A deep learning-based gcm-like temperature field prediction method and system
By employing a deep learning-based GCM-like temperature field prediction method, utilizing a multi-channel spatial encoder, long short-term memory network, and multilayer perceptron, combined with a carbon dioxide weight map, the high computational cost and model matching issues of GCMs are resolved, achieving high-precision and low-cost climate field prediction.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- INST OF GEOGRAPHICAL SCI & NATURAL RESOURCE RES CAS
- Filing Date
- 2026-02-24
- Publication Date
- 2026-06-12
Smart Images
Figure CN122196495A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of temperature prediction technology, and specifically to a deep learning-based GCM-like temperature field prediction method and system. Background Technology
[0002] Accurate prediction of the global temperature field is crucial for applications such as climate science research, future scenario projection, extreme event attribution, and online data assimilation in paleoclimate reconstruction. Currently, the mainstream prediction tools in this field are physical first-principles climate system models (GCMs). These models simulate the evolution of the Earth's climate system by solving complex fluid dynamics and thermodynamic equations, possessing a solid physical foundation. However, GCMs have a fundamental limitation: their computational cost is extremely high. A single simulation typically requires running on high-performance computing clusters for days or even weeks, and this enormous computational overhead severely restricts their application in scenarios requiring large-scale, rapid generation of background fields.
[0003] In recent years, artificial intelligence technologies, represented by deep learning, have provided a new solution for building rapid proxy models of paleoclimate climate (GCM). Among them, the full-element AI meteorological model has demonstrated the ability to complete global weather forecasts within seconds, realizing a paradigm shift from "equation-driven" to "data and knowledge fusion-driven". However, full-element models have large parameters and are difficult to apply directly due to incompatibility with paleoclimate proxy data, while single-element models are difficult to effectively respond to key external forcing signals and suffer from spatiotemporal coupling defects. Summary of the Invention
[0004] (a) Purpose of the invention The purpose of this invention is to provide a deep learning-based GCM-like temperature field prediction method and system that can improve the accuracy of spatiotemporal prediction of specific climate fields in scenarios with limited input variables and sparse data. It can understand and respond to physical mechanisms, and has strong generalization ability and interpretability.
[0005] (II) Technical Solution To address the above problems, this invention provides a deep learning-based GCM-like temperature field prediction method, comprising: The input variables for the prediction are determined, including surface temperature, carbon dioxide concentration, and aerosol optical thickness; Using a preset multi-channel spatial encoder, feature extraction is performed on the input variables to obtain surface temperature features, carbon dioxide concentration features, and aerosol optical thickness features. Based on the preset long short-term memory network, the surface temperature characteristics, carbon dioxide concentration characteristics, and aerosol optical thickness characteristics, a temporal context vector is obtained; Based on the preset multilayer perceptron and the carbon dioxide concentration characteristics, a radiation forcing embedding vector is obtained; The first predicted temperature field is obtained based on the preset decoder network, the temporal context vector, and the radiation forcing embedding vector.
[0006] In another aspect of the present invention, preferably, the method further includes: using a preset carbon dioxide weighting map to correct the first predicted temperature field to obtain a second predicted temperature field.
[0007] In another aspect of the present invention, preferably, the step of using a preset multi-channel spatial encoder to extract features from the input variables to obtain surface temperature features, carbon dioxide concentration features, and aerosol optical thickness features includes: The preset multi-channel spatial encoder includes a first feature extractor, a second feature extractor, and a third feature extractor; The surface temperature, carbon dioxide concentration, and aerosol optical thickness are synchronized in time and divided according to a preset time step to obtain time-aligned surface temperature, carbon dioxide concentration, and aerosol optical thickness sequences. The surface temperature sequence is input into the first feature extractor to obtain the surface temperature feature sequence. The carbon dioxide concentration sequence is input into the second feature extractor to obtain the surface temperature feature sequence. The aerosol optical thickness sequence is input into the third feature extractor to obtain the aerosol optical thickness feature sequence.
[0008] In another aspect of the present invention, preferably, obtaining the temporal context vector based on a preset long short-term memory network, the surface temperature characteristics, carbon dioxide concentration characteristics, and aerosol optical thickness characteristics includes: The surface temperature feature sequence, carbon dioxide concentration feature sequence, and aerosol optical thickness feature sequence are spliced together according to time to obtain a fused feature sequence; The fused feature sequence is sequentially input into a preset long short-term memory network, and a temporal context vector is output at the final time step.
