A method, system, equipment, and medium for precipitation prediction and probabilistic forecasting based on physical consistency constraints.

Abstract

This invention relates to the field of precipitation forecasting technology, and discloses a precipitation forecasting and probabilistic prediction method, system, equipment, and medium based on physical consistency constraints. The method includes: in the training phase, inputting preprocessed first meteorological data into the model to obtain preliminary prediction results, and constructing physical consistency constraints by combining the physical fields in the results; constructing classification prediction losses for precipitation magnitude intervals; performing joint optimization by combining various losses to obtain a trained precipitation forecasting model; inputting preprocessed second meteorological data into the trained model, and outputting continuous prediction results of future precipitation intensity through a physical enhancement feature construction module, a multi-scale circulation feature fusion module, a spatiotemporal feature fusion module, and a feature reconstruction and prediction module. This invention can improve the stability and generalization ability of the model, and provide direct and usable magnitude event probability information for precipitation forecast results, thereby meeting the needs for precipitation risk classification and early warning and extreme event probability assessment.

Description

Technical Field

[0001] This invention relates to the field of precipitation prediction technology, and in particular to a precipitation prediction and probability forecasting method, system, equipment and medium based on physical consistency constraints. Background Technology

[0002] In recent years, artificial intelligence (AI) technologies, especially deep learning algorithms, have developed rapidly in the field of weather forecasting. Existing technologies primarily rely on data-driven AI weather forecasting models, which typically employ deep learning models to learn statistical mapping relationships from historical meteorological data to predict future meteorological elements. These methods usually use reanalysis data or numerical weather prediction model data as input, utilizing structures such as convolutional neural networks, spatiotemporal recurrent networks, Transformers, or graph neural networks to directly learn the mapping relationship from historical meteorological fields to future meteorological fields, outputting a single deterministic prediction of future precipitation intensity or related meteorological variables. Training is typically performed by minimizing loss functions such as mean squared error (MSE) and mean absolute error (MAE), emphasizing prediction accuracy under the overall mean. This approach has indeed achieved certain results in short- to medium-term weather forecasting and also has advantages in computational efficiency compared to traditional numerical models.

[0003] However, there are still some shortcomings in the prediction of precipitation, especially heavy precipitation and extreme precipitation events. These shortcomings are mainly reflected in the following aspects: (1) Existing data-driven AI weather forecast models mainly rely on historical samples for statistical learning. The training objectives of the models usually only focus on minimizing numerical errors, without introducing clear atmospheric physical constraints at the model structure or loss function level. This approach can easily lead to deviations between the predicted results and the actual atmospheric processes in precipitation forecasting. For example, there are problems such as inconsistencies between water vapor transport and precipitation intensity, and a lack of reasonable dynamic support for local precipitation enhancement. This affects the stability and reliability of the model under different climate backgrounds or extreme weather conditions. (2) Most models adopt a deterministic prediction framework. Their training objectives are guided by minimizing the overall mean error. In precipitation problems with uneven sample distribution, they tend to optimize small and medium precipitation events. Their fitting ability for heavy precipitation and extreme precipitation samples is insufficient. Therefore, it is easy to lead to low predicted amplitudes for extreme precipitation events, making it difficult to provide reliable probabilistic decision-making basis for disaster prevention, mitigation and risk assessment. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a precipitation prediction and probabilistic forecasting method and system based on physical consistency constraints to solve the problems that current data-driven models generally lack physical constraints, have insufficient physical consistency in prediction results, and that deterministic prediction frameworks tend to underestimate extreme precipitation events and have insufficient detection rates of extreme events.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: In a first aspect, the present invention provides a precipitation prediction and probabilistic forecasting method based on physical consistency constraints, comprising: Precipitation prediction model training phase and forecast inference phase; The precipitation prediction model training phase includes: Acquire multivariate meteorological data for the target area within a preset time window; The preprocessed first meteorological data is input into the precipitation prediction model to obtain preliminary prediction results; the precipitation prediction model includes at least a physical enhancement feature construction module, a multi-scale circulation feature fusion module, a spatiotemporal feature fusion module, and a feature reconstruction and prediction module connected in sequence; Based on the specific humidity and horizontal wind field components in the preliminary prediction results, the vertical integral of water vapor flux is calculated, and the difference between the corresponding calculation amount and the vertical integral of the actual water vapor flux in the training data is used as a physical consistency constraint to participate in the model training. Based on the preliminary forecast results, a classification-based predicted loss is constructed for precipitation level intervals; Based on the precipitation prediction loss, combined with the physical consistency constraint loss and / or classification prediction loss, the model parameters are jointly optimized to obtain the trained precipitation prediction model. The forecast inference stage includes: The preprocessed second meteorological data is input into the trained precipitation prediction model, and the physical quantities in the input data are structured and modeled to extract physical enhancement features. Multi-scale circulation feature fusion is performed on the physical enhancement features to extract and fuse atmospheric circulation structure features at different spatial scales. Joint modeling of spatial structure and temporal evolution information of multi-scale circulation characteristics forms spatiotemporal fusion characteristics; Based on the spatiotemporal fusion features, feature reconstruction is performed, and continuous prediction results of future precipitation intensity are output.

