Carbon flux prediction method and apparatus

Abstract

The present application relates to a kind of carbon flux prediction method and device, belong to carbon flux prediction technical field, method includes obtaining the flux tower observation sequence data of target area;Flux tower observation sequence data is reconstructed to dialogue sample set;Dialogue sample set is input to sequence-to-sequence big model and is trained, to learn the time-space dependence between multivariate in dialogue history sequence by self-attention mechanism, and generate the carbon flux prediction sequence corresponding to expected response sequence;Based on the error between carbon flux prediction sequence and expected response sequence, the parameters of big model are iteratively optimized by back propagation algorithm until model converges, obtain the trained prediction big model;New flux tower observation sequence data is obtained, and dialogue history sequence based on new flux tower observation sequence data is input to prediction big model, obtain the carbon flux prediction sequence of future P time steps, provide more accurate scientific basis for carbon management and climate change response.

Description

Technical Field

[0001] This invention relates to the field of carbon flux prediction technology, and in particular to a carbon flux prediction method and apparatus. Background Technology

[0002] Global climate change is one of the major challenges facing humanity today, with greenhouse gas emissions (especially carbon dioxide) being a primary driver. Accurately estimating and forecasting carbon fluxes in terrestrial ecosystems is crucial for understanding the global carbon cycle, assessing carbon sink potential, and developing effective climate change response strategies. Terrestrial ecosystems are a vital component of the global carbon cycle, absorbing carbon dioxide from the atmosphere through photosynthesis and releasing it through respiration and decomposition. Therefore, accurately monitoring and simulating the carbon budget of terrestrial ecosystems has profound implications for achieving carbon neutrality goals. Summary of the Invention

[0003] In view of the above-mentioned shortcomings and deficiencies of the prior art, the present invention provides a carbon flux prediction method and apparatus, which solves the technical problem of accurately predicting carbon balance in the prior art.

[0004] To achieve the above objectives, the main technical solutions adopted by the present invention include:

[0005] The first aspect of this invention provides a method for predicting carbon flux.

[0006] The carbon flux prediction method proposed in this embodiment of the invention includes:

[0007] Acquire flux tower observation sequence data for the target area; wherein, the flux tower observation sequence data includes multimodal environmental variable observations at continuous time steps and carbon flux values ​​corresponding to the multimodal environmental variable observations;

[0008] With a predetermined sliding step size, a dialog window with a time step size of W is slid along the time axis. The observation values ​​of multimodal environmental variables in the dialog window are arranged in chronological order to form a dialog history sequence. The carbon flux data in the time period with a time step size of P after the dialog window are used to form the expected response sequence, thus obtaining a dialog sample set.

[0009] The dialogue sample set is input into a sequence-to-sequence model based on the Transformer architecture for training, so as to learn the spatiotemporal dependencies between multiple variables in the dialogue history sequence through the self-attention mechanism and generate a carbon flux prediction sequence corresponding to the expected response sequence.

[0010] Based on the error between the carbon flux prediction sequence and the expected response sequence, the parameters of the large model are iteratively optimized through the backpropagation algorithm until the model converges, thus obtaining the trained prediction large model.

[0011] New flux tower observation sequence data is acquired, and the dialogue history sequence constructed based on the new flux tower observation sequence data is input into the prediction big model to obtain the carbon flux prediction sequence for the next P time steps.

[0012] In some instances, the process of sliding a dialog window with a predetermined sliding step size of W along the time axis, arranging the multimodal environmental variable observations within the dialog window in chronological order to form a dialog history sequence, includes:

[0013] Multiple dialogue windows of different scales are used to capture carbon cycle processes at different time scales, so that the observation values ​​of multimodal environmental variables within the dialogue windows are arranged in chronological order to form a dialogue history sequence; wherein, the multiple dialogue windows of different scales include a first window, a second window and a third window;

[0014] The window length W1 of the first window is set to N1 time steps to capture the intraday variation rhythm of photosynthesis and respiration.

[0015] The window length W2 of the second window is set to N2 time steps to capture the impact of weather processes on carbon flux;

[0016] The window length W3 of the third window is set to N3 time steps to capture the cumulative effect of phenological changes on carbon flux.

[0017] In some instances, the step of inputting the dialogue sample set into a sequence-to-sequence large model based on the Transformer architecture for training includes: in the embedding layer of the sequence-to-sequence large model, when encoding the observation data at each time step, simultaneously fusing the spatial background information of the flux tower site; wherein the spatial background information is obtained by learningable embedding encoding of the site's geographical location, land cover type, and terrain features.

[0018] In some instances, the learning of spatiotemporal dependencies among multiple variables in the dialogue history sequence through a self-attention mechanism includes...

[0019] The large model learns the spatiotemporal dependencies between multiple variables in the dialogue history sequence through a multi-head self-attention mechanism in the encoder; each attention head independently focuses on different feature subspaces to capture the association patterns between different types of environmental variables and carbon flux.

[0020] In some instances, the attention head includes:

[0021] The first set of attention points is used to capture short-term fluctuation dependencies between adjacent time steps;

[0022] The second set of attention is used to capture weather process dependencies spanning multiple days to several weeks;

[0023] The third set of attention is used to capture long-term trend dependencies across seasons and years.

[0024] In some instances, when encoding the observation data at each time step in the embedding layer of the sequence-to-sequence large model, the spatial background information of the flux tower sites is simultaneously fused, including:

[0025] Encode the spatial feature data of GIS (Geographic Information System) into the spatial background vector;

[0026] The spatial background vector is concatenated with the observation data at each time step in the dialogue history sequence to form an enhanced dialogue history sequence with spatial context information.

[0027] In some instances, encoding GIS spatial feature data into spatial background vectors includes:

[0028] Discrete spatial features are mapped to dense vectors using learnable embedding encoding.

[0029] Standardize the features of the continuous space to obtain standard vectors;

[0030] The dense vector and the standard vector are concatenated to form a unified spatial background vector.

[0031] In some instances, the sequence-to-sequence large model based on the Transformer architecture includes, in sequence, an input embedding layer, a positional encoding layer, an encoder, a decoder, and an output layer; wherein:

[0032] The input embedding layer is used to map the multivariate observation vector of each time step in the dialogue history sequence to a high-dimensional vector space;

[0033] The location coding layer is used to inject the temporal location information of each time step into the output of the input embedding layer;

[0034] The encoder consists of multiple stacked encoder layers, each containing a multi-head self-attention mechanism and a feedforward neural network, used to encode a dialogue history sequence with location information into a hidden representation containing context information;

[0035] The decoder consists of multiple stacked decoder layers. Each decoder layer includes a masked multi-head self-attention mechanism, an encoder-decoder attention mechanism, and a feedforward neural network, which are used to generate a carbon flux prediction sequence for future time steps based on the output of the encoder.

