). -1 These multivariate data provide a rich foundation for subsequent modeling. Considering the significant impact of different plant species, light conditions, and environmental factors on the stabilization time of CO2 concentration, to improve the intelligence level of the measurement process, this system innovatively designs a real-time steady-state determination algorithm based on a sliding window. This algorithm quantifies the volatility of the data by calculating the standard deviation of the CO2 concentration within the sliding window. Its mathematical expression is:

[0068] ,

[0069] in, This represents the CO2 concentration value at the i-th sampling point within the sliding window. This represents the average concentration of 20 sampling points within the window. This formula effectively captures the dispersion of the concentration sequence; the standard deviation calculated over 20 consecutive sliding windows must all meet the specified value. When this condition is met, the system determines that the current CO2 concentration has entered a steady-state phase. At this point, the control module immediately triggers the solenoid valve to automatically switch to the next measurement mode, realizing intelligent closed-loop control of the measurement process.

[0070] In this embodiment, under three different light intensities, systematic gas exchange measurements were performed on 40 leaves of each plant species, resulting in a total of 960 sets of high-quality experimental data. Figure 3 As shown, analysis of this data revealed that different plants exhibited significantly different gas exchange characteristics under different light conditions. Meanwhile, as... Figure 4 As shown, the steady-state settling time of the measurement system varies significantly under different conditions, ranging from approximately 220 to 350 seconds.

[0071] To construct a time-series sample dataset suitable for training deep learning models, the original measurement data were systematically labeled according to plant species and light intensity, adding plant type and light level labels to each time series. Subsequently, based on the solenoid valve status information, the continuous measurement data were precisely divided into a decreasing segment (CO2 concentration decrease process) and an increasing segment (CO2 concentration increase process), and by combining the data from these two stages, a complete time-series sample of photosynthetic rate measurement was constructed.

[0072] After the sample construction was completed, a strict data partitioning strategy was adopted. All data samples for each plant under different light conditions were randomly divided into a training set (70%), a validation set (10%), and a test set (20%) in a 7:1:2 ratio to ensure the reliability and independence of model training, parameter tuning, and performance evaluation. Finally, each standardized sample contained the following data structure:

[0073] down: The descending segment input sequence records the dynamic process of CO2 concentration decrease after the measurement mode is switched to blade chamber airflow measurement;

[0074] up: Input sequence for the rising segment, recording the dynamic process of CO2 concentration rise after the measurement mode is switched to reference airflow measurement;

[0075] t1 and t2 represent the steady-state concentration reference values ​​corresponding to the falling and rising segments, respectively, obtained by calculating the arithmetic mean of the last 20 sampling points of each segment.

[0076] A systematic data acquisition and processing workflow lays the data foundation for the subsequent training and validation of deep learning models, ensuring the reliability and practicality of the prediction method.

[0077] Furthermore, after completing data collection and sample construction, a systematic preprocessing process is implemented, including two core steps: normalization and data augmentation, to ensure data quality and improve model performance. Data normalization is the first step in the preprocessing process. Since the collected physical quantity data have completely different dimensions and numerical ranges, to eliminate the adverse effects of these dimensional differences on model training and to prevent certain features from dominating the model training process due to their large numerical ranges, this embodiment uses a min-max normalization method to linearly map all features to the [0,1] interval. Its mathematical expression is:

[0078] ,

[0079] in, Represents the normalized eigenvalues. These are the original eigenvalues. and These are the minimum and maximum values ​​of the feature in the entire dynamic time-series dataset, respectively. This process projects the original data onto a unified numerical space through linear transformation, effectively eliminating the magnitude differences between different features. This avoids the problems of unstable gradient updates and optimization difficulties caused by inconsistent feature scales during model training, while significantly accelerating the model's convergence process and laying a solid foundation for the stable training of subsequent deep learning models.