[0007] As a preferred embodiment of the precipitation prediction and probabilistic forecasting method based on physical consistency constraints described in this invention, the method includes: inputting preprocessed second meteorological data into a trained precipitation prediction model; performing structured modeling on the physical quantities in the input data; and extracting physical enhancement features, including: Based on pre-defined meteorological and physical priors, the input physical quantities are divided along the channel dimension into... Each physical group yields the recombined physical tensor; For the recombined physical tensor, a combination of channel transformation and spatial convolution is used to extract features to obtain physical enhancement features.

[0008] As a preferred embodiment of the precipitation prediction and probabilistic forecasting method based on physical consistency constraints described in this invention, the method involves: fusing multi-scale circulation features to the physical enhancement features, extracting and fusing atmospheric circulation structure features at different spatial scales, including: Based on feature transformation operators with different spatial perception ranges, the physical enhancement features are processed to extract multi-scale circulation features on a continuous spatial scale spectrum to characterize atmospheric circulation and precipitation-related processes. The spatial scale spectrum covers different spatial scales from planetary scales to local disturbances. Based on the global statistical information of the physical enhancement features, scale weights corresponding to each spatial scale are adaptively generated, and the circulation features at different spatial scales are weighted and fused based on the scale weights to form a unified multi-scale circulation feature representation.

[0009] As a preferred embodiment of the precipitation prediction and probabilistic forecasting method based on physical consistency constraints described in this invention, the method involves: jointly modeling the spatial structure and temporal evolution information of multi-scale circulation characteristics to form spatiotemporal fusion features, including: Spatial convolution operators are applied to multi-scale circulation features to obtain spatially enhanced features; Time-enhanced features are obtained by processing multi-scale circulation features using time feature modeling operators; By fusing spatial augmentation features with temporal augmentation features, preliminary spatiotemporal joint features are obtained; The preliminary spatiotemporal joint features are subjected to channel projection and residual enhancement processing to obtain spatiotemporal fusion features.

[0010] As a preferred embodiment of the precipitation prediction and probabilistic forecasting method based on physical consistency constraints described in this invention, the method includes: calculating the vertical integral of water vapor flux based on the specific humidity and horizontal wind field components in the preliminary prediction results, and using the difference between the calculated amount and the vertical integral of the actual water vapor flux in the training data as a physical consistency constraint in model training, including: Based on the model-predicted specific humidity and horizontal wind field components, the vertical integral water vapor transport vector is calculated. Based on the consistency relationship between the vertical integral water vapor transport vector and the actual water vapor transport volume, a physical consistency constraint loss term is constructed. , is represented as: in, This represents a distance or difference metric function used to measure the consistency between the two. Represents the vertical integral water vapor transport vector; This indicates the actual amount of water vapor transported.

[0011] As a preferred embodiment of the precipitation prediction and probabilistic forecasting method based on physical consistency constraints described in this invention, the method comprises: constructing a classification prediction loss for precipitation magnitude intervals based on preliminary prediction results, including: While outputting preliminary prediction results, the probability prediction of precipitation level ranges is carried out based on the preliminary spatiotemporal fusion features generated during the training phase. A classification loss function is introduced to conduct supervised learning on the precipitation level intervals of each grid point, thereby constructing the classification prediction loss.

[0012] As a preferred embodiment of the precipitation prediction and probability forecasting method based on physical consistency constraints described in this invention, it further includes setting several precipitation intensity thresholds in response to the model outputting preliminary prediction results or outputting continuous prediction results of future precipitation intensity. Future precipitation is divided into multiple mutually exclusive precipitation level intervals at each spatial grid point according to the precipitation intensity threshold. Based on the spatiotemporal fusion features shared by the model, the probability of each precipitation level interval is predicted and the occurrence probability of each precipitation level interval is output.

[0013] Secondly, the present invention provides a precipitation prediction and probabilistic forecasting system based on physical consistency constraints, comprising: The acquisition module is used to acquire multivariate meteorological data of the target area within a preset time window; The physical enhancement feature construction module is used to input the preprocessed second meteorological data into the trained precipitation prediction model, perform structured modeling of the physical quantities in the input data, and extract physical enhancement features. The multi-scale circulation feature fusion module is used to perform multi-scale circulation feature fusion on the physical enhancement features, and extract and fuse atmospheric circulation structure features at different spatial scales respectively. The spatiotemporal feature fusion module is used to jointly model the spatial structure and temporal evolution information of multi-scale circulation features to form spatiotemporal fused features. The feature reconstruction and prediction module is used to perform feature reconstruction based on the spatiotemporal fusion features and output continuous prediction results of future precipitation intensity.

[0014] Thirdly, the present invention provides a computer device, comprising: Memory and processor; The memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions. When the computer-executable instructions are executed by the processor, they implement the steps of a precipitation prediction and probability forecasting method based on physical consistency constraints.

[0015] Fourthly, the present invention provides a computer-readable storage medium storing computer-executable instructions that, when executed by a processor, implement the steps of the precipitation prediction and probability forecasting method based on physical consistency constraints.

