A method for predicting nuclear diffusion weather fields by fusing regional self-attention and modified ConvLSTM
By integrating regional self-attention and modified ConvLSTM, the problem of insufficient foresight and global synergy in nuclear emergency meteorological field modeling is solved, achieving efficient nuclear diffusion meteorological field prediction, which is suitable for rapid decision support in nuclear accident emergency scenarios.
Patent Information
- Application Number
- CN202611610179.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-10-19
- Publication Date
- 2026-11-17
Smart Images

Figure CN123116875A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of nuclear accident emergency technology, specifically to a method for predicting nuclear diffusion meteorological fields by fusing regional self-attention and modified ConvLSTM. Background Technology
[0002] In nuclear accident emergency response, risk assessment is required based on the dose field of nuclide diffusion. The dose field is often derived from the diffusion of nuclides in the meteorological field. For the simulation of the meteorological field, conventional methods rely on real-time observation data such as air pressure, humidity, wind speed, and wind direction obtained from ground meteorological stations and upper-air sounding platforms. The meteorological field is constructed through physical calculations such as atmospheric dynamics equations and turbulence parameterization. These methods are usually limited by the spatiotemporal resolution of the monitoring data and can only reconstruct the meteorological field environment on a macroscopic scale. When used as a driving condition for nuclide diffusion, it is difficult to grasp the future trend of change.
[0003] In the field of meteorological field simulation and modeling, existing technologies mainly include several mainstream technical solutions such as Gaussian plume models, Lagrange particle diffusion models, Eulerian grid models, and mesoscale meteorological model downscaling techniques. The following will briefly describe the existing technical solutions most relevant to this application, and provide technical solution diagrams and detailed textual descriptions.
[0004] 1) Gaussian plume model: A simplified process that collects meteorological and source term parameters, calculates diffusion parameters and effective source height, solves for concentration and dose, and outputs the results, enabling rapid early assessment of nuclear emergencies. The specific process is as follows: Figure 1 As shown in the figure. This method is suitable for early emergency scenarios with flat terrain and stable atmosphere. However, it has the drawback of large errors under complex terrain and unsteady weather conditions, and cannot accurately simulate complex airflow. It needs to be optimized in combination with monitoring data. It is mainly used for quickly delineating evacuation areas.
[0005] 2) Lagrange particle diffusion model: After initialization, virtual particles are released and tracked, then the physical process is simulated and the dosage is calculated. The results are then analyzed to simulate pollutant diffusion. The specific process is as follows: Figure 2 As shown. This method has a clear physical process and strong terrain adaptability, making it suitable for medium-range nuclear emergency assessment under complex terrain and non-uniform weather conditions. However, this method has high computational resource requirements, and the number of particles affects the accuracy, resulting in low efficiency in long-term simulations.
[0006] 3) Eulerian mesh model: This model sets up the mesh and parameters to solve the governing equations, then integrates multi-physics processes to calculate the dose, and finally outputs the decision results to achieve large-scale nuclear emergency assessment. The specific process is as follows: Figure 3As shown. This method can couple multiple physics processes and is suitable for long-duration, large-scale nuclear emergency scenarios with distances greater than 50km. However, this method has a greater computational cost in high-resolution grid scenarios and is sensitive to boundary conditions, so it only supports the formulation of large-scale emergency strategies.
[0007] 4) Downscaling techniques for mesoscale meteorological models: First, data preprocessing and downscaling calculations are performed. Then, the meteorological field is processed and a diffusion model is fitted. The optimization process is verified, and a refined meteorological field is generated. The specific process is as follows: Figure 4 As shown. This method integrates large-scale forecasts with local features and is suitable for nuclear emergency modeling at regional scales and in complex terrains. Its drawbacks are that it is computationally time-consuming and relies on high-quality initial data. However, it has been validated by observational data and can support accurate dose assessment in complex terrains.
[0008] Existing methods can meet the needs of some post-event review simulations, but some application scenarios, such as nuclear emergency response, require control over future field information and capture of global change patterns, while also having high performance to support emergency decision-making. These methods are all insufficient for dealing with such tasks, while the technical solution proposed in this application can specifically solve or improve these shortcomings.