[0036] The output layer is used to map the output of the decoder to the final carbon flux prediction value.

[0037] In some instances, the encoder and decoder each comprise multiple stacked layers; wherein,

[0038] The multiple stacked layers of the encoder and decoder are connected through residual and layer normalization mechanisms, respectively.

[0039] A second aspect of the present invention provides a carbon flux prediction device, comprising:

[0040] A data acquisition unit is used to acquire flux tower observation sequence data of a target area; wherein, the flux tower observation sequence data includes multimodal environmental variable observation values ​​at continuous time steps and carbon flux values ​​corresponding to the multimodal environmental variable observation values;

[0041] The data reconstruction unit is used to slide a dialog window with a time step of W on the time axis with a predetermined sliding step size, arrange the multimodal environmental variable observations in the dialog window in chronological order to form a dialog history sequence, and form the expected response sequence from the carbon flux data in the time period with a time step of P after the dialog window to obtain a dialog sample set.

[0042] The model training unit is used to input the dialogue sample set into a sequence-to-sequence large model based on the Transformer architecture for training, so as to learn the spatiotemporal dependencies between multiple variables in the dialogue history sequence through a self-attention mechanism and generate a carbon flux prediction sequence corresponding to the expected response sequence; and based on the error between the carbon flux prediction sequence and the expected response sequence, iteratively optimize the parameters of the large model through a backpropagation algorithm until the model converges, thus obtaining the trained prediction large model.

[0043] The carbon flux prediction unit is used to acquire new flux tower observation sequence data and input the dialogue history sequence constructed based on the new flux tower observation sequence data into the prediction big model to obtain the carbon flux prediction sequence for the next P time steps.

[0044] A carbon flux prediction method of the present invention includes: acquiring flux tower observation sequence data of a target region; wherein the flux tower observation sequence data includes multimodal environmental variable observations at continuous time steps and carbon flux values ​​corresponding to the multimodal environmental variable observations; reconstructing the flux tower observation sequence data into a dialogue sample set; wherein each dialogue sample contains a dialogue history sequence consisting of observation data for W consecutive time steps, and an expected response sequence consisting of carbon flux values ​​for P consecutive time steps thereafter; inputting the dialogue sample set into a sequence-to-sequence large model based on a Transformer architecture for training, so as to learn the spatiotemporal dependencies between multiple variables in the dialogue history sequence through a self-attention mechanism, and generating a carbon flux prediction sequence corresponding to the expected response sequence; based on the error between the carbon flux prediction sequence and the expected response sequence, iteratively optimizing the parameters of the large model through a backpropagation algorithm until the model converges, obtaining the trained prediction large model; acquiring new flux tower observation sequence data, and inputting the dialogue history sequence constructed based on the new flux tower observation sequence data into the prediction large model to obtain a carbon flux prediction sequence for the next P time steps. This invention transforms massive amounts of spatiotemporal data into "dialogue samples" for end-to-end training, enabling the large-scale "carbon search" model to autonomously learn and understand complex carbon cycle mechanisms without the need for pre-setting complex physical equations or empirical parameters. This gives the model stronger adaptability and generalization ability, allowing it to better adapt to different ecosystem types and environmental conditions, and effectively handle large-scale, multi-source, heterogeneous spatiotemporal data, thus surpassing existing technologies in prediction accuracy. Especially in capturing the fine-grained responses of ecosystems to climate change and extreme events, this invention provides more reliable prediction results, offering more accurate scientific evidence for carbon management and climate change response. Attached Figure Description

[0045] Figure 1 A flowchart of a carbon flux prediction method provided in an embodiment of the present invention;

[0046] Figure 2 A flowchart for constructing a dialogue sample is provided in an embodiment of the present invention;

[0047] Figure 3 This is a schematic diagram of the large model architecture provided in an embodiment of the present invention;

[0048] Figure 4 A flowchart of the large model training process provided in this embodiment of the invention;

[0049] Figure 5 This is a schematic diagram of the carbon flux prediction device provided in an embodiment of the present invention. Detailed Implementation

[0050] To better explain and facilitate understanding of the present invention, the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0051] The carbon flux prediction method proposed in this invention addresses the problem of accurately predicting carbon balance. By transforming massive spatiotemporal data into "dialogue samples" for end-to-end training, the "carbon search" large-scale model can autonomously learn and understand complex carbon cycle mechanisms without the need for pre-setting complex physical equations or empirical parameters. This gives the model stronger adaptability and generalization ability, enabling it to better adapt to different ecosystem types and environmental conditions, and effectively handle large-scale, multi-source, heterogeneous spatiotemporal data, thus surpassing existing technologies in prediction accuracy. Especially in capturing the fine-grained responses of ecosystems to climate change and extreme events, this invention provides more reliable prediction results, offering more accurate scientific evidence for carbon management and climate change response.

[0052] To better understand the above technical solutions, exemplary embodiments of the present invention will be described in more detail below with reference to the accompanying drawings. Although exemplary embodiments of the present invention are shown in the drawings, it should be understood that the present invention can be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that the present invention can be understood more clearly and thoroughly, and that the scope of the present invention can be fully conveyed to those skilled in the art.

[0053] Figure 1 This is a flowchart illustrating a carbon flux prediction method provided in an embodiment of the present invention. Figure 1 As shown, the carbon flux prediction method proposed in this embodiment of the invention includes:

[0054] Step 100: Obtain flux tower observation sequence data for the target area; wherein, the flux tower observation sequence data includes multimodal environmental variable observations at continuous time steps and carbon flux values ​​corresponding to the multimodal environmental variable observations;

[0055] Step 110: Slide a dialog window with a time step of W on the time axis with a predetermined sliding step size, arrange the multimodal environmental variable observations in the dialog window in chronological order to form a dialog history sequence, and form the expected response sequence with the carbon flux data in the time period with a time step of P after the dialog window to obtain the dialog sample set.

[0056] Step 120: Input the dialogue sample set into a sequence-to-sequence large model based on the Transformer architecture for training, so as to learn the spatiotemporal dependencies between multiple variables in the dialogue history sequence through the self-attention mechanism, and generate a carbon flux prediction sequence corresponding to the expected response sequence.

[0057] Step 130: Based on the error between the carbon flux prediction sequence and the expected response sequence, iteratively optimize the parameters of the large model using the backpropagation algorithm until the model converges, thus obtaining the trained prediction large model.

[0058] Step 140: Obtain new flux tower observation sequence data, and input the dialogue history sequence constructed based on the new flux tower observation sequence data into the prediction big model to obtain the carbon flux prediction sequence for the next P time steps; W and P are both positive integers greater than 1.