[0016] Compared with existing technologies, the beneficial effects of this invention are as follows: This invention introduces physical constraints into the precipitation prediction model, which improves the physical consistency between precipitation prediction results and related atmospheric physical processes while maintaining high computational efficiency, thereby enhancing the model's stability and generalization ability. Overall, by organically combining multivariate meteorological physical elements, physical enhancement feature modeling, multi-scale circulation analysis, spatiotemporal feature fusion, and physical constraint training mechanisms, this invention overcomes the problem of systematic underestimation of extreme precipitation events in traditional deterministic prediction frameworks, providing direct and usable magnitude event probability information for precipitation forecast results, thus meeting the needs of actual operational applications for precipitation risk classification and early warning, and extreme event probability assessment. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a schematic diagram of the precipitation prediction model training phase in a precipitation prediction and probabilistic forecasting method based on physical consistency constraints, as described in one embodiment of the present invention.

[0019] Figure 2 This is a schematic diagram of the precipitation prediction model inference stage in a precipitation prediction and probabilistic forecasting method based on physical consistency constraints, as described in one embodiment of the present invention. Detailed Implementation

[0020] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0021] Example 1, referring to Figures 1-2 As an embodiment of the present invention, a precipitation prediction and probabilistic forecasting method based on physical consistency constraints is provided, including: a precipitation prediction model training stage and a forecast inference stage; The precipitation prediction model training phase includes: S100: Acquire multivariate meteorological data for the target area within a preset time window; S200: Input the preprocessed first meteorological data into the precipitation prediction model to obtain preliminary prediction results; the precipitation prediction model includes at least a physical enhancement feature construction module, a multi-scale circulation feature fusion module, a spatiotemporal feature fusion module, and a feature reconstruction and prediction module connected in sequence; S300: Based on the specific humidity and horizontal wind field components in the preliminary prediction results, calculate the vertical integral of water vapor flux, and use the difference between the corresponding calculation amount and the vertical integral of the actual water vapor flux in the training data as a physical consistency constraint to participate in model training. S400: Based on the preliminary forecast results, construct the classification prediction loss for precipitation level intervals; S500: Based on precipitation prediction loss, combined with physical consistency constraint loss and / or classification prediction loss, the model parameters are jointly optimized to obtain the trained precipitation prediction model. The prediction inference stage includes: A100: Input the preprocessed second meteorological data into the trained precipitation prediction model, perform structured modeling of the physical quantities in the input data, and extract physical enhancement features; A200: Perform multi-scale circulation feature fusion on the physical enhancement features, and extract and fuse atmospheric circulation structure features at different spatial scales respectively; A300: Jointly modeling the spatial structure and temporal evolution information of multi-scale circulation characteristics to form spatiotemporal fusion features; A400: Based on the spatiotemporal fusion features, feature reconstruction is performed, and the continuous prediction results of future precipitation intensity are output.

[0022] It should be noted that S100-S500 mainly represent the model training phase. The training phase includes data preparation, model building, forward computation, calculating the predicted IVT, obtaining the true IVT, constructing the physical constraint loss, and optimizing by combining other losses. The forecast inference phase includes data preprocessing, inputting the trained model, and outputting precipitation intensity and probability. Structurally, the training and inference phases are the same; the difference lies in the introduction of a vertical integral based on water vapor flux as a physical consistency constraint only during the model training phase. This constraint is used to constrain the consistency between the derived underlying physical field and the actual macroscopic physical quantities, thus ensuring that the precipitation prediction results maintain physical consistency with the corresponding atmospheric circulation and water vapor transport processes. This constraint is not involved in the calculations during the model inference phase.

[0023] Example 2, refer to Figures 1-2This is one embodiment of the present invention. Based on the above embodiment, a precipitation prediction and probabilistic forecasting method based on physical consistency constraints is provided.

[0024] like Figure 1 As shown, the precipitation prediction model training phase includes: S100: Acquire multivariate meteorological data for the target area within a preset time window; Specifically, multivariate meteorological data can cover the surface layer and multiple isobaric layers; the isobaric layers can include 13 pressure layers (1000hPa, 925hPa, 850hPa, 700hPa, 600hPa, 500hPa, 400hPa, 350hPa, 250hPa, 200hPa, 150hPa, 100hPa, 50hPa).

[0025] For example, meteorological data includes at least geopotential height. ,temperature Horizontal wind field components and Vertical wind speed , specific moisture The vertical integral of water vapor flux, and may also include 2m temperature and 10m wind field at the ground. , Variables include (components), dew point temperature, sea level pressure, sea surface temperature, soil moisture, historical precipitation, and topographic elevation.

[0026] In one optional implementation, the acquired data is preprocessed in S100 by performing unified temporal alignment and spatial interpolation on the multivariate meteorological elements covering the ground layer and multiple isobaric layers, so that different physical elements are aligned under a unified spatial grid and temporal resolution; and the physical elements are normalized and missing values ​​are processed to construct a continuous and consistent historical meteorological field sequence as model input, providing a stable data foundation for subsequent precipitation prediction and classification prediction training based on physical constraints.

[0027] Specifically, let the preprocessed multivariate meteorological field be represented as a tensor: in, Indicates the length of a historical time step; Indicates the height and width of the spatial grid; This indicates the number of channels for meteorological physical variables.