[0009] 1) Limitations in Forward-Looking Prediction Capabilities: Existing nuclear emergency meteorological and dose field modeling technologies are mostly driven by real-time monitoring data, focusing on the post-event reproduction and reconstruction of existing field states. For example, based on observational data such as air pressure, humidity, and wind fields during past events, physical calculations are used to reconstruct the distribution of meteorological and nuclide diffusion dose fields for historical periods. These methods are essentially fitting and restoring historical states, lacking the ability to predict future meteorological field evolution and nuclide diffusion trends. They cannot support advance deployment and pre-emptive arrangements of protective measures in nuclear emergency scenarios, and are insufficient to meet the forward-looking decision-making needs during the dynamic development of nuclear accidents.
[0010] 2) Lack of global collaborative capability: Existing nuclear emergency meteorological field and dose field modeling technologies mostly adopt localized computational logic with modular and scale-based approaches. For example, they independently solve meteorological parameters for specific regions or time periods, or only focus on the single physical process of nuclide diffusion. It is difficult to explore the intrinsic correlation between meteorological field evolution, nuclide transport and dose field changes from a global spatiotemporal dimension. It is impossible to achieve global collaborative modeling at multiple scales, resulting in insufficient characterization of field change mechanisms in complex scenarios and difficulty in supporting systematic and global nuclear emergency risk assessment.
[0011] 3) Low efficiency in solving problems: Existing nuclear emergency meteorological field and dose field modeling technologies require high-resolution numerical simulations and refined calculations of multiple physical processes when facing complex scenarios such as complex terrain and unsteady weather. For example, mesoscale meteorological downscaling simulation and large-scale Eulerian grid global solution are required. These calculations have extremely high computational resource requirements and long calculation cycles, making it difficult to quickly output accurate results during the golden emergency period after a nuclear accident. This fails to meet the timeliness requirements for rapid decision-making in nuclear emergency scenarios and affects the timeliness and effectiveness of emergency response. Summary of the Invention
[0012] The present invention proposes a method, device and storage medium for predicting nuclear diffusion meteorological fields by fusing regional self-attention and modified ConvLSTM, which can at least solve one of the technical problems in the background art.
[0013] To achieve the above objectives, the present invention adopts the following technical solution: A method for predicting nuclear diffusion weather fields by fusing regional self-attention and modified ConvLSTM includes the following steps: S1. Collect multivariate meteorological observation data of the target area and construct an initial meteorological field dataset with three-dimensional spatiotemporal attributes of time, longitude, and latitude; S2. The initial meteorological field dataset is preprocessed by outlier removal, missing value interpolation and filtering for noise reduction. After time-series alignment, a multi-channel spatiotemporal data matrix is formed. The input samples and the corresponding future target meteorological field output samples are obtained by using a sliding window to construct a training sample set. S3. Input the input sample into the region self-attention module. First, divide the feature map into non-overlapping regions and extract the region features through average pooling. Perform self-attention operation on the region features to obtain the region attention features. Then, restore the original size spatial weight matrix through unpooling. Element-wise weighted with the input sample, obtain the spatial feature enhancement data. S4. Input the spatial feature enhancement data into the correction Conv-LSTM module, calculate the difference frame between the predicted frame of the previous time step and the current actual frame, add the difference frame as the compensation input to the input gate of Conv-LSTM, complete the cell state and hidden state update, and then perform directional correction on the output based on the ratio of the difference frame to the pre-output to obtain the temporally optimized hidden features. S5. Input the temporally optimized hidden features into a 1×1 convolutional layer, and after channel fusion and dimension mapping, output the nuclear diffusion meteorological field prediction results for future time moments.
[0014] In another aspect, the present invention also discloses a computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the steps of the method described above.
[0015] In another aspect, the present invention also discloses a computer device, including a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor performs the steps of the method described above.