[0080] Based on the completion of data normalization, in order to further improve the robustness and generalization ability of the model and enable it to adapt to various complex and changing situations in the actual measurement environment, this embodiment implements a multi-level and systematic data augmentation strategy in the data loading stage, specifically including: (1) random truncation of the descent segment: randomly truncate a data segment from the complete descent segment time series as the descent segment input sequence to simulate the scenario of collecting data from different times in the actual measurement; (2) end concentration recording: record the CO2 concentration value at the end of the truncation segment data segment, denoted as . This value serves as a key alignment parameter for subsequent matching and truncation of the rising segment, ensuring the physical continuity between data at different stages; (3) Rising segment matching and truncation: The nearest neighbor search algorithm is used in the rising segment sequence to accurately find the matching and truncation of the rising segment. The position with the smallest concentration difference is used to randomly extract a segment of the corresponding length as the input for the rising segment. This ingenious design ensures the continuity and physical rationality of the concentration change in the two stages of rising and falling, and accurately simulates the natural transition of concentration change during the actual gas path switching process; (4) Bernoulli sampling of the rising segment: a random masking mechanism based on Bernoulli distribution is introduced to generate a binary random variable in the interval [0,1]. .when When the rising segment input is replaced with a specific padding value (-1) and does not participate in the model training of the current round; when In this case, valid rising segment samples are retained for training.

[0081] The preprocessing described above ensures consistent input data quality and feature scale, while data augmentation strategies significantly expand the diversity and coverage of the training data. This comprehensive preprocessing approach provides high-quality, diverse input data for the subsequent training of the dual-branch temporal coding network.

[0082] Furthermore, in step S2, a pre-trained dual-branch temporal coding network model (DBE-TSNet) is used to process the pre-processed dynamic time-series data. The design of DBE-TSNet fully considers the physical characteristics of dynamic changes in CO2 concentration and actual measurement requirements. Through innovative network architecture and training strategies, it achieves high-precision and high-efficiency steady-state concentration prediction.

[0083] like Figure 5 As shown, DBE-TSNet employs an innovative dual-branch parallel architecture, specifically designed for the two different dynamic characteristics of the descending and ascending segments present in gas exchange measurements.

[0084] The method for processing pre-processed dynamic time-series data using a pre-trained two-branch temporal coding network model includes the following steps:

[0085] S2.1: Set up two parallel encoder branches, input the falling segment sequence in the dynamic time series data into the first encoder branch, and input the rising segment sequence into the second encoder branch, and process them independently respectively;

[0086] S2.2: In each encoder branch, the input sequence is sequentially processed by differential channel enhancement, local convolution feature extraction, and linear channel fusion, so that each branch outputs high-order temporal features that can characterize the dynamic characteristics of its corresponding sequence.

[0087] S2.3: The high-order temporal features output by the first encoder branch and the second encoder branch are concatenated along the feature dimension to obtain the fused global feature representation;

[0088] S2.4: Perform a fully connected mapping on the fused global feature representation to obtain the predicted steady-state concentration value of the falling segment, the prediction uncertainty of the falling segment, the predicted steady-state concentration value of the rising segment, and the prediction uncertainty of the rising segment.

[0089] In step S2.2, the method for sequentially performing differential channel enhancement, local convolution feature extraction, and linear channel fusion on the input sequence in each encoder branch is as follows:

[0090] Differential channel enhancement is performed on the input multivariate time series data to simultaneously extract the absolute state and the trend of change;

[0091] One-dimensional convolution and non-linear activation operations are performed on the enhanced original and differential sequences to extract cross-channel joint features;

[0092] The cross-channel joint features are fully connected to compress the temporal dimension and capture long-range dependencies, resulting in dimensionally compressed temporal features.

[0093] The difference between adjacent time steps in the time dimension of the compressed temporal features is calculated, and the difference is concatenated with the original features to enhance the perception of dynamic concentration changes, thus obtaining dynamically enhanced high-order temporal features.

[0094] Specifically, the model adopts a design that combines a dual-branch parallel coding architecture with a fusion prediction module to simultaneously process dynamic time-series data of CO2 concentration decrease and increase, and output steady-state concentration prediction values ​​and corresponding prediction uncertainties.