[0028] S200: Input the preprocessed first meteorological data into the precipitation prediction model to obtain preliminary prediction results; the precipitation prediction model includes at least a physical enhancement feature construction module, a multi-scale circulation feature fusion module, a spatiotemporal feature fusion module, and a feature reconstruction and prediction module connected in sequence; Based on the above implementation method, at each time step The meteorological field is composed of various physical elements. Let the current reporting time be... The target of prediction is the future. The meteorological element field within a time period is denoted as ; In time Using a multivariate meteorological element field as input, a mapping relationship is established from the historical meteorological field to the future meteorological field, denoted as […]. ;in, This represents the precipitation prediction model function; This represents the future weather field predicted by the model.

[0029] In another alternative implementation, based on the above implementation, when forecasts over longer time spans are required, the model can be... The prediction result at time step one is used as the input for the next time step, and is gradually generated through an autoregressive approach. The weather forecast results at any given time.

[0030] S300: Based on the specific humidity and horizontal wind field components in the preliminary prediction results, calculate the vertical integral of water vapor flux, and use the difference between the corresponding calculation amount and the vertical integral of the actual water vapor flux in the training data as a physical consistency constraint to participate in model training. In this embodiment of the application, S300 includes: Based on the model-predicted specific humidity and horizontal wind field components, the vertical integral water vapor transport vector is calculated. Specifically, precipitation prediction models simultaneously predict multiple layers of physical variables, including: specific humidity. and horizontal wind field components Define the vertical integral water vapor transport vector (IVT) obtained from the physical field calculation of the model output as follows: in: This refers to the ground pressure (or the lower boundary pressure of the integral). The integral is the upper boundary air pressure. is the gravitational acceleration constant.

[0031] Meanwhile, the training data directly provides the actual vertical integral water vapor transport at the corresponding time point, denoted as .

[0032] Based on the consistency relationship between the vertical integral water vapor transport vector and the actual water vapor transport volume, a physical consistency constraint loss term is constructed. , is represented as: in, This represents a distance or difference metric function used to measure the consistency between the two. Represents the vertical integral water vapor transport vector; This indicates the actual amount of water vapor transported.

[0033] In one alternative implementation, a distance or difference metric function is used to measure the consistency between the two. Mean squared error, weighted norm, or other continuous difference measures can be used.

[0034] It should be noted that, during the training phase, by introducing physical consistency constraints, the model can suppress non-physical results that do not match the water vapor transport conditions during precipitation prediction, thereby improving the physical reliability and stability of the prediction results.

[0035] S400: Based on the preliminary forecast results, construct the classification prediction loss for precipitation level intervals; In this embodiment of the application, S400 includes: While outputting preliminary prediction results, the probability prediction of precipitation level ranges is carried out based on the preliminary spatiotemporal fusion features generated during the training phase. A classification loss function is introduced to conduct supervised learning on the precipitation level intervals of each grid point, thereby constructing the classification prediction loss.

[0036] In one alternative implementation, to alleviate the problem of uneven distribution of precipitation samples, the classification loss function can adopt the weighted cross-entropy loss function, which can assign weights to each precipitation level interval that are inversely proportional to the sample frequency.

[0037] In another alternative implementation, the classification loss function can be the Focal Loss function, which uses a modulation factor to make the model training focus more on the difficult-to-classify heavy precipitation samples.

[0038] It should be noted that the classification loss function is not limited to the example above, and other loss function forms suitable for multi-class classification tasks can be used.

[0039] S500: Based on precipitation prediction loss, combined with physical consistency constraint loss and / or classification prediction loss, the model parameters are jointly optimized to obtain the trained precipitation prediction model. Specifically, during the model training phase, based on all the output results from the above steps, a combined loss function is constructed, expressed as: in: The precipitation prediction loss is specifically the mean square error loss between the model's predicted continuous physical variables (including precipitation intensity and other output physical fields) and their corresponding true values. The physical consistency constraint loss is based on vertical integral water vapor transport; Losses are categorized by precipitation level; Here are the weight coefficients for each loss term. The model parameters are jointly optimized by minimizing the above combined loss function.

[0040] In one alternative implementation, the loss function in S500 can be constructed by combining only the precipitation prediction loss and the physical consistency constraint loss; or by combining the precipitation prediction loss and the classification loss for optimization.

[0041] In another alternative implementation, S500 can also be optimized solely through the precipitation prediction loss function.

[0042] like Figure 2 As shown, the prediction inference stage includes: A100: Input the preprocessed second meteorological data into the trained precipitation prediction model, perform structured modeling of the physical quantities in the input data, and extract physical enhancement features; It should be noted that the acquisition and preprocessing methods of A100 are the same as those in step S100, and will not be repeated here.

[0043] Based on the training phase, it can be seen that, based on the preprocessed multivariate meteorological field Based on this, the prediction model function It further includes physical enhancement steps and spatiotemporal feature fusion steps.

[0044] Among them, the physical enhancement step is used to construct physical prior features for the input multivariate meteorological field, embedding physical quantities closely related to precipitation formation into the internal feature space of the model in a structured manner, as described below; In this embodiment of the application, in step A100, the preprocessed second meteorological data is input into the trained precipitation prediction model, and the physical quantities in the input data are structured and modeled to extract physical enhancement features, including the following steps A1-A2: A1: Based on pre-defined meteorological and physical priors, the input physical quantities are divided along the channel dimension into... Each physical group yields the recombined physical tensor; Specifically, A1 uses the initially acquired output As input, considering the functional differences of different physical quantities in the precipitation formation process, based on pre-set meteorological and physical priors, the physical variables are divided along the channel dimension. There are 3 physical groups; each physical group corresponds to a set of variables with similar physical meanings. By reshaping the channel dimension, the reshaped physical tensor is obtained, represented as: in, This indicates the number of physical groups, and its value can be set based on prior meteorological knowledge. Prior meteorological physical knowledge refers to the functional roles and inherent coupling relationships of different meteorological physical elements in atmospheric motion and precipitation formation.