[0016] As can be seen from the above technical solution, the nuclear diffusion meteorological field prediction method of the present invention, which integrates regional self-attention and modified ConvLSTM, relates to the field of nuclear emergency meteorological field prediction technology. Addressing the problems of insufficient forward-looking prediction capability, weak global correlation, and low computational efficiency in existing technologies, this invention employs a data-driven deep learning architecture to achieve rapid and accurate meteorological field prediction. The method includes: collecting multi-source meteorological variables to construct a spatiotemporal dataset; generating time-series samples through data cleaning, interpolation, filtering, and sliding window segmentation; improving the response of key regional features by pooling, attention calculation, and unpooling of spatial feature regions through regional self-attention; introducing a prediction difference feedback mechanism using modified Conv-LSTM to compensate for difference frames at the input gate and directionally correct the output, strengthening the learning of temporal evolution patterns; and finally outputting the future meteorological field prediction result through 1×1 convolution. This invention combines global spatial correlation and temporal prediction capabilities, has high computational efficiency and good prediction accuracy, and can meet the needs of rapid and forward-looking meteorological field extrapolation in nuclear accident emergency scenarios, making it suitable for nuclide diffusion risk assessment and emergency decision support.
[0017] Compared with existing technologies, the nuclear diffusion meteorological field prediction model that integrates regional self-attention and modified Conv-LSTM proposed in this invention has the following significant advantages and effects: 1) Adapts to the needs of forecasting tasks and has a forward-looking perspective. The method of this invention constructs a structure based on a length of The sliding window time-series sample construction mechanism shifts the model training objective from the traditional weather forecasting approach of replicating the current weather field to the trend projection of the future weather field. In the sample construction phase, the input samples are... The model uses multivariate climate data from consecutive time steps, with the output sample being the target meteorological field at the next time step after the input window. This time-shifted sample partitioning method forces the model to learn the evolution patterns of multivariates over time during training, rather than relying solely on static features at the current moment for prediction. Simultaneously, the modified difference feedback mechanism in Conv-LSTM enhances the ability to capture future trends. By calculating the difference between the predicted frame from the previous time step and the current actual frame, the difference frame is used as compensation input to the input gate. The pre-output is then targeted and corrected based on the ratio of the difference frame to the actual frame, enabling the model to dynamically adjust the direction of its predictions for future changes. This allows for more accurate predictions of the future evolution trend of the meteorological field, resulting in stronger forward-looking application value.
[0018] 2) Grasp the patterns of change from a holistic perspective; there is a stronger correlation. This invention achieves global correlation modeling across spatial and temporal dimensions through the synergistic effect of a region self-attention mechanism and a modified Conv-LSTM. Spatially, the region self-attention mechanism first divides the input data into multiple spatial regions, extracts representative features from each region using average pooling, calculates attention weights between regions, and finally maps these weights back to the original spatial dimensions using unpooling, achieving a weighted fusion of global spatial features. This mechanism differs from the local receptive field of traditional convolutional operations, capturing the correlation between distant spatial regions and avoiding the loss of spatial correlation caused by local feature extraction. Temporally, Conv-LSTM replaces fully connected operations with convolutional operations, preserving the structural information of the data in the spatial dimension. Simultaneously, through the transmission of cell states and hidden states, it learns the dynamic evolution of multiple variables over time. The modification module integrates temporal difference feedback into the Conv-LSTM's gating mechanism, enabling the model to dynamically adjust based on prediction errors when learning temporal patterns, thus strengthening temporal correlation. This joint modeling of spatial and temporal dimensions enables the prediction results to take into account both the physical correlation between variables and the spatiotemporal evolution patterns, avoiding the loss of correlation caused by dimensional splitting in traditional methods. The final output meteorological field is smoother in spatial distribution and more reasonable in temporal evolution.
[0019] 3) More efficient performance This invention significantly improves prediction efficiency by replacing the physical formula calculations in traditional weather forecasting with an end-to-end neural network architecture. Traditional nuclear diffusion weather field prediction typically relies on numerical weather prediction models, requiring the solution of complex fluid dynamics and thermodynamic equations. This computation involves numerous iterative calculations and grid partitioning, resulting in high computational costs and slow response times, making it difficult to meet real-time prediction requirements. In contrast, this invention directly drives the neural network for prediction using preprocessed multivariate data. A regional self-attention mechanism achieves spatial feature weighting, and a modified Conv-LSTM enables temporal pattern learning. The entire prediction process involves only basic mathematical operations such as matrix multiplication and convolution, eliminating the need for solving complex physical equations. Furthermore, the model has already completed the mining of multivariate dependencies during the training phase, and the inference phase only requires real-time collected multivariate data to quickly output prediction results. This data-driven prediction approach significantly reduces computational resource consumption while maintaining prediction accuracy, making it suitable for nuclear emergency scenarios requiring rapid response and possessing greater engineering application value. Attached Figure Description
[0020] Figure 1 This is a schematic diagram of the existing Gaussian plume model process; Figure 2 This is a schematic diagram of the existing Lagrange particle diffusion model process; Figure 3This is a schematic diagram of the existing Eulerian mesh model process; Figure 4 This is a schematic diagram of the downscaling technology process for existing mesoscale meteorological models; Figure 5 This is a schematic diagram of the prediction method according to an embodiment of the present invention; Figure 6 This is a schematic diagram of the region self-attention of the present invention; Figure 7 This is a schematic diagram of the modified Conv-LSTM of the present invention; Figure 8 This is a heatmap of U-Wind data from an embodiment of the present invention; Figure 9 This is a V-Wind data heatmap of an embodiment of the present invention. Detailed Implementation
[0021] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are some embodiments of the present invention, but not all embodiments.