[0095] The model input consists of multivariate time-series data collected during plant photosynthesis measurement, represented as follows: ,

[0096] Where t is the time series length and d is the feature dimension. Each sample contains six physical quantity feature vectors, which can be represented as:

[0097] ,

[0098] in, This represents the CO2 concentration value (unit: ppm = μmol·mol). -1 ), This indicates the temperature of the CO2 sensor (in °C). H represents the temperature of the leaf chamber (°C), P represents the air humidity (%), P represents the air pressure (Pa), and F represents the gas flow rate (ml·min). -1 ).

[0099] The DBE-TSNet model learns a nonlinear mapping function through training. To achieve dynamic sequence to steady-state CO2 concentration The prediction is expressed as:

[0100] .

[0101] The core components of the DBE-TSNet model include:

[0102] Diff differential channel enhancement layer: The original input sequence is concatenated with its first-order difference sequence in the time dimension and jointly input into the subsequent network, thereby capturing the absolute state and changing trend of CO2 concentration at the same time and enhancing the model's ability to perceive dynamic processes.

[0103] AC-LC cross-coding block: In each encoder branch, cross-channel feature extraction and temporal structure compression are performed sequentially. First, joint features among multiple variables are extracted through one-dimensional convolution and non-linear activation operations; then, the temporal structure is compressed through a fully connected layer to capture long-range dependencies; finally, the compressed features are concatenated with the difference features of adjacent time steps to further enhance the modeling of concentration dynamics.

[0104] Global Average Pooling (GAP) layer: Global average pooling is performed on the encoded high-order time series features in the time dimension to extract the overall steady-state features of the sequence, avoids direct input of high-dimensional time series data into the fully connected layer, reduces the number of model parameters and improves generalization ability;

[0105] Multilayer Perceptron (MLP): The features output by the dual-branch encoder are concatenated in different dimensions and then input into the MLP for nonlinear combination and mapping. The final output consists of four target values: the predicted steady-state concentration in the descent phase. Uncertainty in predicting the descent segment Predicted steady-state concentration during the rising phase Uncertainty in the rising segment prediction .

[0106] In the training process of DBE-TSNet, in order to simultaneously optimize the steady-state concentration prediction value and its corresponding uncertainty, this embodiment adopts a loss function based on Gaussian negative log-likelihood, with the following basic form:

[0107] ,

[0108] in, This represents the true steady-state concentration value. The steady-state concentration value predicted by the model. The prediction uncertainty of the model output. The magnification factor (in this embodiment) This is used to enhance the sensitivity of the prediction residuals to the loss, thereby guiding the model to pay more attention to samples that are difficult to fit.

[0109] Considering that the Bernoulli random sampling strategy was used for the rising segment samples during data preprocessing, the rising segment only has about a 50% probability of participating in model training, resulting in a systematic weakening of its expected contribution to the overall loss. To avoid the model becoming overly biased towards learning the falling segment due to the sampling strategy, this embodiment introduces a weight factor of 2 for the rising segment loss term when constructing the total loss function, ensuring that the falling segment and the rising segment have equivalent importance in the overall training objective. Therefore, the final loss function of the model is defined as the sum of the falling segment loss and the weighted rising segment loss:

[0110] ,

[0111] in, The total loss of the model, and The losses for the descent and ascent segments are calculated using the Gaussian loss formula described above. The composite loss function design ensures that the model can learn effectively in both different stages of the dynamic sequence, thus improving the overall prediction performance.

[0112] For model training, DBE-TSNet uses the Adam optimizer for parameter optimization, with an initial learning rate set to... The training process incorporates a dynamic learning rate scheduling strategy: if the validation set loss does not decrease within 20 consecutive training epochs, the current learning rate is reduced to 50% of its original value, but the minimum learning rate is never lower than [a certain value]. The maximum number of iterations in the training process is set to 300 rounds, and the batch size is set to 8. After each round of training, the average loss and mean absolute error (MAE) are calculated on an independent validation set, and the best-performing model parameters are selected and saved to prevent overfitting and ensure the model's generalization ability.