[0045] Specifically, different physical elements can be attributed to different physical processes or controlling factors in the precipitation formation mechanism; for example, water vapor transport processes, dynamic processes, thermal processes, and underlying surface influence processes.

[0046] It should be noted that, based on the above-mentioned division of physical processes, grouping variables with similar physical meanings or strong correlations in physical mechanisms into the same physical group helps to enhance the model's ability to jointly model similar physical information and suppress interference between unrelated physical variables.

[0047] A2: For the recombined physical tensor, feature extraction is performed by combining channel transformation and spatial convolution to obtain physical enhancement features; Specifically, regarding the recombined physical tensor Feature extraction is performed by combining channel transformation and spatial convolution. This allows the model to model the interaction between different physical quantities and extract spatial features while maintaining the grouping structure of physical variables. The mapping process is represented as follows: in, and The point convolution kernel is used for channel-level compression and restoration. It uses depthwise separable convolution kernels to extract local spatial structure features. " indicates convolution operation, It is a non-linear activation function.

[0048] The final output of A2 can be represented as: ;in, The feature channel dimension is used; the physically enhanced features serve as inputs for subsequent spatiotemporal feature extraction and prediction steps.

[0049] It should be noted that the physical augmentation step introduces physical prior information at the feature level, while the physical consistency constraint step imposes consistency restrictions on the physical quantities output by the model at the prediction level. Both act on different levels of the model, jointly ensuring the physical rationality of the model's feature representation and prediction results.

[0050] A200: Perform multi-scale circulation feature fusion on the physical enhancement features, and extract and fuse atmospheric circulation structure features at different spatial scales respectively; In this embodiment of the application, A200 performs multi-scale circulation feature fusion on the physical enhancement features, extracting and fusing atmospheric circulation structure features at different spatial scales, including the following steps B1-B2: B1: Based on feature transformation operators with different spatial perception ranges, the physical enhancement features are processed to extract multi-scale circulation features that characterize atmospheric circulation and precipitation-related processes on a continuous spatial scale spectrum, which covers different spatial scales from planetary scales to local disturbances. Specifically, let the feature tensor output by A2 be: ; By applying spatial feature transformations of different scales to the input features, a multi-scale circulation feature representation is constructed. The circulation characteristics at each scale are defined as follows: in, Indicates the first Spatial feature transformation operators at various scales; different scales correspond to different spatial perception ranges, used to characterize the structure of circulation and weather systems at different scales.

[0051] It should be noted that, through the above process, the model can identify features associated with precipitation events at multiple spatial scales, such as planetary-scale circulation background, weather-scale systems, and local-scale convection triggering and disturbances.

[0052] Furthermore, at every spatial scale Above, intra-scale modeling is performed on the corresponding circulation features to obtain scale-enhanced features, represented as: in, This indicates a feature extraction operator within a given scale, used to characterize the spatial correlation and self-similarity structure of meteorological fields at that scale. This process is used to explicitly model the mechanism by which circulation structures at different spatial scales influence precipitation formation.

[0053] B2: Based on the global statistical information of the physical enhancement features, adaptively generate the scale weights corresponding to each spatial scale, and perform weighted fusion of circulation features at different spatial scales based on the scale weights to form a unified multi-scale circulation feature representation. Specifically, based on the above implementation methods, a unified multi-scale circulation characteristic is formed, which can be expressed as: in: This indicates a cross-scale feature fusion operation; Indicates the first Weight adjustment function for each scale.

[0054] In an optional implementation, to dynamically reflect the importance of circulation characteristics at different spatial scales in current precipitation forecasting, a scale weighting coefficient can be introduced: Among them, scale weight Adaptively generated based on global statistical information of scale features, represented as ; This represents the global feature aggregation operator. This represents the normalized mapping function.

[0055] In summary, the characteristics of multi-scale circulation can be expressed as follows: As input for the subsequent step of spatial feature fusion, it is used to further model the circulation evolution process and precipitation temporal variation characteristics.

[0056] It should be noted that step A200 extracts and fuses circulation features at different spatial scales, enabling the model to explicitly characterize the synergistic effects of multi-scale weather systems on precipitation formation, and can provide a unified representation of circulation features for subsequent spatiotemporal evolution modeling.

[0057] A300: Jointly modeling the spatial structure and temporal evolution information of multi-scale circulation characteristics to form spatiotemporal fusion features; In this embodiment of the application, A300 performs joint modeling of spatial structure and temporal evolution information of multi-scale circulation characteristics to form spatiotemporal fusion characteristics, including the following steps C1-C4: C1: Applying a spatial convolution operator to multi-scale circulation features yields spatially enhanced features; Specifically, spatial augmentation features can be represented as: ;in," " indicates spatial convolution operation; This represents a spatial feature extraction operator used to capture spatial structure information within a local and neighborhood range; this spatial path is used to model the spatial morphological features and organizational structure of precipitation systems.