[0022] The purpose of this invention is to solve the problem of predicting the meteorological field of nuclide diffusion in nuclear emergency scenarios. The specific process is as follows: Figure 5 As shown, the collected observation data is assimilated and cleaned, and the spatial features of the field data are weighted using regional self-attention. Then, the weighted data is captured by modified Conv-LSTM to capture the patterns of temporal features. The model learns from historical data cases and finally predicts the nuclide diffusion meteorological field based on the actual observation data input and output.
[0023] The specific steps of this plan are as follows: Step 1: Through a multi-type monitoring point sensor network deployed in the target area, collect and integrate multi-source climate variable data in real time, including 2-meter temperature (…). ), 2-meter dew point temperature ( ), surface air pressure ( ) and total cloud cover ( Each climate variable data is correlated with three-dimensional information of time, longitude, and latitude, forming a multivariate meteorological dataset with spatiotemporal attributes. Specifically, 2-meter temperature reflects the near-surface atmospheric thermal state, 2-meter dew point temperature characterizes the near-surface atmospheric water vapor content, surface air pressure reflects the vertical distribution characteristics of atmospheric mass, and total cloud cover reflects the degree of cloud cover in the region. Based on this multivariate data, an initial meteorological field dataset is constructed to provide fundamental data support for subsequent meteorological field predictions. The goal is to uncover the spatiotemporal correlations among the multiple variables and achieve high-precision predictions of the future meteorological field for a specified region.
[0024] Step 2: Preprocess the raw multivariate climate data obtained in Step 1. First, remove outlier data points that exceed the physically reasonable range. Then, use bilinear interpolation to fill in missing data. Finally, use wavelet transform filtering to eliminate random interference during data acquisition. After preprocessing, align and combine the multivariate data along the time dimension to ensure that the data at the same time step contains... The gridded data of all variables are used to form a multi-channel spatiotemporal data matrix. Further, a length of... The sliding window divides the input samples along the time dimension, with each input sample consisting of... The output sample consists of multivariate data from consecutive time steps, and the corresponding output sample is the target meteorological field data of the next time step after the current input sample time window. The meteorological field is constructed by calculating historical data through the CALPUFF model. A large-scale training sample set is generated based on historical meteorological data according to the above sample construction rules, while real-time collected multivariate data is retained as input for model inference to drive the trained model to output the future meteorological field prediction results.
[0025] Step 3: Use region self-attention to spatially weight the input data.
[0026] like Figure 6 As shown, region self-attention is based on the self-attention mechanism to weight the spatial features of multivariate data, enabling the model to more accurately identify important information in space. Unlike self-attention, region attention first divides the data into average regions, with each region represented by its average value, which is equivalent to average pooling. The pooling result is then used for attention operations to obtain a weight matrix. Finally, the weight matrix is depooled to the size of the input data and weighted element-wise with the input data.