[0113] To comprehensively and objectively evaluate the overall performance of the DBE-TSNet model, this embodiment uses multi-dimensional evaluation metrics:

[0114] Mean Absolute Error (MAE): Measures the average absolute deviation between model-predicted values ​​and actual steady-state concentration values;

[0115] Root mean square error (RMSE): reflects the stability of the model's prediction results and the degree of dispersion of the error;

[0116] Mean Absolute Percentage Error (MAPE): Assessing the accuracy of predictions from the perspective of relative error;

[0117] Coefficient of determination (R²): Characterizes the goodness of fit between predicted and actual values, and evaluates the model's ability to explain data variation.

[0118] Furthermore, in step S3, the steady-state concentration prediction result is determined and adaptively controlled based on the prediction uncertainty.

[0119] Specifically, after obtaining the steady-state concentration prediction value and its corresponding prediction uncertainty from the dual-branch time-series coding network model, the system enters the intelligent judgment stage. In this embodiment, uncertainty thresholds, verified through extensive experiments, are preset for both the falling and rising segments. The uncertainty threshold for the falling segment is set to 0.0267, and the uncertainty threshold for the rising segment is set to 0.0146. These thresholds are determined based on statistical analysis of the test set data, ensuring that the model's prediction error remains stable within 1 ppm under these thresholds, meeting the accuracy requirements for calculating the photosynthetic CO2 assimilation rate.

[0120] When the prediction uncertainty output by the model is lower than the preset threshold corresponding to the current sequence type (declining or rising segment), the system determines that the prediction result has high reliability. At this time, the determination and control module will immediately output the steady-state concentration prediction value corresponding to the current moment. or The data is used as the final prediction result, and can automatically trigger the switching of the solenoid valve or terminate the data acquisition of the current stage according to the preset strategy. The core advantage of this mechanism is that it allows the system to make decisions in advance based on the high confidence prediction of the model before the CO2 concentration has fully reached physical steady state, thereby significantly shortening the measurement waiting time.

[0121] When the prediction uncertainty output by the model is higher than or equal to a preset threshold, the system determines that the reliability of the current prediction result is insufficient and the conditions for early termination of data acquisition are not yet met. At this time, the judgment and control module will not output a prediction value, but will instead instruct the data acquisition system to continue acquiring subsequent time-series data. Subsequently, the system automatically returns to step S2, and inputs the time-series sequence, which includes new data points and has been length-expanded, back into the pre-trained dual-branch time-series coding network model for a new round of prediction and uncertainty evaluation. This process is repeated until the prediction uncertainty meets the condition of being lower than the threshold.

[0122] This uncertainty-based adaptive cyclic prediction mechanism constitutes the core control logic of the method of this invention. It dynamically determines the termination time of data acquisition through the confidence index output by the model, thereby achieving a high degree of intelligence and maximizing efficiency in the measurement process while ensuring prediction accuracy. It is especially suitable for complex measurement scenarios where the steady-state establishment time is dynamically affected by various factors such as plant species and light intensity, demonstrating excellent adaptability and robustness.

[0123] Experiment 1:

[0124] To evaluate the performance advantage of the DBE-TSNet model in the CO2 steady-state prediction task, this experiment compared it with six typical time series prediction models, including one-dimensional convolutional models (1D-CNN), recurrent neural network models (LSTM, GRU), linear time series models (D-Linear), lightweight convolutional models (LightTS), and Transformer-like models (i-Transformer).

[0125] All comparison models were trained and evaluated on the same training, validation, and test sets. The experiments used multidimensional evaluation metrics: mean absolute error (MAE), root mean square error (RMSE), mean absolute percentage error (MAPE), and coefficient of determination (R²) to comprehensively measure the prediction accuracy, stability, and goodness of fit of each model.