[0009] In another aspect of the present invention, preferably, obtaining the radiation forcing embedding vector based on a preset multilayer perceptron and the carbon dioxide concentration features includes: Obtain the carbon dioxide concentration features of the last time step in the carbon dioxide concentration feature sequence; The carbon dioxide concentration features of the last time step are flattened to obtain a one-dimensional feature vector. Using a pre-defined multilayer perceptron, the one-dimensional feature vector is nonlinearly mapped to obtain a radiative forcing embedding vector.
[0010] In another aspect of the present invention, preferably, obtaining the first predicted temperature field based on the preset decoder network, the temporal context vector, and the radiative forcing embedding vector includes: The temporal context vector and the radiative forcing embedding vector are concatenated to obtain the climate feature vector; The climate feature vector is upsampled step by step based on a preset decoder network to obtain the first predicted temperature field.
[0011] In another aspect of the present invention, preferably, the preset carbon dioxide weight map is a trained response weight map that has the same spatial dimension as the first predicted temperature field. The training of the response weight map includes: The response weight map is initialized based on prior climatological knowledge; The response weight map is iteratively optimized using backpropagation and gradient descent methods to obtain a preset carbon dioxide weight map.
[0012] In another aspect of the present invention, preferably, the step of correcting the first predicted temperature field using a preset carbon dioxide weighting map to obtain a second predicted temperature field includes: The second predicted temperature field is obtained by processing the preset carbon dioxide weight map and the first predicted temperature field grid by grid point.
[0013] In another aspect of the present invention, preferably, the method further includes: The multi-channel spatial encoder, long short-term memory network, decoder network, and carbon dioxide weight map are trained according to a preset training strategy. The preset training strategy includes: In the first stage, the decoder network and carbon dioxide weight map are unfrozen and trained. In the second stage, the Long Short-Term Memory network is unfrozen and trained. In the third stage, the multi-channel spatial encoder is unfrozen and trained. In the fourth stage, the multi-channel spatial encoder, long short-term memory network, decoder network, and carbon dioxide weight map are unfrozen and trained. Different learning rates are used in each stage, and the learning rate gradually decreases as the stage progresses.
[0014] In another aspect, preferably, a deep learning-based GCM-like temperature field prediction system includes: Determining module: Determines the input variables for prediction, including surface temperature, carbon dioxide concentration, and aerosol optical thickness; Feature extraction module: Using a preset multi-channel spatial encoder, features are extracted from the input variables to obtain surface temperature features, carbon dioxide concentration features, and aerosol optical thickness features; The first acquisition module obtains a temporal context vector based on a preset long short-term memory network, the surface temperature characteristics, carbon dioxide concentration characteristics, and aerosol optical thickness characteristics. The second acquisition module obtains a radiation forcing embedding vector based on a preset multilayer perceptron and the carbon dioxide concentration features. Prediction module: Based on the preset decoder network, the temporal context vector and the radiation forcing embedding vector, the first predicted temperature field is obtained.
[0015] (III) Beneficial Effects The above-described technical solution of the present invention has the following beneficial technical effects: This invention integrates key climate elements such as surface temperature, carbon dioxide concentration, and aerosol optical thickness to comprehensively reflect the impact of different external forcings on the temperature field, achieving higher prediction accuracy with fewer input variables. By embedding a carbon dioxide radiative forcing vector, the physical mechanism of key external forcings is reflected during the prediction process, making the carbon dioxide radiative forcing signal significant at the feature level. This avoids it being submerged by high-frequency fluctuations in time-series modeling, thus effectively compensating for the model's response capability to long-term forcing. Attached Figure Description
[0016] Figure 1 This is an overall flowchart of one embodiment of the present invention; Figure 2 This is a schematic diagram of the overall structure of one embodiment of the present invention. Detailed Implementation
[0017] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments and the accompanying drawings. It should be understood that these descriptions are merely exemplary and not intended to limit the scope of the invention. Furthermore, descriptions of well-known structures and techniques are omitted in the following description to avoid unnecessarily obscuring the concept of the invention.
[0018] Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.
[0019] In the description of this invention, it should be noted that the terms "first," "second," and "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.
[0020] Furthermore, the technical features involved in the different embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.
[0021] The invention will now be described in more detail with reference to the accompanying drawings. In the various drawings, the same elements are indicated by similar reference numerals. For clarity, the various parts in the drawings are not drawn to scale.