[0027] By employing a regional self-attention mechanism to weight the spatial features of the input multivariate meteorological data, the model's ability to perceive key spatial regions is enhanced. Let the input feature tensor output in step 2 be... Its dimensions are ,in For the number of channels, and These represent the original height and width of the input data, respectively. The input features are then used with a partitioning function. Dimensionality reduction is performed. This function divides the original feature map into [variables] using average pooling. Divide the regions into non-overlapping regions and extract representative features from each region. (The resulting region features are shown.) Represented as:
[0028] in, For region partitioning function, and These represent the number of rows and columns in the divided region, respectively. After this operation, the data size is reduced from... Compress to The data for each region is the average of the data within that region after partitioning. The dimension is This step significantly reduces the matrix size for subsequent attention calculations, performing self-attention operations on the compressed region data to capture long-range dependencies between regions. This is achieved through a learnable weight matrix. A linear transformation is performed on the region features to generate a query, key, and value matrix, and the attention output is calculated. :
[0029] in, This is the scaling factor, typically taken as the square root of the key vector dimension. The function is used to normalize the attention weights, enabling the model to focus on regional features that are more important to the current prediction task. The low-dimensional attention feature map is then processed using a reduction function. Map back to the original spatial resolution. This function uses unpooling to... Feature map expansion to The size of the input data is used to generate a spatial weight matrix with the same size as the input data. :
[0030] in, and The original input data's row and column numbers are used as the restoration function. The parameters guide the upsampling process. The restored weight matrix... It contains global context information, reflecting the relative importance of each pixel in space. The generated spatial weight matrix... Compared with the original input data Perform the Hadamard product to obtain the spatially feature-weighted output. Weighted output features Represented as:
[0031] in, The Hadamard product represents element-wise multiplication. By applying regional feature weights to the input data, the model can adaptively enhance the feature responses of key regions, providing spatially enhanced input for subsequent time-series prediction modules.
[0032] Step 4: Use modified Conv-LSTM to learn the temporal patterns of the weighted data.
[0033] like Figure 7 As shown, the modified Conv-LSTM is based on Conv-LSTM to capture the temporal variation patterns of data, enabling the model to learn the correlation between climate variables and meteorological fields. A modification module is introduced on top of Conv-LSTM. This module provides direct difference feedback to the Conv-LSTM structure based on the difference between the predicted frame from the previous time step and the current actual frame. The difference frame is used as compensation input to the input gate of the Conv-LSTM cell, and the ratio of the difference frame to the actual frame is used as directional repair for the current pre-output, resulting in the cell state and hidden feature representation of the final predicted meteorological field. The modified Conv-LSTM module is used to learn the temporal patterns of the weighted feature data output from step 3 to capture the dynamic evolution of the meteorological field over time. This module introduces a bidirectional modification mechanism on top of the traditional Conv-LSTM, using the difference information between the predicted frame from the previous time step and the current actual frame to perform guided optimization of the model's internal state update and final output, thereby improving the accuracy of temporal extrapolation. Let the current time step be... The weighted input data output in step 3 is The hidden state at the previous moment was The current prediction is The actual observed true value at the current moment is The model calculates and predicts the difference frames. The difference frame represents the deviation between the prediction result at the previous time step and the current actual situation, as shown in the following formula:
[0034] The difference frame It serves as the guiding signal for the correction module, acting as a learnable compensation term during memory updates in the input gating system, enabling the model to converge faster and introducing learnable weight parameters. The model can dynamically adjust the impact of differential information on the activation level of the input gate, thereby suppressing or enhancing feature inflow into specific regions through differential feedback during the information filtering stage. (Input gate) Forgotten Gate Candidate memory units and the Gate of Oblivion The calculation formula is as follows:
[0035] in, for Activation function and These are the corresponding weight tensors and bias terms, respectively. The model updates the cell state. :
[0036] During the output correction phase, the model first calculates the uncorrected hidden states. Based on this, a preliminary prediction output is obtained, and then the difference frame is used. A correction coefficient is constructed by comparing the ratio of the initial predicted output to the corrected output, using learnable weight parameters. The final output at the current moment Scale correction is performed. This mechanism allows the model to adaptively scale the current output based on the error magnitude and direction from the previous time step. Final output. The calculation logic is as follows:
[0037] in, To smooth out terms and prevent the denominator from being zero, the modified Conv-LSTM introduces error feedback at the input to optimize state updates and uses ratio correction at the output to achieve targeted repair of the prediction results, thereby obtaining the cell states and hidden feature representations of the final predicted meteorological field.