[0126] As shown in Table 1, the experimental results demonstrate that DBE-TSNet achieves optimal or near-optimal performance across all metrics. The MAE in the descent phase is only 1.22 ppm, and the RMSE is 2.43 ppm, significantly outperforming traditional deep learning models; the error in the ascending phase is also the lowest (MAE = 2.07 ppm, RMSE = 3.95 ppm). Furthermore, the R² values ​​reach 0.996 (descent phase) and 0.984 (ascending phase), indicating a very high fit to the dynamic changes in CO2. In contrast, D-Linear and i-Transformer perform poorly, suggesting that steady-state prediction tasks have highly nonlinear characteristics, and Transformer-type models are prone to overfitting in short-sequence scenarios. While LightTS and GRU demonstrate some modeling ability, their overall performance is still significantly weaker than the DBE-TSNet model.

[0127] Table 1:

[0128] Experiment 2:

[0129] To verify the contribution of the core components (AC cross-channel convolution, LC temporal linear compression, and Diff neighbor difference enhancement) in the DBE-TSNet model to the prediction performance, a systematic ablation experiment was designed.

[0130] As shown in Table 2, the experiment constructed six variant models by sequentially removing each core component for comparative analysis. The experimental results show that when all three components are removed simultaneously, the model performance drops significantly (MAE in the descent segment reaches 31.19 ppm, R² is only 0.148); introducing the Diff component alone can improve performance to some extent; and the combination of AC and LC components is particularly crucial for performance improvement. Ultimately, the fully configured DBE-TSNet model (AC+LC+Diff) achieved the best performance (MAE in the descent segment = 1.22 ppm, R² = 0.996), fully validating the effectiveness and complementarity of each core component.

[0131] Table 2:

[0132] The AC component enhances the model's ability to model multivariate coupling relationships through cross-channel convolution; the LC component effectively captures the dependency structure of long sequences through temporal linear compression; and the Diff component strengthens the model's perception of concentration change trends through difference enhancement. The organic combination of these three components enables the model to comprehensively capture the complex features of dynamic changes in CO2 concentration.

[0133] Experiment 3:

[0134] To verify the correlation between the prediction uncertainty of the model output and the actual prediction error, this experiment divided all test samples into ten equal parts according to the uncertainty range of the model output, and statistically analyzed the MAE, RMSE, MAPE and input segment ratio of the falling and rising segments in each interval.

[0135] As shown in Table 3, the experimental results indicate that the prediction error (MAE) monotonically increases with increasing uncertainty, regardless of whether it is in the decreasing or increasing uncertainty range. In the low uncertainty range (decreasing segment: 0.0048-0.0086; increasing segment: 0.0047-0.0077), the model's MAE is less than 1 ppm, and the RMSE remains stable below 1.2 ppm, indicating that the model's predictions are highly reliable at this level. As uncertainty increases, the prediction error gradually worsens, while the input segment ratio decreases significantly, indicating that the model can correctly reflect the decrease in prediction reliability when data information is insufficient.

[0136] Table 3:

[0137] In addition to statistical analysis of the uncertainty intervals of the overall dataset, the stability of the model under different physiological structures, leaf characteristics, and gas exchange features was further evaluated at the individual plant level. To this end, all test set samples of the eight plant species were partitioned and statistically analyzed according to the aforementioned uncertainty intervals. For example... Figures 6-13 The figure shows the residual box plots for each plant in the descending and ascending segments.

[0138] Regardless of the plant species, the prediction error shows a monotonically increasing trend with the increase of uncertainty, which is consistent with the overall pattern.

[0139] Based on the above statistical patterns, this embodiment determines the uncertainty threshold for the decreasing segment to be 0.0267 and the uncertainty threshold for the increasing segment to be 0.0146. When the uncertainty of the model output is lower than the threshold of the corresponding stage, the prediction error can be stably maintained below 1 ppm, meeting the accuracy requirements for calculating the photosynthetic CO2 assimilation rate, and can serve as a reliable basis for early termination of sampling in actual measurements.

[0140] Experiment 4:

[0141] To evaluate the acceleration effect of the method of the present invention in actual measurement, this experiment statistically analyzes the advance prediction time of all samples in the test dataset based on a determined uncertainty threshold (falling segment: 0.0267, rising segment: 0.0146).