[0022] Example 1 A deep learning-based GCM-like temperature field prediction method Figure 1 An overall flowchart of one embodiment of the present invention is shown. Figure 2 A schematic diagram of the overall structure of an embodiment of the present invention is shown, as follows. Figure 1 and Figure 2 As shown, it includes: The input variables for prediction are determined, including surface temperature, carbon dioxide concentration, and aerosol optical thickness; here, surface temperature is the temperature at two meters above the surface, and aerosol optical thickness is the optical thickness of volcanic aerosols. The surface temperature, carbon dioxide concentration, and aerosol optical thickness at multiple consecutive time steps are used as input variables to obtain the global temperature field at the next time step.
[0023] Using a preset multi-channel spatial encoder, feature extraction is performed on the input variables to obtain surface temperature features, carbon dioxide concentration features, and aerosol optical thickness features. Specifically, the multi-channel spatial encoder extracts features from the input variables to obtain surface temperature features, carbon dioxide concentration features, and aerosol optical thickness features, including: The preset multi-channel spatial encoder includes a first feature extractor, a second feature extractor, and a third feature extractor; each feature extractor can employ a convolutional network, an attention mechanism, or other suitable deep neural network structure to extract the features of the input variables in the spatial dimension, while retaining the spatial distribution information of each variable.
[0024] The surface temperature, carbon dioxide concentration, and aerosol optical thickness are synchronized in time and divided according to a preset time step to obtain time-aligned surface temperature, carbon dioxide concentration, and aerosol optical thickness sequences. The time alignment operation ensures the correspondence of each input variable within the same time step.
[0025] The surface temperature sequence is input into the first feature extractor to obtain a surface temperature feature sequence, which can reflect the local structure, pattern and evolution trend of the temperature field in the spatial dimension.
[0026] The carbon dioxide concentration sequence is input into the second feature extractor to obtain the surface temperature feature sequence, which can capture the distribution differences of carbon dioxide in different spatial regions and its potential driving effect on the temperature field.
[0027] The aerosol optical thickness sequence is input into the third feature extractor. The aerosol optical thickness feature sequence can reflect the spatial distribution characteristics of aerosols and their moderating effect on radiation and temperature fields.
[0028] For example, consider an independent residual network encoder, such as ResNet18, configured for each of the three input variables. These encoders run in parallel at each time step t. For any one of the input variables... , (where v∈{ The corresponding encoder maps it to a low-dimensional feature vector: in, This represents the feature of the input variable v at time step t. Let H represent the feature extractor for the input variable v, where H and W are the height and width of the input variable, respectively. This refers to the dimension of the encoded feature vector. By extracting features from the three input variables separately, specific features of different input variables—temperature (system state), CO2 (long-term forcing), and AOD (short-term impulse forcing)—were achieved, laying the foundation for subsequent accurate modeling of their different impacts on the prediction results.
[0029] Based on a pre-defined long short-term memory (LSTM) network, the surface temperature characteristics, carbon dioxide concentration characteristics, and aerosol optical thickness characteristics, a temporal context vector is obtained. The LSTM network can capture the evolution of climate variables over time, enabling temporal dependency modeling of the temperature field while considering both long-term trends and short-term fluctuations. In this embodiment, the pre-defined LSTM network is a multi-layer LSTM network. Based on the pre-defined LSTM network, the surface temperature characteristics, carbon dioxide concentration characteristics, and aerosol optical thickness characteristics, the temporal context vector is obtained, including: The surface temperature feature sequence, carbon dioxide concentration feature sequence, and aerosol optical thickness feature sequence are spliced together according to time to obtain a fused feature sequence; The fused feature sequences are sequentially input into a pre-defined long short-term memory (LSTM) network, and a temporal context vector is output at the final time step. This temporal context vector is a compressed climate temporal context vector, containing comprehensive dynamic evolution information of the climate system over the past T time steps. The LSM network effectively learns the internal variability of the climate system (such as the quasi-periodicity of modes like ENSO and NAO) and its time-lag response to external forcing.
[0030] Based on a preset multilayer perceptron and the carbon dioxide concentration characteristics, a radiative forcing embedding vector is obtained. This radiative forcing embedding vector addresses the problem of insufficient learning by long short-term memory networks for low-frequency, small-amplitude but significantly cumulative CO2 forcing signals, enhancing sensitivity to key human-induced forcing. In this embodiment, obtaining the radiative forcing embedding vector based on a preset multilayer perceptron and the carbon dioxide concentration characteristics includes: Obtain the carbon dioxide concentration feature of the last time step of the carbon dioxide concentration feature sequence; here, the carbon dioxide concentration feature of the last time step in the input sequence is processed separately.