[0058] C2: Time-enhanced features are obtained by processing multi-scale circulation features through time feature modeling operators; Specifically, time-enhanced features can be represented as: ;in, This represents a time-feature modeling operator used to capture the dependencies between different time steps; this time path is used to characterize the propagation, evolution, and persistence of precipitation systems in the time dimension.

[0059] In an alternative implementation, temporal feature modeling in C2 can be achieved through a self-attention mechanism, where the attention weights are defined as: in, , The projection vector of the time feature; Relative positional information related to time intervals is used to characterize the time decay or duration of precipitation processes.

[0060] C3: The spatial enhancement features are fused with the temporal enhancement features to obtain preliminary spatiotemporal joint features; Specifically, the preliminary spatiotemporal joint characteristics can be expressed as: in, This represents the normalization operator, used to stabilize the characteristic distribution.

[0061] C4: Perform channel projection and residual enhancement processing on the preliminary spatiotemporal joint features to obtain spatiotemporal fusion features; Specifically, to further enhance the nonlinear expressive power and maintain the stability of information transmission, the fused features are subjected to channel projection and residual enhancement processing, as follows: in, It represents the feature projection and nonlinear transformation operator; the residual connection is used to suppress gradient vanishing and preserve the original feature information.

[0062] Finally, the spatiotemporal fusion features output by C4 are represented as follows: The spatiotemporal joint features serve as inputs for subsequent steps, such as A400 feature reconstruction and precipitation prediction. In A400, multi-scale feature representations can be further constructed through scale transformation operations to generate high-resolution precipitation prediction results and related probability products.

[0063] A400: Based on the spatiotemporal fusion features, feature reconstruction is performed, and continuous prediction results of future precipitation intensity are output; Specifically, A400 applies a scale transformation operation to the spatiotemporal fusion features, reconstructing the multi-scale, low-resolution spatiotemporal features into precipitation prediction results at the target spatial resolution. Through scale-aware feature fusion, spatial resolution restoration, and residual enhancement mechanisms, it achieves a fine characterization of precipitation intensity distribution and outputs precipitation forecast results for subsequent prediction and analysis. The multi-scale feature representation constructed from the C4 output is as follows: in, Indicates the first Each scale has a feature transformation operator; different scale features correspond to different spatial resolutions, which are used to support the subsequent multi-scale feature reconstruction process.

[0064] To fuse multi-scale features into a unified resolution representation, a scale-aware feature fusion strategy is introduced. Features at different scales are upsampled to the target spatial resolution and then weighted and fused, as shown below: in, This represents the upsampling operator, used to restore spatial resolution; This represents element-wise multiplication; For the first The weight tensor corresponding to each scale.

[0065] Scale weight Adaptively generated from global statistical information of physically enhanced features, as follows: in, This indicates the output of the physics enhancement module; This is the normalized mapping function. This mechanism is used to dynamically adjust the contribution of physical features at different scales to the reconstruction of precipitation distribution.

[0066] To further suppress non-physical oscillations and enhance feature robustness, a residual gating enhancement mechanism is introduced into the fused features: in: and Represents the characteristic transformation operator; This indicates a normalization operation; the residual structure is used to maintain the stability of the physical enhancement features during the reconstruction process.

[0067] Based on enhanced feature representation Generate precipitation forecasts for future time periods: in, Represents the physical field of predicted precipitation; This represents the historical precipitation observations at the end of the input sequence; This represents the precipitation prediction mapping function.

[0068] In an optional implementation, the prediction mapping function can be expressed as: in, This indicates the feature concatenation operation. The feature reconstruction and precipitation prediction module is used in the model inference phase to generate precipitation prediction results for the target time period. It should be noted that during the model training phase, the output of the model structure involved in A400 is the same as the output of S200, and together with the true values ​​of precipitation observations for the corresponding time period, they are used to construct the model training loss function in order to optimize the model parameters.

[0069] In this embodiment of the application, the preprocessing stage S200 and the inference stage A400 further include: In response to the model outputting preliminary prediction results or continuous prediction results of future precipitation intensity, several precipitation intensity thresholds are set; For example, let the actual precipitation intensity at the future target time / period be... Unit: mm / forecast period; Threshold setting is expressed as: The future precipitation is divided into multiple mutually exclusive precipitation level intervals at each spatial grid point according to the precipitation intensity threshold, and the probability of occurrence of each precipitation level interval is predicted based on the spatiotemporal fusion features shared by the model. For example, the intervals are divided into four mutually exclusive categories based on the threshold, as follows: in, Representing spatial grid points The corresponding category label for the actual precipitation level range.

[0070] To facilitate multi-class probability prediction and supervised learning, the class labels are represented as one-hot label vectors: Wherein, the vector of the first Each component Used to indicate grid points Does the precipitation intensity at this location belong to the first category? The range of magnitudes is defined as follows: That is, in the label vector, only the component corresponding to the actual precipitation level range is set to 1, and the rest are set to 0, which is used to clearly indicate the actual precipitation level category of the grid point.