[0142] The experimental procedure is as follows: For the descending segment of each sample, starting from a random starting position, the input sequence is gradually expanded with a single sampling point as the step size, and the uncertainty of the model output is calculated in real time. When the uncertainty is lower than the threshold for the first time, it is recorded as the "predicted point of the descending segment". Then, in the ascending segment, the position closest to the CO2 concentration of the predicted point of the descending segment is found as the starting point, and the input is expanded with a single sampling point as the step size until the uncertainty is lower than the threshold of the ascending segment.

[0143] As shown in Table 4, the experimental results demonstrate that the model maintains high prediction accuracy even when the thresholds are reached (MAE = 0.69 ppm for the descent segment and 0.97 ppm for the ascending segment), and reliable predictions can be achieved using only a portion of the input data: the average input times for the descent and ascending segments are 54.14 seconds (40% of the total time) and 71.54 seconds (61% of the total time), respectively. Based on this, measurements can be stopped 77.81 seconds and 51.06 seconds earlier, respectively, resulting in a total time saving of 128.87 seconds, reducing the measurement cycle by approximately 50%.

[0144] Table 4:

[0145]

[0146] These results fully demonstrate that the uncertainty-based early termination strategy can significantly improve CO2 measurement efficiency while ensuring prediction accuracy.

[0147] To further evaluate the stability of the model under different physiological characteristics and photosynthetic intensities, this study statistically analyzed the MAE and input time ratio (Ratio) for each plant at which reliable predictions were achieved under low, medium, and high light intensities. The results are as follows: Figure 14 As shown.

[0148] Despite differences in plant species and light intensity, the model's MAE remained within a low error range under all conditions, indicating that the model has reliable cross-plant generalization ability.

[0149] Example 2:

[0150] like Figure 15 As shown, based on the same inventive concept as in Embodiment 1, the present invention also provides a system for predicting the steady-state concentration of carbon dioxide in photosynthetic leaf chambers, used to implement the steps of the method for predicting the steady-state concentration of carbon dioxide in photosynthetic leaf chambers described in Embodiment 1, including the following modules:

[0151] The data acquisition and preprocessing module is used to acquire dynamic time-series data of carbon dioxide concentration in leaf chambers during plant photosynthesis measurement, and to preprocess the dynamic time-series data, wherein the dynamic time-series data includes a concentration decrease sequence and an increase sequence generated by periodic gas path switching.

[0152] The prediction and comparison module is used to process the pre-processed dynamic time series data using a pre-trained dual-branch temporal coding network model, obtain the steady-state concentration prediction value corresponding to the current sequence and the prediction uncertainty of the steady-state concentration prediction value, and compare the prediction uncertainty with the uncertainty threshold preset for the current sequence type.

[0153] The determination and control module is used to determine that the prediction result is reliable when the prediction uncertainty is lower than the preset uncertainty threshold for the current sequence type, and output the steady-state concentration prediction value as the final prediction value; when the prediction uncertainty is higher than or equal to the preset uncertainty threshold for the current sequence type, the prediction result is determined to be unreliable, and subsequent time series data is collected and returned to the prediction and comparison module for a new round of prediction and determination.

[0154] The data acquisition and preprocessing module, prediction and comparison module, and judgment and control module of the photosynthetic leaf chamber carbon dioxide steady-state concentration prediction system proposed in this embodiment are respectively used to implement steps S1, S2, and S3 in the photosynthetic leaf chamber carbon dioxide steady-state concentration prediction method in Embodiment 1. To avoid redundancy, they will not be described again here.

[0155] Example 3:

[0156] The present invention also provides an electronic device, which includes a processor, a memory, and a bus system. The processor and the memory are connected through the bus system. The memory is used to store instructions, and the processor is used to execute the instructions stored in the memory to implement the steady-state carbon dioxide concentration prediction method for photosynthetic leaf chambers described in Embodiment 1.