[0059] In this exemplary embodiment, the flux tower observation sequence data includes at least one of the following:

[0060] Environmental parameters: air temperature, relative humidity, wind speed, photosynthetically active radiation, soil temperature, soil moisture content, precipitation;

[0061] Ecosystem response parameters: net ecosystem exchange, gross primary productivity, and ecosystem respiration.

[0062] In this exemplary embodiment, before constructing the dialogue sample, the following is also included:

[0063] Extract derived features from the observation data, including at least one of daily average temperature, daily cumulative precipitation, seasonal indicators, and diurnal cycle indicators;

[0064] Integrate GIS spatial features, including at least one of geographical location, altitude, land cover type, vegetation index, and topographic features.

[0065] In this exemplary embodiment, the carbon-based large-scale model includes an encoder and a decoder;

[0066] The encoder uses a multi-head self-attention mechanism to encode the input sequence and outputs a context representation;

[0067] The decoder employs a masked self-attention mechanism and an encoder-decoder attention mechanism to generate a carbon flux prediction sequence for the next P time steps.

[0068] In this exemplary embodiment, the loss function used during model training is mean squared error or mean absolute error, and at least one of batch training, learning rate scheduling, Dropout regularization, or early stopping strategy is employed during optimization.

[0069] In this exemplary embodiment, after the model training is completed, the method further includes:

[0070] The generated prediction results shall be verified for compliance, including at least one of the following: comparison with measured data, uncertainty assessment, or simulation testing using engineering tools;

[0071] Based on the validation results, the model is incrementally learned or its parameters are fine-tuned.

[0072] In this exemplary embodiment, the method also supports cross-site generalization prediction by introducing site identifier embedding or spatial attention mechanisms, enabling the model to be applied to flux tower sites that were not trained.

[0073] In this exemplary embodiment, the carbon flux prediction results are used in at least one of the following application scenarios:

[0074] Short-term carbon flux early warning and ecosystem management;

[0075] Regional carbon budget assessment and carbon sink potential analysis;

[0076] Carbon flux simulation under climate change scenarios;

[0077] Carbon credit assessment and decision support in the carbon trading market.

[0078] This invention transforms massive amounts of spatiotemporal data into "dialogue samples" for end-to-end training, enabling the large-scale "carbon search" model to autonomously learn and understand complex carbon cycle mechanisms without the need for pre-setting complex physical equations or empirical parameters. This gives the model stronger adaptability and generalization ability, allowing it to better adapt to different ecosystem types and environmental conditions, and effectively handle large-scale, multi-source, heterogeneous spatiotemporal data, thus surpassing existing technologies in prediction accuracy. Especially in capturing the fine-grained responses of ecosystems to climate change and extreme events, this invention provides more reliable prediction results, offering more accurate scientific evidence for carbon management and climate change response.

[0079] In some instances, the process of sliding a dialog window with a predetermined sliding step size of W along the time axis, arranging the multimodal environmental variable observations within the dialog window in chronological order to form a dialog history sequence, includes:

[0080] Multiple dialogue windows of different scales are used to capture carbon cycle processes at different time scales, so that the observation values ​​of multimodal environmental variables within the dialogue windows are arranged in chronological order to form a dialogue history sequence; wherein, the multiple dialogue windows of different scales include a first window, a second window and a third window;

[0081] The window length W1 of the first window is set to N1 time steps to capture the intraday variation rhythm of photosynthesis and respiration.

[0082] The window length W2 of the second window is set to N2 time steps to capture the impact of weather processes on carbon flux;

[0083] The window length W3 of the third window is set to N3 time steps to capture the cumulative effect of phenological changes on carbon flux.

[0084] In this exemplary embodiment, a multi-scale dialogue sample set is constructed using multiple dialogue windows of different scales. These dialogue windows include multiple windows of different scales, each used to capture carbon cycle processes at different time scales.

[0085] The first window, with a historical window length W1 set to N1 (e.g., 24~48) time steps corresponding to 24 hours, is used to capture the intraday circadian rhythm of photosynthesis and respiration.

[0086] The second window, with a historical window length W2 set to N2 (e.g., 244~336) time steps corresponding to 7 days, is used to capture the impact of weather processes (including cold fronts, persistent droughts, and precipitation events) on carbon flux.

[0087] The third window, with a historical window length W3 set to N3 (e.g., 1000~1440) time steps corresponding to 30 days, is used to capture the cumulative effect of phenological changes (including leaf unfolding, senescence, and growing season length) on carbon flux.

[0088] The dialogue sample set includes multi-scale dialogue samples constructed based on the first window, the second window, and the third window, respectively, and are alternately or jointly input into the large model during training, enabling the model to learn spatiotemporal dependencies at different time scales simultaneously.

[0089] In this exemplary embodiment, by employing a sliding window, the model can learn the 72-hour environmental evolution pattern starting from any initial time and its corresponding carbon flux response over the next 24 hours, significantly improving the model's predictive ability under different seasons and weather conditions. Validated on a test set, the sliding window effectively improves the model's prediction accuracy compared to sampling from a fixed starting point.

[0090] In some instances, the step of inputting the dialogue sample set into a sequence-to-sequence large model based on the Transformer architecture for training includes: in the embedding layer of the sequence-to-sequence large model, when encoding the observation data at each time step, simultaneously fusing the spatial background information of the flux tower site; wherein the spatial background information is obtained by learningable embedding encoding of the site's geographical location, land cover type, and terrain features.

[0091] In this exemplary embodiment, static spatial background information is stitched together with dynamic observation data at each time step, so that spatial information runs through the entire model processing process, achieving true spatiotemporal coupling rather than simple stitching.

[0092] By using learnable embedding encoding of spatial features, the model can automatically discover the similarities and differences in carbon cycle functions among different ecosystem types, forming a spatial representation space with ecological significance.

[0093] In some instances, the learning of spatiotemporal dependencies among multiple variables in the dialogue history sequence through a self-attention mechanism includes...

[0094] The large model learns the spatiotemporal dependencies between multiple variables in the dialogue history sequence through a multi-head self-attention mechanism in the encoder; each attention head independently focuses on different feature subspaces to capture the association patterns between different types of environmental variables and carbon flux.

[0095] In this exemplary embodiment, multiple attention heads work in parallel, with each head independently focusing on different feature subspaces, enabling the simultaneous capture of various correlation patterns between environmental variables, such as temperature-humidity coupling and light-temperature synergy.

[0096] The features learned by different attention heads complement each other, together forming a rich understanding of the carbon cycle mechanism and avoiding the limitations of a single attention head.

[0097] In some instances, the attention head includes:

[0098] The first set of attention points is used to capture short-term fluctuation dependencies between adjacent time steps;

[0099] The second set of attention is used to capture weather process dependencies spanning multiple days to several weeks;

[0100] The third set of attention is used to capture long-term trend dependencies across seasons and years.