[0031] The carbon dioxide concentration features at the last time step are flattened to obtain a one-dimensional feature vector. Flattening can be achieved through vectorization operations, forming a unified representation suitable for fully connected network inputs while preserving the original feature information.
[0032] A pre-defined multilayer perceptron is used to perform a nonlinear mapping on the one-dimensional feature vector to obtain a radiative forcing embedding vector. This pre-defined multilayer perceptron is independent of the long short-term memory network and includes two fully connected layers. It transforms the spatial distribution information of carbon dioxide concentration into a compact high-level semantic representation to characterize the radiative forcing effect of carbon dioxide changes on the temperature field. This can be expressed using the following formula: in, Represents the radiative forcing embedding vector. This indicates the carbon dioxide concentration characteristics at the last time step.
[0033] By processing carbon dioxide concentration characteristics separately, an independent and enhanced external forcing channel is provided, which makes the radiation forcing signal of carbon dioxide significant at the feature level, avoiding its being submerged by high-frequency fluctuations in the long short-term memory network, thus effectively supplementing the response capability to long-term forcing.
[0034] Based on a preset decoder network, the temporal context vector, and the radiation forcing embedding vector, a first predicted temperature field is obtained. The decoder can use upsampling, deconvolution, or attention mechanisms to convert the high-dimensional feature vector into a temperature field output with the same spatial resolution as the target. Further, in this embodiment, obtaining the first predicted temperature field based on the preset decoder network, the temporal context vector, and the radiation forcing embedding vector includes: The climate feature vector is obtained by concatenating the temporal context vector and the radiative forcing embedding vector, expressed by the following formula: in, Represents the climate feature vector. h T Represents the time-series context vector. This represents the radiative forcing embedding vector.
[0035] The climate feature vector is upsampled stepwise based on a pre-defined decoder network to obtain the first predicted temperature field. The pre-defined decoder network consists of multiple layers of transposed convolutions. The decoder performs a stepwise upsampling process, first mapping the feature vector to a low-resolution pattern at the planetary scale, then gradually refining the regional scale and local features through intermediate layers, and finally restoring the resolution to the target grid (e.g., 2°×2°), outputting the first predicted temperature field. This achieves high-fidelity reconstruction from abstract features to specific spatial distribution, while also incorporating information from the system's internal dynamics, i.e., the time-series context vector and the external forcing background, i.e., the radiative forcing embedding vector.
[0036] Furthermore, this embodiment also includes: using a preset carbon dioxide weight map to correct the first predicted temperature field to obtain a second predicted temperature field. The preset carbon dioxide weight map is a trained response weight map with the same spatial dimension as the first predicted temperature field, and the training of the response weight map includes: The response weight map is initialized based on prior climatological knowledge. Latitude-weighted initialization of the response weight map is also performed using prior climatological knowledge, that is, higher initial weights are assigned to high-latitude regions and lower weights to low-latitude regions, in order to reflect the polar amplification effect.
[0037] The response weight map is iteratively optimized using backpropagation and gradient descent methods to obtain a preset carbon dioxide weight map. This map can be adaptively fine-tuned based on the data to learn more accurate regional sensitivity, thereby effectively alleviating the data drift problem at different carbon dioxide concentration levels, such as before and after the Industrial Revolution, and significantly improving cross-period generalization ability and long-term stability.
[0038] The correction involves mapping the preset carbon dioxide weight map to the first predicted temperature field grid by grid to obtain the second predicted temperature field. The process is represented by the following formula: in, This represents the adjustment function, which is equivalent to applying a spatially heterogeneous bias term. This represents the second predicted temperature field. This represents the first predicted temperature field. The key physical knowledge of how the carbon dioxide forcing response varies with latitude is directly encoded into the model structure, ensuring that the model's predictions spatially conform to basic climatological understanding and addressing the issue of unreliable physics in black-box models. Furthermore, the visualization of the carbon dioxide weighting map provides an intuitive physical window into the analysis of the decision-making mechanism.