[0071] Based on spatiotemporal feature representation Constructing a probability prediction head for a range of magnitude Output the classification logits corresponding to each spatial grid point: The interval probability distribution is obtained through softmax mapping: It should be noted that during the model inference (A400) stage, the interval probability distribution of each spatial grid point is directly output. Furthermore, it can generate threshold event probability products; for example: probability of heavy rainfall events (≥50): Probability of extreme precipitation events (≥100): The probability results can be used to generate heavy precipitation risk distribution maps and extreme precipitation early warning products.

[0072] In summary, this invention, in terms of model structure, groups and models meteorological variables according to their physical functions, and accurately describes the physical process of precipitation formation through multi-scale circulation fusion and spatiotemporal joint modeling. At the training and supervision level, it introduces a physical consistency constraint loss based on vertical integral water vapor transport, forcing the underlying physical field predicted by the model to conform to the laws of atmospheric motion. In terms of output, it employs a multi-task learning framework to jointly optimize precipitation intensity regression and precipitation magnitude probability classification, enabling the simultaneous output of physically more reliable continuous precipitation forecasts and valuable extreme precipitation probability warnings, thus improving the physical rationality of the forecast results.

[0073] Example 3 illustrates a schematic scheme for a precipitation prediction and probabilistic forecasting method based on physical consistency constraints. It should be noted that the technical solution of this precipitation prediction and probabilistic forecasting system based on physical consistency constraints is based on the same concept as the aforementioned precipitation prediction and probabilistic forecasting method based on physical consistency constraints. Details not described in detail in this example can be found in the description of the aforementioned precipitation prediction and probabilistic forecasting method based on physical consistency constraints.

[0074] This embodiment also provides another precipitation prediction and probabilistic forecasting system based on physical consistency constraints, including: The acquisition module is used to acquire multivariate meteorological data of the target area within a preset time window; The physical enhancement feature construction module is used to input the preprocessed second meteorological data into the trained precipitation prediction model, perform structured modeling of the physical quantities in the input data, and extract physical enhancement features. The multi-scale circulation feature fusion module is used to perform multi-scale circulation feature fusion on the physical enhancement features, and extract and fuse atmospheric circulation structure features at different spatial scales respectively. The spatiotemporal feature fusion module is used to jointly model the spatial structure and temporal evolution information of multi-scale circulation features to form spatiotemporal fused features. The feature reconstruction and prediction module is used to perform feature reconstruction based on the spatiotemporal fusion features and output continuous prediction results of future precipitation intensity.

[0075] Based on the above system modules, it also includes: The physical consistency constraint module is used to calculate the vertical integral of water vapor flux based on the specific humidity and horizontal wind field components in the preliminary prediction results, and the difference between the corresponding calculation amount and the vertical integral of the actual water vapor flux in the training data is used as the physical consistency constraint to participate in the model training. The classification prediction loss construction module is used to construct the classification prediction loss for precipitation level intervals based on the preliminary prediction results. The joint loss construction module is used to jointly optimize the model parameters based on precipitation prediction loss, combined with physical consistency constraint loss and / or classification prediction loss, to obtain the trained precipitation prediction model. This embodiment also provides a computer device applicable to a precipitation prediction and probability forecasting method based on physical consistency constraints, comprising: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the precipitation prediction and probability forecasting method based on physical consistency constraints as proposed in the above embodiment.

[0076] This embodiment also provides a storage medium on which a computer program is stored. When the program is executed by a processor, it implements a precipitation prediction and probability forecasting method based on physical consistency constraints as proposed in the above embodiments.

[0077] The storage medium proposed in this embodiment and the method for implementing precipitation prediction and probabilistic forecasting based on physical consistency constraints proposed in the above embodiments belong to the same inventive concept. Technical details not described in detail in this embodiment can be found in the above embodiments, and this embodiment has the same beneficial effects as the above embodiments.

[0078] Based on the above description of the implementation methods, those skilled in the art will clearly understand that the present invention can be implemented using software and necessary general-purpose hardware. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as a computer floppy disk, read-only memory (ROM), random access memory (RAM), flash memory, hard disk, or optical disk, etc., including several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods of the various embodiments of the present invention.

[0079] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A precipitation prediction and probabilistic forecasting method based on physical consistency constraints, characterized in that, include: Precipitation prediction model training phase and forecast inference phase; The precipitation prediction model training phase includes: Acquire multivariate meteorological data for the target area within a preset time window; The preprocessed first meteorological data is input into the precipitation prediction model to obtain preliminary prediction results; the precipitation prediction model includes at least a physical enhancement feature construction module, a multi-scale circulation feature fusion module, a spatiotemporal feature fusion module, and a feature reconstruction and prediction module connected in sequence; Based on the specific humidity and horizontal wind field components in the preliminary prediction results, the vertical integral of water vapor flux is calculated, and the difference between the corresponding calculation amount and the vertical integral of the actual water vapor flux in the training data is used as a physical consistency constraint to participate in the model training. Based on the preliminary forecast results, a classification-based predicted loss is constructed for precipitation level intervals; Based on the precipitation prediction loss, combined with the physical consistency constraint loss and / or classification prediction loss, the model parameters are jointly optimized to obtain the trained precipitation prediction model. The forecast inference stage includes: The preprocessed second meteorological data is input into the trained precipitation prediction model, and the physical quantities in the input data are structured and modeled to extract physical enhancement features. Multi-scale circulation feature fusion is performed on the physical enhancement features to extract and fuse atmospheric circulation structure features at different spatial scales. Joint modeling of spatial structure and temporal evolution information of multi-scale circulation characteristics forms spatiotemporal fusion characteristics; Based on the spatiotemporal fusion features, feature reconstruction is performed, and continuous prediction results of future precipitation intensity are output.