[0101] In this exemplary embodiment, signals at different scales require different processing mechanisms, with different attention heads automatically dividing the work to focus on short-term, medium-term, and long-term patterns respectively. By focusing on different time scales from hourly to yearly, the model can simultaneously capture short-term fluctuations, medium-term processes, and long-term trends in carbon flux, achieving a complete understanding of the carbon cycle. Accurately predicting multi-scale changes in carbon flux does not rely on a single model or mechanism, but rather on the automatic division of labor among multi-scale attention heads, joint training using multi-window dialogue samples, and the organic integration of local and global mechanisms, enabling the model to possess both the ability to "see details" (using a microscope) and the ability to "see trends" (using a telescope).

[0102] Short-term dependencies provide the foundation for long-term dependencies, while long-term dependencies provide the context for short-term predictions, forming a hierarchical feature extraction structure.

[0103] In some instances, when encoding the observation data at each time step in the embedding layer of the sequence-to-sequence large model, the spatial background information of the flux tower sites is simultaneously fused, including:

[0104] Encode the GIS spatial feature data into the spatial background vector;

[0105] The spatial background vector is concatenated with the observation data at each time step in the dialogue history sequence to form an enhanced dialogue history sequence with spatial context information.

[0106] In this exemplary embodiment, encoding GIS spatial feature data into a spatial background vector includes:

[0107] Discrete spatial features are mapped to dense vectors using learnable embedding encoding.

[0108] Standardize the features of the continuous space to obtain standard vectors;

[0109] The dense vector and the standard vector are concatenated to form a unified spatial background vector.

[0110] In this exemplary embodiment, different types of spatial features, such as discrete (land cover) and continuous (elevation, slope), are uniformly encoded into vector form to facilitate unified processing by the model.

[0111] Learnable embedding encoding preserves the similarity between discrete categories in the vector space; for example, the vector distance between evergreen forests and deciduous forests is smaller than the distance between forests and grasslands.

[0112] By embedding dimensions, a balance is struck between information capacity and computational efficiency (too small a dimension results in insufficient information, while too large a dimension increases the computational burden).

[0113] The spatial background vector is optimized along with other model parameters during training, and is continuously adjusted to better serve the prediction task.

[0114] In this exemplary embodiment, sliding a dialog window with a time step of W on the time axis with a predetermined sliding step size, and arranging the multimodal environmental variable observations within the dialog window in chronological order to form the dialog history sequence, includes:

[0115] For each time point t in the time series, the multivariate observation data of W consecutive time steps within the time interval [t-W+1,t] are arranged in chronological order to form the dialogue history sequence X_t=[x_{t-W+1},x_{t-W+2},...,x_t], where x_i is the observation vector of time step i, which contains multimodal environmental variables and the corresponding historical carbon flux values.

[0116] In this exemplary embodiment, the carbon flux data within a time step of P following the dialogue window is used to form the expected response sequence, including...

[0117] Arrange the carbon flux values ​​of P consecutive time steps within the time interval [t+1, t+P] in chronological order to form the expected response sequence Y_t=[y_{t+1},y_{t+2},...,y_{t+P}], where y_i is the carbon flux value at time step i.

[0118] In this exemplary embodiment, the prediction window length P of the sliding window is set to 24 time steps, corresponding to 12 hours, to provide the carbon flux prediction for the next day and support carbon sink management decisions.

[0119] In this exemplary embodiment, the flux tower observation sequence data exhibits multiple heterogeneities, including: time scale heterogeneity, spatial scale heterogeneity, and variable type heterogeneity.

[0120] Temporal heterogeneity refers to the fact that carbon cycle processes encompass everything from hourly photosynthetic responses and day-level weather events to monthly phenological changes.

[0121] Spatial scale heterogeneity refers to the significant differences in carbon flux response patterns among different ecosystem types and climate zones;

[0122] Variable heterogeneity refers to the fact that different environmental variables have different response times and lag effects on carbon flux;

[0123] The large model automatically divides the work through a multi-head self-attention mechanism, and autonomously learns to handle the multiple heterogeneities during the training process. Different attention heads focus on the dependencies of different time scales, different variable types, and different ecosystem characteristics.

[0124] Figure 2 This is a flowchart illustrating the construction of a dialogue sample according to an embodiment of the present invention. Figure 2 As shown, the dialogue sample construction process includes:

[0125] Step 20: Obtain raw flux tower observation data;

[0126] Step 21: The raw flux tower observation data undergoes data cleaning and missing value processing;

[0127] Step 22: Time series alignment and standardization of flux tower observation data;

[0128] Step 23: Extract multivariate features from flux tower observation data;

[0129] Step 24: Obtain GIS spatial data;

[0130] Step 25: Extract spatial features from GIS spatial data;

[0131] Step 26: After extracting spatial features from the GIS spatial data, fuse it with the time series data;

[0132] Step 27: The fused data is constructed using a sliding time window;

[0133] Step 28: Obtain dialogue samples.

[0134] The data acquisition and cleaning process involves obtaining raw high-frequency (e.g., 30-minute, 1-hour) observational data from global flux observation networks or local flux towers. This includes, but is not limited to, net ecosystem exchange, gross primary productivity, ecosystem respiration, air temperature, relative humidity, wind speed, photosynthetically active radiation, soil temperature, soil moisture content, and precipitation. Quality control of the raw data is performed, including outlier detection and removal, and missing value imputation (e.g., using linear interpolation, spline interpolation, or machine learning-based methods). Timestamps of all variables are aligned, and sampling frequencies are standardized. Variables with different dimensions are standardized to eliminate dimensional influences and accelerate model convergence. In addition to raw observational variables, derived features can be extracted, such as daily average temperature, daily cumulative precipitation, seasonal indicators (e.g., month, seasonal unique thermal coding), and diurnal cycle indicators. For GIS spatial data, spatial features such as geographical location (latitude and longitude), altitude, land cover type, vegetation index, and topographic features (slope, aspect) of flux tower stations can be extracted. These spatial features can serve as static contextual information and be fused with time-series data. This is the core step in constructing the "dialogue samples." A fixed-length "dialogue window" (e.g., data from the past 24 hours, 48 ​​hours, or 7 days) and a prediction step size (e.g., the next 1 hour, 3 hours, or 1 day) are defined. For each time point t, the observation data from the past W time steps (including all environmental parameters and ecosystem response data, as well as fused spatial features) are treated as a "dialogue history" or "input sequence," denoted as X_t=[x_{t-W+1},...,x_t]. Here, x_t is the multivariate observation vector at time t. The carbon fluxes from the next P time steps are treated as the "dialogue response" or "target sequence," denoted as Y_t=[y_{t+1},...,y_{t+P}]. By sliding this window across the entire time series, a large number of (X_t,Y_t) pairs, i.e., "dialogue samples," are generated. Each X_t can be seen as what the Earth system "said" in the past, while Y_t is the future "response" that we hope the model will "understand" and "predict." This sliding window mechanism enables the model to learn long-term dependencies and dynamic patterns in time-series data, and to incorporate discrete GIS spatial information as contextual information for each time step into the dialogue.