[0157] Example 4:

[0158] The present invention also provides a computer storage medium storing a computer software product, the computer software product including several instructions for causing a computer device to execute the steady-state concentration prediction method for carbon dioxide in photosynthetic leaf chambers as described in Embodiment 1.

[0159] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0160] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of 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, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0161] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0162] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0163] Obviously, the above embodiments are merely illustrative examples for clear explanation and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations here. However, obvious variations or modifications derived therefrom are still within the scope of protection of this invention.

Claims

1. A method for predicting the steady-state concentration of carbon dioxide in a photosynthetic leaf chamber, characterized in that, Includes the following steps: S1: Collect dynamic time-series data of carbon dioxide concentration in leaf chambers during plant photosynthesis measurement. The dynamic time-series data includes concentration decrease and increase sequences generated by periodic gas path switching, as well as environmental data. The dynamic time-series data is preprocessed. S2: The pre-trained dual-branch temporal coding network model is used to process the pre-processed dynamic time series data to obtain the steady-state concentration prediction value corresponding to the current sequence and the prediction uncertainty of the steady-state concentration prediction value; the prediction uncertainty is compared with the uncertainty threshold preset for the current sequence type. Among them, the methods for processing pre-processed dynamic time-series data using a pre-trained dual-branch temporal coding network model include: S2.1: Set up two parallel encoder branches, input the falling segment sequence in the dynamic time series data into the first encoder branch, and input the rising segment sequence into the second encoder branch, and process them independently respectively; S2.2: In each encoder branch, the input sequence is sequentially enhanced across differential channels, extracted from local convolutional features, and fused with linear channels, so that each branch outputs high-order temporal features that can characterize the dynamic characteristics of its corresponding sequence. S2.3: The high-order temporal features output by the first encoder branch and the second encoder branch are concatenated along the feature dimension to obtain the fused global feature representation; S2.4: Perform a fully connected mapping on the fused global feature representation to obtain the predicted steady-state concentration value in the falling segment, the prediction uncertainty in the falling segment, the predicted steady-state concentration value in the rising segment, and the prediction uncertainty in the rising segment; The method for sequentially performing cross-differential channel enhancement, local convolution feature extraction, and linear channel fusion on the input sequence is as follows: Differential channel enhancement is performed on the input multivariate time series data to simultaneously extract the absolute state and the trend of change; One-dimensional convolution and non-linear activation operations are performed on the enhanced original and differential sequences to extract cross-channel joint features; The cross-channel joint features are fully connected to compress the temporal dimension and capture long-range dependencies, resulting in dimensionally compressed temporal features. S3: When the prediction uncertainty is lower than the preset uncertainty threshold for the current sequence type, the prediction result is determined to be reliable, and the steady-state concentration prediction value is output as the final prediction value; when the prediction uncertainty is higher than or equal to the preset uncertainty threshold for the current sequence type, the prediction result is determined to be unreliable, and subsequent time series data is collected, returning to step S2 for a new round of prediction and determination.

2. The method for predicting steady-state carbon dioxide concentration in photosynthetic leaf chambers according to claim 1, characterized in that: In step S1, the method for preprocessing the dynamic time-series data is as follows: The collected dynamic time-series data is normalized, and each feature in the dynamic time-series data is linearly mapped to the [0,1] interval. The normalization formula is: ， in, These are the normalized eigenvalues. These are the original eigenvalues. and These are the minimum and maximum values ​​of this feature in dynamic time-series data, respectively.

3. The method for predicting steady-state carbon dioxide concentration in photosynthetic leaf chambers according to claim 1 or 2, characterized in that: In step S1, the preprocessing of the dynamic time-series data also includes data augmentation, specifically as follows: The descending segment sequence is randomly truncated to obtain the descending segment input sequence; Locate the position in the rising segment sequence that matches the carbon dioxide concentration value at the end of the falling segment input sequence, and randomly extract from that position to obtain the rising segment input sequence; The rising segment input sequence is selectively sampled using a random mask. When the random mask indicates a first state, the rising segment input sequence is replaced with padding values ​​and does not participate in model training. When the random mask indicates a second state, the rising segment input sequence is retained for model training.