[0039] Furthermore, the method also includes: The multi-channel spatial encoder, long short-term memory network, decoder network, and carbon dioxide weight map are trained according to a preset training strategy. During training, all input variables are resampled to a uniform grid (e.g., 2°×2°) and subjected to minimum-maximum normalization. Training and test sets are constructed based on a mode membership partitioning strategy to ensure the reliability of generalization ability evaluation. Mean squared error (MSE) is used as the optimization objective to directly minimize the grid error between the predicted temperature field and the true field. Dropout, early stopping, and adaptive learning rate scheduling strategies are used in combination to prevent overfitting and promote model convergence to a better generalization solution.
[0040] The preset training strategy includes: In the first stage, the decoder network and carbon dioxide weight map are unfrozen and trained to quickly adapt to the systematic biases of the new dataset; In the second stage, the Long Short-Term Memory network is unfrozen and trained to learn specific internal variable dynamic sequences (such as real ENSO evolution) in new data. In the third stage, the multi-channel spatial encoder is unfrozen, trained, and the large-scale circulation and teleconnection modes are adjusted. In the fourth stage, the multi-channel spatial encoder, long short-term memory network, decoder network and carbon dioxide weight map are unfrozen and trained, and global fine-tuned with an extremely low learning rate to achieve optimal integration. Different learning rates are used in each stage, and the learning rate gradually decreases as the stage progresses.
[0041] The training strategy in this embodiment can quickly and efficiently adapt the pre-trained model to new observational or reanalysis data using a very small number of samples (hundreds), significantly reducing prediction errors (e.g., 38.8%), demonstrating its practical value as a "base model." To prevent catastrophic forgetting, a phased, physically guided unfreezing strategy ensures that the model absorbs new data features without forgetting the general physical laws learned from massive amounts of simulation data. The physically meaningful optimization process links the transfer learning process in machine learning to the physical level of the climate system, making the optimization process more controllable and reliable.
[0042] Example 2 A deep learning-based GCM-like temperature field prediction system includes: Determining module: Determines the input variables for prediction, including surface temperature, carbon dioxide concentration, and aerosol optical thickness; Feature extraction module: Using a preset multi-channel spatial encoder, features are extracted from the input variables to obtain surface temperature features, carbon dioxide concentration features, and aerosol optical thickness features; The first acquisition module obtains a temporal context vector based on a preset long short-term memory network, the surface temperature characteristics, carbon dioxide concentration characteristics, and aerosol optical thickness characteristics. The second acquisition module obtains a radiation forcing embedding vector based on a preset multilayer perceptron and the carbon dioxide concentration features. Prediction module: Based on the preset decoder network, the temporal context vector and the radiation forcing embedding vector, the first predicted temperature field is obtained.
[0043] It should be understood that the specific embodiments described above are merely illustrative or explanatory of the principles of the invention and do not constitute a limitation thereof. Therefore, any modifications, equivalent substitutions, improvements, etc., made without departing from the spirit and scope of the invention should be included within the protection scope of the invention. Furthermore, the appended claims are intended to cover all variations and modifications falling within the scope and boundaries of the appended claims, or equivalent forms of such scope and boundaries.
[0044] The present invention has been described above with reference to embodiments thereof. However, these embodiments are merely illustrative and not intended to limit the scope of the invention. The scope of the invention is defined by the appended claims and their equivalents. Various substitutions and modifications can be made by those skilled in the art without departing from the scope of the invention, and all such substitutions and modifications should fall within the scope of the invention.
[0045] Although embodiments of the present invention have been described in detail, it should be understood that various changes, substitutions, and modifications can be made to the embodiments of the present invention without departing from the spirit and scope of the invention.
[0046] 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 deep learning-based GCM-like temperature field prediction method, characterized in that, include: The input variables for the prediction are determined, including surface temperature, carbon dioxide concentration, and aerosol optical thickness; Using a preset multi-channel spatial encoder, feature extraction is performed on the input variables to obtain surface temperature features, carbon dioxide concentration features, and aerosol optical thickness features. Based on the preset long short-term memory network, the surface temperature characteristics, carbon dioxide concentration characteristics, and aerosol optical thickness characteristics, a temporal context vector is obtained; Based on the preset multilayer perceptron and the carbon dioxide concentration characteristics, a radiation forcing embedding vector is obtained; The first predicted temperature field is obtained based on the preset decoder network, the temporal context vector, and the radiation forcing embedding vector.
2. The deep learning-based GCM-like temperature field prediction method according to claim 1, characterized in that, The method further includes: using a preset carbon dioxide weighting map to correct the first predicted temperature field to obtain a second predicted temperature field.