[0135] In some instances, the sequence-to-sequence large model based on the Transformer architecture includes, in sequence, an input embedding layer, a positional encoding layer, an encoder, a decoder, and an output layer; wherein:

[0136] The input embedding layer is used to map the multivariate observation vector of each time step in the dialogue history sequence to a high-dimensional vector space;

[0137] The location coding layer is used to inject the temporal location information of each time step into the output of the input embedding layer;

[0138] The encoder consists of multiple stacked encoder layers, each containing a multi-head self-attention mechanism and a feedforward neural network, used to encode a dialogue history sequence with location information into a hidden representation containing context information;

[0139] The decoder consists of multiple stacked decoder layers. Each decoder layer includes a masked multi-head self-attention mechanism, an encoder-decoder attention mechanism, and a feedforward neural network, which are used to generate a carbon flux prediction sequence for future time steps based on the output of the encoder.

[0140] The output layer is used to map the output of the decoder to the final carbon flux prediction value.

[0141] In this exemplary embodiment, the encoder progressively abstracts the input dialogue history sequence into a context-rich hidden representation through multi-layer stacking, capturing dependencies from local to global. The decoder generates future sequences based on historical context through a masking mechanism and encoder-decoder attention, ensuring the temporal coherence and causality of the predictions.

[0142] In some instances, the encoder and decoder each comprise multiple stacked layers; wherein,

[0143] The multiple stacked layers of the encoder and decoder are connected through residual and layer normalization mechanisms, respectively.

[0144] In this exemplary embodiment, residual connections provide a "highway" for gradients, enabling efficient propagation of gradients in deep networks and supporting the stacking of more layers (e.g., 12 or 24 layers) to capture more complex patterns. Key information from lower layers (e.g., "rain begins to fall on day 1") is directly transmitted to higher layers through residual connections, preventing information attenuation or loss during layer-by-layer transmission. Layer normalization stabilizes the input distribution of each layer, avoiding internal covariate shifts and significantly improving training speed and stability. The combination of these two features allows the model to be stacked with dozens of layers without degradation, providing sufficient model capacity to capture the complex patterns of the carbon cycle.

[0145] In this exemplary embodiment, the multi-head self-attention mechanism includes a heterogeneous window attention head, which is configured as follows:

[0146] For the first type of environment variables, the time dependency is captured using the first time window length;

[0147] For the second type of environment variables, the second time window length is used to capture their time dependencies;

[0148] The length of the first time window is different from that of the second time window, and is adaptively determined based on the physical characteristics of the environmental variables and their response time to carbon flux.

[0149] In this exemplary embodiment, at least one layer of the encoder includes a cross-modal attention mechanism, which is used to:

[0150] The dynamic environmental variable features in the dialogue history sequence are used as queries, and the spatial background information of the flux tower site is used as the key and value.

[0151] The attention weights between the dynamic environmental variables and spatial context information are calculated so that the model can dynamically pay attention to and integrate the corresponding spatial context information when analyzing environmental changes.

[0152] Because the input data formats and structures differ, they are referred to as heterogeneous. Heterogeneous window attention primarily enables the model to learn the weights of different heterogeneous data to be adopted in different scenarios;

[0153] Cross-modal attention is used to learn data weights between different modalities (images, grids, text).

[0154] Heterogeneous window attention addresses the heterogeneity of the temporal dimension, while cross-modal attention addresses the heterogeneity of the spatial dimension. Together, they enable the model to truly understand "what kind of carbon flux response will occur in what place and what type of environmental change sequence".

[0155] Figure 3 This is a schematic diagram of the large model architecture provided for an embodiment of the present invention. For example... Figure 3 As shown, the large model architecture includes:

[0156] The system comprises an input processing layer, a positional encoding layer, an encoder layer, a multi-head self-attention layer, a heterogeneous window attention layer, a cross-modal attention layer, an adaptive normalization layer, a temporal feedforward network layer, a decoder layer, a mask self-attention layer, and an output processing layer. Specifically, the first encoder layer is sequentially connected in series with the first multi-head attention layer, the first adaptive normalization layer, and the first temporal feedforward network. The second encoder layer is sequentially connected in series with the heterogeneous window attention layer, the second adaptive normalization layer, and the second temporal feedforward network. The third encoder layer is sequentially connected in series with the cross-modal attention layer, the third adaptive normalization layer, and the third temporal feedforward network.

[0157] This system architecture diagram illustrates the main components of the "Carbon Seeking" large-scale model. The input "dialogue sample" first passes through an embedding layer and positional encoding, then enters the encoder. The encoder internally contains a multi-head self-attention mechanism and a feedforward neural network. The encoder's output is passed to the decoder. The decoder internally contains a masked multi-head self-attention mechanism, an encoder-decoder attention mechanism, and a feedforward neural network. The decoder's output passes through an output layer to finally obtain the carbon flux prediction result.

[0158] Functions of each module:

[0159] 1. Input layer and embedding layer:

[0160] The multivariate observation vectors at each time step in the "dialogue samples" are mapped to a high-dimensional vector space to make them processable by the model. For categorical features (such as land cover type), one-hot encoding or learned embeddings can be used.

[0161] 2. Position Encoding: Since the Transformer model itself does not have the ability to process sequence order, position encoding is needed to provide the model with absolute or relative position information for each time step in the sequence. This is crucial for capturing the dynamic changes of time series.

[0162] 3. Encoder:

[0163] It consists of multiple identical encoder layers stacked together. Each encoder layer contains two main sublayers: a multi-head self-attention mechanism and a feedforward neural network. The multi-head self-attention mechanism allows the model to simultaneously attend to all other time steps in the input sequence while processing each time step, weighting them according to their correlations. This enables the model to capture complex long-term dependencies and interactions between different variables. The feedforward neural network performs a non-linear transformation on the output of the attention mechanism, increasing the model's expressive power. The encoder's role is to encode the input "dialogue history" X_t into a hidden representation containing rich contextual information.

[0164] 4. Decoder:

[0165] It consists of multiple identical decoder layers stacked together. Each decoder layer contains three main sublayers: a masked multi-head self-attention mechanism, an encoder-decoder attention mechanism, and a feedforward neural network.