[0038] Step 5: Input the modified hidden feature representation from the Conv-LSTM output in Step 4 into a 1×1 convolutional layer. The 1×1 convolution operation compresses the channel dimension of the high-dimensional hidden features and fuses them, mapping the hidden features to the output dimension of the target meteorological field to generate the final meteorological field prediction result. The 1×1 convolutional layer learns the weight relationships between different hidden feature channels, effectively integrating multi-scale spatiotemporal features. This ensures that the output meteorological field data matches the target prediction requirements in terms of spatial resolution and variable dimension, providing high-precision, fine-grained prediction data support for applications such as weather warnings and climate research.
[0039] Meanwhile, there are alternative solutions for some structures, devices, and method steps in this technical solution, which will be described in detail below.
[0040] 1) Alternatives to regional division methods In step 3, after dividing the region into non-overlapping average regions, the average value of the data within each region is used as the regional value. An alternative approach is to irregularly divide the region into custom regions based on actual climate conditions, and the representative value for each region can use representative data such as the mode, maximum, or minimum value.
[0041] The purpose of dividing regions is to reduce the data size while preserving spatial characteristics, thereby improving the efficiency of self-attention computation. Using average division and taking the average value is a common approach. In practice, targeted region division can be carried out according to specific circumstances.
[0042] 2) Replacement of historical meteorological field construction methods In step 2, the current method for constructing a centralized meteorological field is through computational simulation using CALPUFF, inputting historical multivariate climate data into CALPUFF to obtain the meteorological field. An alternative approach is to use physical models such as WRF or HYSPLIT, selecting an appropriate scale physical model based on the characteristics of the actual scenario to achieve more accurate fitting results.
[0043] Building a meteorological field in the dataset provides a learning target for the model. In the actual prediction process of the model, a meteorological field of similar quality can be generated without cumbersome physical calculation steps. Data construction belongs to the offline preprocessing process, and its computational cost does not affect the real-time performance of the model inference stage. The physical consistency and spatial integrity of the data can be prioritized over computational efficiency.
[0044] In summary, the core innovative technologies of the embodiments of the present invention are as follows: 1) A region self-attention weighting system was designed based on spatial attributes. This invention proposes a regional self-attention module that overcomes the computational bottleneck of traditional self-attention through regional partitioning and feature compression mechanisms, while accurately capturing the spatial dependencies of meteorological fields. This module dynamically divides the input feature map into non-overlapping regions. Through average pooling, it compresses the multivariate features of each region into single-valued vectors, significantly reducing the computational complexity of attention. Self-attention operations are performed based on the compressed regional features. A spatial weight matrix is generated by quantifying the importance of different regions to the prediction target. Then, an unpooling operation maps the weight matrix back to the original spatial resolution, achieving pixel-by-pixel fine-grained weighting. This design retains the ability to model global spatial correlations while significantly improving computational efficiency through regional compression strategies. It effectively solves the problem of balancing spatial feature redundancy and long-range dependencies in meteorological forecasting, enabling the model to accurately identify and strengthen the feature contributions of key regions, laying a high-fidelity spatial foundation for subsequent time-series forecasts.
[0045] 2) A modified Conv-LSTM was designed to address timing characteristics. This invention proposes a modified Conv-LSTM module, which optimizes the performance of traditional Conv-LSTM in temporal extrapolation by introducing a difference feedback mechanism based on prediction errors. This module adds a bidirectional correction path to the core Conv-LSTM structure. At the structure entrance, the difference frame between the previous predicted frame and the current actual frame is directly compensated to the input gate. The gating mechanism dynamically adjusts the information filtering weights at the current moment, enabling the model to actively focus on regions with large historical prediction deviations. At the structure exit, the ratio of the difference frame to the actual frame is used to construct directional correction coefficients, which are then used to perform directional weighted correction on the pre-output results, transforming the error signal into a learnable correction gradient. This design explicitly models the spatiotemporal distribution characteristics of prediction errors, guiding the model to learn from the evolution patterns, thus enabling the model to have stronger adaptive correction capabilities and long-term prediction stability during the nonlinear evolution of meteorological fields.