3. The deep learning-based GCM-like temperature field prediction method according to claim 1, characterized in that, The process involves using a pre-set multi-channel spatial encoder to extract features from the input variables, obtaining surface temperature features, carbon dioxide concentration features, and aerosol optical thickness features, including: The preset multi-channel spatial encoder includes a first feature extractor, a second feature extractor, and a third feature extractor; The surface temperature, carbon dioxide concentration, and aerosol optical thickness are synchronized in time and divided according to a preset time step to obtain time-aligned surface temperature, carbon dioxide concentration, and aerosol optical thickness sequences. The surface temperature sequence is input into the first feature extractor to obtain the surface temperature feature sequence. The carbon dioxide concentration sequence is input into the second feature extractor to obtain the surface temperature feature sequence. The aerosol optical thickness sequence is input into the third feature extractor to obtain the aerosol optical thickness feature sequence.
4. The deep learning-based GCM-like temperature field prediction method according to claim 3, characterized in that, The temporal context vector is obtained based on the preset long short-term memory network, the surface temperature characteristics, carbon dioxide concentration characteristics, and aerosol optical thickness characteristics, including: The surface temperature feature sequence, carbon dioxide concentration feature sequence, and aerosol optical thickness feature sequence are spliced together according to time to obtain a fused feature sequence; The fused feature sequence is sequentially input into a preset long short-term memory network, and a temporal context vector is output at the final time step.
5. The deep learning-based GCM-like temperature field prediction method according to claim 4, characterized in that, The process of obtaining a radiation forcing embedding vector based on a preset multilayer perceptron and the carbon dioxide concentration features includes: Obtain the carbon dioxide concentration features of the last time step in the carbon dioxide concentration feature sequence; The carbon dioxide concentration features of the last time step are flattened to obtain a one-dimensional feature vector. Using a pre-defined multilayer perceptron, the one-dimensional feature vector is nonlinearly mapped to obtain a radiative forcing embedding vector.
6. The deep learning-based GCM-like temperature field prediction method according to claim 5, characterized in that, The first predicted temperature field is obtained based on the preset decoder network, the temporal context vector, and the radiation forcing embedding vector, including: The temporal context vector and the radiative forcing embedding vector are concatenated to obtain the climate feature vector; The climate feature vector is upsampled step by step based on a preset decoder network to obtain the first predicted temperature field.
7. The deep learning-based GCM-like temperature field prediction method according to claim 6, characterized in that, The preset carbon dioxide weight map is a trained response weight map that has the same spatial dimension as the first predicted temperature field. The training of the response weight map includes: The response weight map is initialized based on prior climatological knowledge; The response weight map is iteratively optimized using backpropagation and gradient descent methods to obtain a preset carbon dioxide weight map.
8. The deep learning-based GCM-like temperature field prediction method according to claim 7, characterized in that, The step of correcting the first predicted temperature field using a preset carbon dioxide weighting map to obtain the second predicted temperature field includes: The second predicted temperature field is obtained by processing the preset carbon dioxide weight map and the first predicted temperature field grid by grid point.
9. The deep learning-based GCM-like temperature field prediction method according to claim 8, characterized in that, The method further includes: The multi-channel spatial encoder, long short-term memory network, decoder network, and carbon dioxide weight map are trained according to a preset training strategy. The preset training strategy includes: In the first stage, the decoder network and carbon dioxide weight map are unfrozen and trained. In the second stage, the Long Short-Term Memory network is unfrozen and trained. In the third stage, the multi-channel spatial encoder is unfrozen and trained. In the fourth stage, the multi-channel spatial encoder, long short-term memory network, decoder network, and carbon dioxide weight map are unfrozen and trained. Different learning rates are used in each stage, and the learning rate gradually decreases as the stage progresses.
10. A deep learning-based GCM-like temperature field prediction system, characterized in that, include: Determining module: Determines the input variables for prediction, including surface temperature, carbon dioxide concentration, and aerosol optical thickness; Feature extraction module: Using a preset multi-channel spatial encoder, features are extracted from the input variables to obtain surface temperature features, carbon dioxide concentration features, and aerosol optical thickness features; The first acquisition module obtains a temporal context vector based on a preset long short-term memory network, the surface temperature characteristics, carbon dioxide concentration characteristics, and aerosol optical thickness characteristics. The second acquisition module obtains a radiation forcing embedding vector based on a preset multilayer perceptron and the carbon dioxide concentration features. Prediction module: Based on the preset decoder network, the temporal context vector and the radiation forcing embedding vector, the first predicted temperature field is obtained.