[0166] Masked multi-head self-attention mechanism: Similar to the self-attention mechanism in the encoder, but with the addition of a masking mechanism to ensure that when predicting the carbon flux at the current time step, attention can only be paid to information from the current and previous time steps, avoiding information leakage. Encoder-decoder attention mechanism: Allows the decoder to pay attention to the hidden representation of the "dialogue history" output by the encoder when generating predictions, thereby effectively passing input information to the prediction process. The decoder's role is to progressively generate a carbon flux prediction sequence Y_t for future time steps based on the context information provided by the encoder and the already predicted carbon flux.

[0167] 5. Output layer:

[0168] The decoder output is mapped to the final carbon flux prediction. This can be a linear layer, with an appropriate activation function chosen based on the nature of the prediction task (regression or classification).

[0169] The training of the "Carbon Seeking" large model is an end-to-end optimization process with the goal of minimizing the error between the predicted and the true values.

[0170] Figure 4 This is a flowchart illustrating the large model training process provided in an embodiment of the present invention. Figure 4 As shown, the large model training process includes:

[0171] Step 40: Obtain the dialogue sample dataset;

[0172] Step 41: Divide the dialogue sample dataset into data segments;

[0173] Step 42: Model initialization;

[0174] Step 43: Perform iterative training on the model;

[0175] Step 44: In the training loop, perform forward propagation sequentially;

[0176] Step 45: Calculate the loss;

[0177] Step 46: Backpropagation;

[0178] Step 47: Deploy the model;

[0179] Step 48: After reaching convergence or the maximum number of iterations, perform model performance evaluation;

[0180] Step 49: Update the model parameters.

[0181] This flowchart illustrates the training and optimization process of the "Carbon Search" large-scale model. First, the "dialogue sample" dataset is divided into training, validation, and test sets. Then, model initialization is performed, entering the training loop. Within the training loop, forward propagation, loss calculation, backpropagation, and parameter updates are performed sequentially, and the loop returns to the training loop. Training ends when convergence is reached or the maximum number of iterations is reached, followed by model evaluation, and finally, model deployment.

[0182] Specific steps:

[0183] 1. Dataset Partitioning: The constructed "dialogue sample" dataset is divided into a training set, a validation set, and a test set. The training set is used for learning model parameters, the validation set is used for tuning model hyperparameters and early stopping, and the test set is used to evaluate the final performance of the model.

[0184] 2. Model initialization: Randomly initialize all trainable parameters of the model.

[0185] 3. Loss Function: For carbon flux prediction (regression task), the mean squared error or mean absolute error is usually used as the loss function to measure the difference between the predicted value and the actual value.

[0186] 4. Optimizer: Employs optimization algorithms such as Adam and SGD, calculates the gradient of the loss function with respect to the model parameters through backpropagation, and updates the model parameters based on the gradient to minimize the loss function.

[0187] 5. Training Strategies: Batch Training: Divide the training data into small batches for training to improve training efficiency and stability. Learning Rate Scheduling: Dynamically adjust the learning rate, such as using learning rate decay or learning rate warm-up, to help the model converge better. Regularization: Employ techniques such as Dropout and weight decay to prevent overfitting. Early Stopping: Stop training when the performance on the validation set no longer improves to avoid overfitting. The trained "Carbon Seeker" large model can be used to predict carbon flux on new flux tower observation data.

[0188] First, real-time or historical flux tower observation data undergoes the same data preprocessing as during the training phase, and is then used to construct "dialogue samples." These "dialogue samples" are input into the pre-trained "carbon search" large model, which outputs carbon flux predictions for the next P time steps, and finally visualizes and applies the results.

[0189] Specific steps:

[0190] 1. Data Input: Receive real-time or historical flux tower observation data.

[0191] 2. Data preprocessing: The input data is cleaned, aligned, standardized, and feature extracted in the same way as in the training phase.

[0192] 3. Construct “dialogue samples”: Using the same sliding window mechanism as in the training phase, construct “dialogue samples” for prediction from the preprocessed data.

[0193] 4. Model Inference: The constructed "dialogue samples" are input into the pre-trained "carbon search" large model. The model performs forward propagation and outputs the carbon flux prediction sequence for the next P time steps.

[0194] 5. Results Output and Application: The forecast results will be de-standardized (if previously standardized) and presented in the form of charts, reports, etc. The forecast results can be used for: short-term carbon flux early warning and management; regional carbon budget assessment; climate change scenario simulation; carbon sink potential assessment and carbon trading decision support.

[0195] This invention provides a novel and efficient carbon flux prediction solution by transforming flux tower observation data into "dialogue samples" and combining them with the powerful capabilities of large-scale models. It is expected to significantly improve prediction accuracy and model generalization ability.

[0196] This invention proposes and constructs a novel data organization paradigm for "dialogue samples," which treats continuous observational data (including multimodal environmental parameters and ecosystem response data) over a period of time around a flux tower as a continuous "dialogue" between the Earth system and the observation equipment, and uses this data as the input sequence for large-scale models. This paradigm can more effectively capture the contextual dependencies and dynamic changes in time-series data.

[0197] A dialog-window-based method for constructing "dialogue samples" details how to convert spatiotemporal training samples (including point observation data and spatial feature data) of traditional GIS models into "dialogue samples" that can slide along the time axis using a dialog-window mechanism. This method allows discrete GIS spatial information to be integrated into time-series data as contextual information for each time step, forming a unified input format.

[0198] This invention utilizes large-scale model architectures (such as encoder-decoder structures) like Transformer or its variants to perform end-to-end training on constructed "dialogue samples," achieving accurate prediction of carbon flux. This architecture can fully leverage the powerful sequence modeling and generalization capabilities of large-scale models to autonomously learn complex carbon cycle mechanisms from massive amounts of spatiotemporal data.

[0199] This invention achieves deep integration of flux tower high-frequency observation data and GIS spatial feature data. By using "dialogue samples," heterogeneous data from different sources and at different scales are uniformly input into a large model, thereby improving data utilization efficiency and the model's comprehensive predictive capabilities.

[0200] Through the aforementioned innovations, this invention can significantly improve the accuracy of carbon flux prediction, especially in capturing the fine-grained responses of ecosystems to environmental changes. Simultaneously, the application of large-scale models endows them with stronger generalization capabilities, enabling them to better adapt to different ecosystem types and environmental conditions.