[0046] The following are examples: Each point in the meteorological field has wind direction and speed, which are directional vectors. Dividing these vectors into UV components transforms them into two scalars. Learning and predicting the results for each scalar is as follows: Figure 8-9 As shown, the left subplot of each graph represents the actual field data calculated from physics, the middle subplot represents the prediction result from model inference, and the right subplot represents the ratio of the difference between the actual and predicted values to the actual values. The colors of the heatmap for the actual and predicted values, from blue to red, represent the actual data values increasing. In the difference ratio heatmap, blue indicates that the predicted result for that point is too small, red indicates that the predicted result for that point is too large, and the lighter the color, the closer it is to white, the smaller the error in the prediction result for that point, and the more accurate the prediction.
[0047] In another aspect, the present invention also discloses a computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the steps of the method described above.
[0048] In another aspect, the present invention also discloses a computer device, including a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor performs the steps of the method described above.
[0049] It is understood that the systems, devices, and storage media provided in the embodiments of the present invention correspond to the methods provided in the embodiments of the present invention, and the explanations, examples, and beneficial effects of the relevant content can be referred to the corresponding parts of the above methods.
[0050] In the above embodiments, implementation can be achieved entirely or partially through software, hardware, firmware, or any combination thereof. When implemented using software, it can be implemented entirely or partially as a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., solid state disk (SSD)).
[0051] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.
[0052] The various embodiments in this specification are described in a related manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the system embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions of the method embodiments.
[0053] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.
Claims
1. A method for predicting nuclear diffusion meteorological fields by fusing regional self-attention and modified ConvLSTM, characterized in that... Modeling and analyzing meteorological field forecasting tasks from spatial and temporal dimensions includes the following steps. S1. Collect multivariate meteorological observation data of the target area and construct an initial meteorological field dataset with three-dimensional spatiotemporal attributes of time, longitude, and latitude; S2. The initial meteorological field dataset is preprocessed by outlier removal, missing value interpolation and filtering for noise reduction. After time-series alignment, a multi-channel spatiotemporal data matrix is formed. The input samples and the corresponding future target meteorological field output samples are obtained by using a sliding window to construct a training sample set. S3. Input the input sample into the region self-attention module. First, divide the feature map into non-overlapping regions and extract the region features through average pooling. Perform self-attention operation on the region features to obtain the region attention features. Then, restore the original size spatial weight matrix through unpooling. Element-wise weighted with the input sample, obtain the spatial feature enhancement data. S4. Input the spatial feature enhancement data into the correction Conv-LSTM module, calculate the difference frame between the predicted frame of the previous time step and the current actual frame, add the difference frame as the compensation input to the input gate of Conv-LSTM, complete the cell state and hidden state update, and then perform directional correction on the output based on the ratio of the difference frame to the pre-output to obtain the temporally optimized hidden features. S5. Input the temporally optimized hidden features into a 1×1 convolutional layer, and after channel fusion and dimension mapping, output the nuclear diffusion meteorological field prediction results for future time moments.
2. The nuclear diffusion meteorological field prediction method based on fusion of regional self-attention and modified ConvLSTM as described in claim 1, characterized in that: S1 specifically includes, By deploying a multi-type monitoring point sensor network in the target area, multi-source climate variable data are collected and integrated in real time. The climate variables include 2-meter temperature, 2-meter dew point temperature, surface air pressure and total cloud cover. Each climate variable data is associated with three-dimensional information of time, longitude and latitude, forming a multivariate meteorological dataset with spatiotemporal attributes; among them, 2-meter temperature reflects the near-surface atmospheric thermal state, 2-meter dew point temperature characterizes the near-surface atmospheric water vapor content, surface air pressure reflects the vertical distribution characteristics of atmospheric mass, and total cloud cover reflects the degree of cloud cover in the region.
3. The nuclear diffusion meteorological field prediction method based on fusion of regional self-attention and modified ConvLSTM as described in claim 2, characterized in that: S2 specifically includes, The raw multivariate climate data acquired by S1 is preprocessed by first removing outlier data points that exceed the set range, then filling in the missing data using bilinear interpolation, and finally eliminating random interference during the data acquisition process through wavelet transform filtering. After preprocessing, the multivariate data are aligned and combined in the time dimension to form a multi-channel spatiotemporal data matrix containing all variables including 2-meter temperature, 2-meter dew point temperature, surface air pressure and total cloud cover at the same time step. Use a length of The sliding window divides the input samples along the time dimension, with each input sample consisting of... The output sample consists of multivariate data from consecutive time steps, and the corresponding output sample is the target meteorological field data of the next time step after the current input sample time window. The meteorological field is constructed by calculating historical data through the CALPUFF model. A large-scale training sample set is generated based on historical meteorological data according to the above sample construction rules, while real-time collected multivariate data is retained as input for model inference to drive the trained model to output the future meteorological field prediction results.