2. The precipitation prediction and probabilistic forecasting method based on physical consistency constraints as described in claim 1, characterized in that, The preprocessed second meteorological data is input into the trained precipitation prediction model. The physical quantities in the input data are then structured and modeled to extract physical enhancement features, including: Based on pre-defined meteorological and physical priors, the input physical quantities are divided along the channel dimension into... Each physical group yields the recombined physical tensor; For the recombined physical tensor, a combination of channel transformation and spatial convolution is used to extract features to obtain physical enhancement features.

3. The precipitation prediction and probabilistic forecasting method based on physical consistency constraints as described in claim 2, characterized in that, The physical enhancement features are subjected to multi-scale circulation feature fusion, which extracts and fuses atmospheric circulation structure features at different spatial scales, including: Based on feature transformation operators with different spatial perception ranges, the physical enhancement features are processed to extract multi-scale circulation features on a continuous spatial scale spectrum to characterize atmospheric circulation and precipitation-related processes. The spatial scale spectrum covers different spatial scales from planetary scales to local disturbances. Based on the global statistical information of the physical enhancement features, scale weights corresponding to each spatial scale are adaptively generated, and the circulation features at different spatial scales are weighted and fused based on the scale weights to form a unified multi-scale circulation feature representation.

4. The precipitation prediction and probabilistic forecasting method based on physical consistency constraints as described in claim 3, characterized in that, Joint modeling of spatial structure and temporal evolution information of multi-scale circulation characteristics forms spatiotemporal fusion features, including: Spatial convolution operators are applied to multi-scale circulation features to obtain spatially enhanced features; Time-enhanced features are obtained by processing multi-scale circulation features using time feature modeling operators; By fusing spatial augmentation features with temporal augmentation features, preliminary spatiotemporal joint features are obtained; The preliminary spatiotemporal joint features are subjected to channel projection and residual enhancement processing to obtain spatiotemporal fusion features.

5. The precipitation prediction and probabilistic forecasting method based on physical consistency constraints as described in claim 4, characterized in that, Based on the specific humidity and horizontal wind field components in the preliminary prediction results, the vertical integral of water vapor flux is calculated, and the difference between the corresponding calculated amount and the vertical integral of the actual water vapor flux in the training data is used as a physical consistency constraint in model training, including: Based on the model-predicted specific humidity and horizontal wind field components, the vertical integral water vapor transport vector is calculated. Based on the consistency relationship between the vertical integral water vapor transport vector and the actual water vapor transport volume, a physical consistency constraint loss term is constructed. , is represented as: in, This represents a distance or difference metric function used to measure the consistency between the two. Represents the vertical integral water vapor transport vector; This indicates the actual amount of water vapor transported.

6. The precipitation prediction and probabilistic forecasting method based on physical consistency constraints as described in claim 5, characterized in that, Based on preliminary forecast results, a classification-based predicted loss is constructed for precipitation level intervals, including: While outputting preliminary prediction results, the probability prediction of precipitation level ranges is carried out based on the preliminary spatiotemporal fusion features generated during the training phase. A classification loss function is introduced to conduct supervised learning on the precipitation level intervals of each grid point, thereby constructing the classification prediction loss.

7. The precipitation prediction and probabilistic forecasting method based on physical consistency constraints as described in claim 6, characterized in that, It also includes setting several precipitation intensity thresholds in response to the model outputting preliminary prediction results or continuous prediction results of future precipitation intensity; Future precipitation is divided into multiple mutually exclusive precipitation level intervals at each spatial grid point according to the precipitation intensity threshold. Based on the spatiotemporal fusion features shared by the model, the probability of each precipitation level interval is predicted and the occurrence probability of each precipitation level interval is output.

8. A precipitation prediction and probabilistic forecasting system based on physical consistency constraints, using the method described in any one of claims 1-7, characterized in that, include: The acquisition module is used to acquire multivariate meteorological data of the target area within a preset time window; The physical enhancement feature construction module is used to input the preprocessed second meteorological data into the trained precipitation prediction model, perform structured modeling of the physical quantities in the input data, and extract physical enhancement features. The multi-scale circulation feature fusion module is used to perform multi-scale circulation feature fusion on the physical enhancement features, and extract and fuse atmospheric circulation structure features at different spatial scales respectively. The spatiotemporal feature fusion module is used to jointly model the spatial structure and temporal evolution information of multi-scale circulation features to form spatiotemporal fused features. The feature reconstruction and prediction module is used to perform feature reconstruction based on the spatiotemporal fusion features and output continuous prediction results of future precipitation intensity.

9. A computer device, characterized in that, include: Memory and processor; The memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions. When the computer-executable instructions are executed by the processor, they implement the steps of the precipitation prediction and probability forecasting method based on physical consistency constraints as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, It stores computer-executable instructions, which, when executed by a processor, implement the steps of the precipitation prediction and probability forecasting method based on physical consistency constraints as described in any one of claims 1 to 7.

Images