[0201] This invention significantly improves the utilization efficiency of flux tower observation data and the depth of spatiotemporal correlation mining by introducing an innovative "dialogue sample" data organization paradigm. Traditional GIS models and carbon cycle models often treat time and space dimensions independently or in a simplified manner when processing spatiotemporal data, making it difficult to capture the complex nonlinear spatiotemporal dynamics and deep correlation patterns of carbon flux. This invention, however, treats continuous observation data around the flux tower over a period of time as a continuous "dialogue" with the Earth system, and combines this with a dialogue window mechanism, enabling the model to learn long-term dependencies and dynamic patterns in time-series data. Simultaneously, discrete GIS spatial information is integrated into the dialogue as contextual information for each time step. This data organization method allows large-scale models to more comprehensively and deeply understand the formation mechanism of ecosystem carbon flux and its response to environmental changes, thus overcoming the shortcomings of existing technologies in terms of low data utilization efficiency and difficulty in fully mining spatiotemporal correlations. Secondly, this invention leverages the powerful representation learning and generalization capabilities of large-scale models to significantly improve the accuracy of carbon flux prediction and the model's generalization ability. Traditional carbon cycle models typically rely on complex parameterization and empirical models, whose generalization ability is limited by the understanding of specific ecosystems, requiring extensive recalibration when applied to different regions or types of ecosystems. This invention transforms massive spatiotemporal data into "dialogue samples" for end-to-end training, enabling the "Carbon Seeking" large-scale model to autonomously learn and understand complex carbon cycle mechanisms without the need for pre-setting complex physical equations or empirical parameters. This gives the model stronger adaptability and generalization ability, allowing it to better adapt to different ecosystem types and environmental conditions, and effectively handle large-scale, multi-source, heterogeneous spatiotemporal data, thus surpassing existing technologies in prediction accuracy. Especially in capturing the fine-grained responses of ecosystems to climate change and extreme events, this invention provides more reliable prediction results, offering more accurate scientific evidence for carbon management and climate change response.

[0202] This invention provides a carbon flux prediction device. Figure 5 This is a schematic diagram of the carbon flux prediction device provided in an embodiment of the present invention. Figure 5 As shown, the carbon flux prediction device includes:

[0203] The data acquisition unit 50 is used to acquire flux tower observation sequence data of the target area; wherein, the flux tower observation sequence data includes multimodal environmental variable observation values ​​at continuous time steps and carbon flux values ​​corresponding to the multimodal environmental variable observation values;

[0204] The data reconstruction unit 51 is used to slide a dialog window with a time step of W on the time axis with a predetermined sliding step size, arrange the multimodal environmental variable observations in the dialog window in chronological order to form a dialog history sequence, and form the expected response sequence from the carbon flux data in the time period with a time step of P after the dialog window to obtain a dialog sample set.

[0205] The model training unit 52 is used to input the dialogue sample set into a sequence-to-sequence large model based on the Transformer architecture for training, so as to learn the spatiotemporal dependencies between multiple variables in the dialogue history sequence through a self-attention mechanism and generate a carbon flux prediction sequence corresponding to the expected response sequence; and based on the error between the carbon flux prediction sequence and the expected response sequence, iteratively optimize the parameters of the large model through a backpropagation algorithm until the model converges, thereby obtaining the trained prediction large model.

[0206] The carbon flux prediction unit 53 is used to acquire new flux tower observation sequence data and input the dialogue history sequence constructed based on the new flux tower observation sequence data into the prediction big model to obtain the carbon flux prediction sequence for the next P time steps.

[0207] In this exemplary embodiment, the flux tower observation sequence data includes at least one of the following:

[0208] Environmental parameters: air temperature, relative humidity, wind speed, photosynthetically active radiation, soil temperature, soil moisture content, precipitation;

[0209] Ecosystem response parameters: net ecosystem exchange, gross primary productivity, and ecosystem respiration.

[0210] In this exemplary embodiment, before constructing the dialogue sample, the following is also included:

[0211] Extract derived features from the observation data, including at least one of daily average temperature, daily cumulative precipitation, seasonal indicators, and diurnal cycle indicators;

[0212] Integrate GIS spatial features, including at least one of geographical location, altitude, land cover type, vegetation index, and topographic features.

[0213] In this exemplary embodiment, the carbon-based large-scale model includes an encoder and a decoder;

[0214] The encoder uses a multi-head self-attention mechanism to encode the input sequence and outputs a context representation;

[0215] The decoder employs a masked self-attention mechanism and an encoder-decoder attention mechanism to generate a carbon flux prediction sequence for the next P time steps.

[0216] In this exemplary embodiment, the loss function used during model training is mean squared error or mean absolute error, and at least one of batch training, learning rate scheduling, Dropout regularization, or early stopping strategy is employed during optimization.

[0217] In this exemplary embodiment, after the model training is completed, the method further includes:

[0218] The generated prediction results shall be verified for compliance, including at least one of the following: comparison with measured data, uncertainty assessment, or simulation testing using engineering tools;

[0219] Based on the validation results, the model is incrementally learned or its parameters are fine-tuned.

[0220] In this exemplary embodiment, the method also supports cross-site generalization prediction by introducing site identifier embedding or spatial attention mechanisms, enabling the model to be applied to flux tower sites that were not trained.

[0221] In this exemplary embodiment, the carbon flux prediction results are used in at least one of the following application scenarios:

[0222] Short-term carbon flux early warning and ecosystem management;

[0223] Regional carbon budget assessment and carbon sink potential analysis;

[0224] Carbon flux simulation under climate change scenarios;

[0225] Carbon credit assessment and decision support in the carbon trading market.

[0226] This invention transforms massive amounts of spatiotemporal data into "dialogue samples" for end-to-end training, enabling the large-scale "carbon search" model to autonomously learn and understand complex carbon cycle mechanisms without the need for pre-setting complex physical equations or empirical parameters. This gives the model stronger adaptability and generalization ability, allowing it to better adapt to different ecosystem types and environmental conditions, and effectively handle large-scale, multi-source, heterogeneous spatiotemporal data, thus surpassing existing technologies in prediction accuracy. Especially in capturing the fine-grained responses of ecosystems to climate change and extreme events, this invention provides more reliable prediction results, offering more accurate scientific evidence for carbon management and climate change response.

[0227] Since the systems / devices described in the above embodiments of the present invention are systems / devices used to implement the methods of the above embodiments of the present invention, those skilled in the art can understand the specific structure and modifications of the systems / devices based on the methods described in the above embodiments of the present invention, and therefore will not be repeated here. All systems / devices used in the methods of the above embodiments of the present invention fall within the scope of protection of the present invention.

[0228] In the description of this invention, it should be understood that the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0229] In this invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," "linking," and "fixing," etc., should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral part; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal communication of two components or the interaction between two components. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0230] In this invention, unless otherwise explicitly specified and limited, "above" or "below" the second feature can mean that the first and second features are in direct contact, or that they are in indirect contact through an intermediate medium. Furthermore, "above," "over," or "on top" the second feature can mean that the first feature is directly above or diagonally above the second feature, or simply indicates that the first feature is at a higher horizontal level than the second feature. "Below," "below," or "beneath" the second feature can mean that the first feature is directly below or diagonally below the second feature, or simply indicates that the first feature is at a lower horizontal level than the second feature.