4. The method for predicting steady-state carbon dioxide concentration in photosynthetic leaf chambers according to claim 1, characterized in that: In step S2, the dual-branch temporal coding network model is trained using a loss function based on Gaussian likelihood. The expression is: ， in, This represents the true steady-state concentration value. The steady-state concentration value predicted by the two-branch temporal coding network model. The prediction uncertainty is the output of the two-branch temporal coding network model. This is the magnification factor.

5. The method for predicting steady-state carbon dioxide concentration in photosynthetic leaf chambers according to claim 1, characterized in that: In step S1, the dynamic time series data includes a variety of physical quantities related to the gas exchange process, including gas concentration parameters, environmental parameters, and flow parameters.

6. A system for predicting the steady-state concentration of carbon dioxide in a photosynthetic leaf chamber, characterized in that, Includes the following modules: The data acquisition and preprocessing module is used to acquire dynamic time-series data of carbon dioxide concentration in leaf chambers during plant photosynthesis measurement. The dynamic time-series data includes concentration decrease and increase sequences generated by periodic gas path switching, as well as environmental data. The dynamic time-series data is preprocessed. The prediction and comparison module is used to process the pre-processed dynamic time series data using a pre-trained dual-branch temporal coding network model, obtain the steady-state concentration prediction value corresponding to the current sequence and the prediction uncertainty of the steady-state concentration prediction value, and compare the prediction uncertainty with the uncertainty threshold preset for the current sequence type. Among them, the methods for processing pre-processed dynamic time-series data using a pre-trained dual-branch temporal coding network model include: S2.1: Set up two parallel encoder branches, input the falling segment sequence in the dynamic time series data into the first encoder branch, and input the rising segment sequence into the second encoder branch, and process them independently respectively; S2.2: In each encoder branch, the input sequence is sequentially enhanced across differential channels, extracted from local convolutional features, and fused with linear channels, so that each branch outputs high-order temporal features that can characterize the dynamic characteristics of its corresponding sequence. S2.3: The high-order temporal features output by the first encoder branch and the second encoder branch are concatenated along the feature dimension to obtain the fused global feature representation; S2.4: Perform a fully connected mapping on the fused global feature representation to obtain the predicted steady-state concentration value in the falling segment, the prediction uncertainty in the falling segment, the predicted steady-state concentration value in the rising segment, and the prediction uncertainty in the rising segment; The method for sequentially performing cross-differential channel enhancement, local convolution feature extraction, and linear channel fusion on the input sequence is as follows: Differential channel enhancement is performed on the input multivariate time series data to simultaneously extract the absolute state and the trend of change; One-dimensional convolution and non-linear activation operations are performed on the enhanced original and differential sequences to extract cross-channel joint features; The cross-channel joint features are fully connected to compress the temporal dimension and capture long-range dependencies, resulting in dimensionally compressed temporal features. The determination and control module is used to determine that the prediction result is reliable when the prediction uncertainty is lower than the preset uncertainty threshold for the current sequence type, and output the steady-state concentration prediction value as the final prediction value; when the prediction uncertainty is higher than or equal to the preset uncertainty threshold for the current sequence type, the prediction result is determined to be unreliable, and subsequent time series data is collected and returned to the prediction and comparison module for a new round of prediction and determination.

7. An electronic device, characterized in that, The electronic device includes a processor, a memory, and a bus system. The processor and the memory are connected through the bus system. The memory is used to store instructions, and the processor is used to execute the instructions stored in the memory to implement the steady-state carbon dioxide concentration prediction method for photosynthetic leaf chambers as described in any one of claims 1 to 5.

8. A computer storage medium, characterized in that, The computer storage medium stores a computer software product, the computer software product including several instructions for causing a computer device to execute the steady-state concentration prediction method for carbon dioxide in photosynthetic leaf chambers as described in any one of claims 1 to 5.

Images