4. The nuclear diffusion meteorological field prediction method based on fusion of regional self-attention and modified ConvLSTM as described in claim 3, characterized in that: S3 specifically includes, By using a regional self-attention mechanism to weight the spatial features of the input multivariate meteorological data, the model's ability to perceive key spatial regions is enhanced. Let the input feature tensor of the output of S2 be... Its dimensions are ,in For the number of channels, and These represent the original height and width of the input data, respectively; the input features are then processed using a partitioning function. Dimensionality reduction is performed; this partitioning function divides the original feature map into segments using average pooling. Extract representative features from each of three non-overlapping regions. Regional characteristics after division Represented as: ; in, and These represent the number of rows and columns in the divided region, respectively; the data size is determined by... Compress to The data for each region is the average of the data within that region after partitioning. The dimension is ; Through a learnable weight matrix A linear transformation is performed on the region features to generate a query, key, and value matrix, and the attention output is calculated. : ; in, Scaling factor The function is used to normalize the attention weights, enabling the model to focus on regional features that are more important to the current prediction task; the reduction function is used to transform the low-dimensional attention feature map. Mapping back to the original spatial resolution; this restoration function employs unpooling to... Feature map expansion to The size of the input data is used to generate a spatial weight matrix with the same size as the input data. : ; in, and The number of rows and columns of the original input data serves as the parameter of the restoration function to guide the upsampling process; the restored weight matrix... It contains global context information, reflecting the relative importance of each pixel in space; the generated spatial weight matrix Compared with the original input data Perform the Hadamard product to obtain the spatially feature-weighted output. Weighted output features Represented as: ; in, The Hadamard product represents element-wise multiplication.
5. The nuclear diffusion meteorological field prediction method based on fusion of regional self-attention and modified ConvLSTM according to claim 4, characterized in that: S4 includes, The modified Conv-LSTM module is used to learn the temporal patterns of the weighted feature data output by S3 in order to capture the dynamic evolution of the meteorological field over time. Let the current time be The weighted input data output by S3 is The hidden state at the previous moment was The current prediction is The actual observed true value at the current moment is Model calculation predicts difference frames The difference frame represents the deviation between the prediction result at the previous time step and the current actual situation, as shown in the following formula: ; The difference frame It serves as the guiding signal for the correction module, acting as a learnable compensation term during memory updates in the input gating system, enabling the model to converge faster and introducing learnable weight parameters. The model can dynamically adjust the impact of differential information on the activation level of the input gate, thereby suppressing or enhancing feature inflow into specific regions through differential feedback during the information filtering stage; input Door Forgotten Gate Candidate memory units and the Gate of Oblivion The calculation formula is as follows: ; in, for Activation function and These are the corresponding weight tensors and bias terms, respectively; Model updates cell state : ; During the output correction phase, the model first calculates the uncorrected hidden states. Based on this, a preliminary prediction output is obtained, and then the difference frame is used. A correction coefficient is constructed by comparing the ratio of the initial predicted output to the corrected output, using learnable weight parameters. The final output at the current moment Perform proportional correction; final output The calculation logic is as follows: ; in, To smooth out terms and prevent the denominator from being zero, the modified Conv-LSTM introduces error feedback at the input to optimize state updates and implements targeted repair of the prediction results at the output through ratio correction, thereby obtaining the cell state and hidden feature representation of the final predicted meteorological field.
6. The nuclear diffusion meteorological field prediction method based on fused region self-attention and modified ConvLSTM according to claim 5, characterized in that: Step S5 includes, The hidden feature representation of the modified Conv-LSTM output in S4 is input into a 1×1 convolutional layer. The high-dimensional hidden features are compressed in channel dimension and fused through 1×1 convolution operation. The hidden features are then mapped to the output dimension of the target meteorological field to generate the final meteorological field prediction result.
7. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, the processor performs the steps of the method as described in any one of claims 1 to 6.
8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the computer program is executed by the processor, the processor performs the steps of the method as described in any one of claims 1 to 6.