[0231] In the description of this specification, the terms "one embodiment," "some embodiments," "embodiment," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0232] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make modifications, alterations, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A method for predicting carbon flux, characterized in that, include: Acquire flux tower observation sequence data for the target area; wherein, the flux tower observation sequence data includes multimodal environmental variable observations at continuous time steps and carbon flux values ​​corresponding to the multimodal environmental variable observations; With a predetermined sliding step size, a dialog window with a time step size of W is slid along the time axis. The observation values ​​of multimodal environmental variables in the dialog window are arranged in chronological order to form a dialog history sequence. The carbon flux data in the time period with a time step size of P after the dialog window are used to form the expected response sequence, thus obtaining a dialog sample set. The dialogue sample set is input into a sequence-to-sequence model based on the Transformer architecture for training, so as to learn the spatiotemporal dependencies between multiple variables in the dialogue history sequence through the self-attention mechanism and generate a carbon flux prediction sequence corresponding to the expected response sequence. Based on the error between the carbon flux prediction sequence and the expected response sequence, the parameters of the large model are iteratively optimized through the backpropagation algorithm until the model converges, thus obtaining the trained prediction large model. New flux tower observation sequence data is acquired, and the dialogue history sequence constructed based on the new flux tower observation sequence data is input into the prediction big model to obtain the carbon flux prediction sequence for the next P time steps.

2. The carbon flux prediction method as described in claim 1, characterized in that, The process of sliding a dialog window with a predetermined sliding step size of W on the time axis, arranging the multimodal environmental variable observations within the dialog window in chronological order to form a dialog history sequence, includes: Multiple dialogue windows of different scales are used to capture carbon cycle processes at different time scales, so that the observation values ​​of multimodal environmental variables within the dialogue windows are arranged in chronological order to form a dialogue history sequence; wherein, the multiple dialogue windows of different scales include a first window, a second window and a third window; The window length W1 of the first window is set to N1 time steps to capture the intraday variation rhythm of photosynthesis and respiration. The window length W2 of the second window is set to N2 time steps to capture the impact of weather processes on carbon flux; The window length W3 of the third window is set to N3 time steps to capture the cumulative effect of phenological changes on carbon flux.

3. The carbon flux prediction method as described in claim 1, characterized in that, The step of inputting the dialogue sample set into a sequence-to-sequence large model based on the Transformer architecture for training includes: in the embedding layer of the sequence-to-sequence large model, when encoding the observation data at each time step, simultaneously fusing the spatial background information of the flux tower site; wherein the spatial background information is obtained by learning-based embedding encoding of the site's geographical location, land cover type, and terrain features.

4. The carbon flux prediction method as described in claim 1, characterized in that, The method of learning the spatiotemporal dependencies among multiple variables in the dialogue history sequence through a self-attention mechanism includes... The large model learns the spatiotemporal dependencies among multiple variables in the dialogue history sequence through a multi-head self-attention mechanism in the encoder. Each attention head independently focuses on a different feature subspace to capture the correlation patterns between different types of environmental variables and carbon flux.

5. The carbon flux prediction method as described in claim 4, characterized in that, The attention head includes: The first set of attention points is used to capture short-term fluctuation dependencies between adjacent time steps; The second set of attention is used to capture weather process dependencies spanning multiple days to several weeks; The third set of attention is used to capture long-term trend dependencies across seasons and years.

6. The carbon flux prediction method as described in claim 3, characterized in that, In the embedding layer of the sequence-to-sequence large model, when encoding the observation data at each time step, the spatial background information of the flux tower site is simultaneously fused, including: Encode the GIS spatial feature data into the spatial background vector; The spatial background vector is concatenated with the observation data at each time step in the dialogue history sequence to form an enhanced dialogue history sequence with spatial context information.

7. The carbon flux prediction method as described in claim 6, characterized in that, The process of encoding GIS spatial feature data into spatial background vectors includes: Discrete spatial features are mapped to dense vectors using learnable embedding encoding. Standardize the features of the continuous space to obtain standard vectors; The dense vector and the standard vector are concatenated to form a unified spatial background vector.

8. The carbon flux prediction method as described in claim 6, characterized in that, The sequence-to-sequence large model based on the Transformer architecture includes, in sequence, an input embedding layer, a positional encoding layer, an encoder, a decoder, and an output layer; wherein: The input embedding layer is used to map the multivariate observation vector of each time step in the dialogue history sequence to a high-dimensional vector space; The location coding layer is used to inject the temporal location information of each time step into the output of the input embedding layer; The encoder consists of multiple stacked encoder layers, each containing a multi-head self-attention mechanism and a feedforward neural network, used to encode a dialogue history sequence with location information into a hidden representation containing context information; The decoder consists of multiple stacked decoder layers. Each decoder layer includes a masked multi-head self-attention mechanism, an encoder-decoder attention mechanism, and a feedforward neural network, which are used to generate a carbon flux prediction sequence for future time steps based on the output of the encoder. The output layer is used to map the output of the decoder to the final carbon flux prediction value.

9. The carbon flux prediction method as described in claim 8, characterized in that, The encoder and decoder each comprise multiple stacked layers; wherein, The multiple stacked layers of the encoder and decoder are connected through residual and layer normalization mechanisms, respectively.

10. A carbon flux prediction device, characterized in that, include: A data acquisition unit is used to acquire flux tower observation sequence data of a target area; wherein, the flux tower observation sequence data includes multimodal environmental variable observation values ​​at continuous time steps and carbon flux values ​​corresponding to the multimodal environmental variable observation values; The data reconstruction unit is used to slide a dialog window with a time step of W on the time axis with a predetermined sliding step size, arrange the multimodal environmental variable observations in the dialog window in chronological order to form a dialog history sequence, and form the expected response sequence from the carbon flux data in the time period with a time step of P after the dialog window to obtain a dialog sample set. The model training unit is used to input the dialogue sample set into a sequence-to-sequence large model based on the Transformer architecture for training, so as to learn the spatiotemporal dependencies between multiple variables in the dialogue history sequence through a self-attention mechanism and generate a carbon flux prediction sequence corresponding to the expected response sequence; and based on the error between the carbon flux prediction sequence and the expected response sequence, iteratively optimize the parameters of the large model through a backpropagation algorithm until the model converges, thus obtaining the trained prediction large model. The carbon flux prediction unit is used to acquire new flux tower observation sequence data and input the dialogue history sequence constructed based on the new flux tower observation sequence data into the prediction big model to obtain the carbon flux prediction sequence for the next P time steps.

Images