A fusion rainfall prediction enhancement method and system based on Beidou GNSS and rain radar

Abstract

The application provides a fusion rainfall prediction enhancement method and system based on Beidou GNSS and a rain radar, and the method comprises the following steps: acquiring radar data and precipitable water vapor (PWV) data of a target area, and extracting a radar observation tensor and a Beidou / GNSS water vapor tensor; extracting water vapor field composite features according to the Beidou / GNSS water vapor tensor, and obtaining water vapor dynamics features by performing nonlinear coding on the water vapor field composite features; inputting the radar observation tensor and the Beidou / GNSS water vapor tensor into a trained fusion rainfall prediction enhancement model, and outputting a fusion rainfall prediction result of the target area.

Description

Technical Field

[0001] This invention belongs to the field of rainfall prediction, specifically relating to a fusion rainfall prediction enhancement method and system based on BeiDou GNSS and rain-measuring radar. Background Technology

[0002] Rainfall radar is the primary source of observational data for short-term precipitation forecasting, providing extremely high spatiotemporal resolution in capturing the spatial structure of precipitation systems. Radar-based short-term precipitation forecasting has been widely applied in meteorology and hydrology. However, traditional radar-based short-term forecasting methods have significant limitations. While traditional physical extrapolation methods, such as optical flow, can effectively capture the trajectory of radar echoes, they are essentially based on kinematic assumptions and perform poorly in modeling the physical growth and dissipation processes of precipitation systems, making it difficult to predict the nonlinear changes of convective systems. Existing deep learning-based radar precipitation forecasting models still face a fundamental flaw: these methods rely heavily on single-source radar data and, to a large extent, on kinematic principles, predicting future states solely based on the historical trajectory of radar echoes, lacking supplementary consideration of atmospheric environmental physical variables. For severe convective weather systems, single radar echo kinematic extrapolation is insufficient to accurately predict their formation and dissipation. To overcome this limitation, introducing auxiliary atmospheric physical variables becomes an inevitable choice. Precipitable water vapor (PWV) retrieved from Global Navigation Satellite Systems (GNSS) provides continuous and high-resolution spatiotemporal variations in atmospheric water vapor. Water vapor is a key precursor to convection development and an important physical constraint on precipitation evolution. Theoretically, fusing high-resolution precipitation radar data with PWV data retrieved from GNSS can significantly improve prediction accuracy.

[0003] To address this issue, this invention proposes a fusion precipitation prediction enhancement method based on water vapor inversion from BeiDou / GNSS and precipitation radar. This method designs a plug-and-play water vapor control module. This module can seamlessly adapt to and integrate with any pre-trained deep learning radar precipitation prediction model with an encoder-decoder structure (e.g., conventional models based on the U-Net architecture, or generative models such as NowcastNet and DiffCast). This invention integrates multi-scale atmospheric water vapor features retrieved from BeiDou / GNSS into a frozen pre-trained backbone network, thereby endowing the general model with enhanced ability to perceive the influence of water vapor on precipitation evolution without compromising its original powerful radar extrapolation capabilities. Summary of the Invention

[0004] To address the shortcomings of existing precipitation prediction models, such as insufficient ability to predict the formation and dissipation of convective systems, severe decline in model accuracy over long periods, and poor fusion of multi-source data, this invention provides a fusion precipitation prediction enhancement method and system based on BeiDou GNSS and rain measuring radar. By setting up a plug-and-play water vapor control module, the multi-scale atmospheric water vapor characteristics retrieved from BeiDou / GNSS are integrated into a pre-trained backbone network in a frozen state. This enhances the ability of the general model to perceive the evolution of precipitation without compromising the original powerful radar extrapolation capability.

[0005] According to one aspect of the present invention, a fusion rainfall prediction enhancement method based on BeiDou GNSS and rainfall radar is provided, comprising the following steps:

[0006] Acquire radar data and precipitable water vapor (PWV) data for the target area, and extract radar observation tensor and BeiDou / GNSS water vapor tensor;

[0007] Based on the water vapor tensor of Beidou / GNSS, the composite features of the water vapor field are extracted, and the water vapor dynamic features are obtained by nonlinear encoding of the composite features of the water vapor field.

[0008] The radar observation tensor and the BeiDou / GNSS water vapor tensor are input into the trained fusion precipitation prediction enhancement model, and the fusion precipitation prediction results for the target area are output.

[0009] The training steps for the fusion rainfall prediction enhancement model include:

[0010] Obtain training samples for the fusion rainfall prediction enhancement model;

[0011] Construct a fusion-enhanced rainfall prediction model;

[0012] A heavy precipitation sample enhancement strategy is introduced to optimize the training samples, and the optimized training samples are used to train the fusion precipitation prediction enhancement model, outputting the trained model.

[0013] As a further technical solution, the radar observation tensor and the BeiDou / GNSS water vapor tensor are extracted, and the steps include:

[0014] Based on the observation timestamp of radar data, the precipitable water vapor (PWV) data is time-synchronized, and the time-synchronized PWV data is spatially resampled to generate a gridded PWV water vapor field that is spatiotemporally aligned with the radar data.

[0015] The radar observation tensor is composed of radar data within a preset continuous time step; the gridded PWV water vapor field within the preset continuous time step is selected to form the BeiDou / GNSS water vapor tensor.

[0016] As a further technical solution, the composite features of the water vapor field are extracted based on the BeiDou / GNSS water vapor tensor, and the composite features of the water vapor field are nonlinearly encoded to obtain the water vapor dynamic features. The steps include:

[0017] The spatial gradient field and the temporal rate of change field of water vapor are calculated based on the BeiDou / GNSS water vapor tensor, and the three are spliced ​​together in the channel dimension to obtain the composite water vapor feature tensor.

[0018] The water vapor dynamics characteristics are obtained by nonlinearly encoding the composite water vapor feature tensor using a three-dimensional spatiotemporal convolutional network.

[0019] As a further technical solution, the training samples of the integrated rainfall prediction enhancement model include radar observation tensor, BeiDou / GNSS water vapor tensor and precipitation prediction target tensor; wherein, the precipitation prediction target tensor is used to characterize the precipitation prediction target in the future preset number of consecutive time steps after the preset number of consecutive time steps, including precipitation intensity and precipitation grid area ratio.

[0020] As a further technical solution, the fusion precipitation prediction enhancement model includes a basic backbone network and a water vapor control module. The encoder of the basic backbone network processes the radar observation tensor to obtain multi-level radar coded features; the water vapor control module processes the water vapor dynamic features to obtain multi-level water vapor regulation features, and the multi-level water vapor regulation features are introduced layer by layer during the decoding of the multi-level radar coded features in the basic backbone network, finally outputting the precipitation prediction result. Among them, the basic backbone network adopts a pre-trained precipitation prediction deep learning model with an encoder-decoder structure, and the encoder parameters of the basic backbone network are frozen.

[0021] As a further technical solution, the water vapor control module is configured as a plug-and-play module, including a homogeneous feature mapping layer and a cross-layer regulation output layer. The homogeneous feature mapping layer adopts a coding branch that is homogeneous with the basic backbone network encoder, receives water vapor dynamic features, and outputs water vapor regulation features that match different levels of the basic backbone network. The cross-layer regulation output layer uses an alignment injection operator to inject water vapor regulation features of different levels into the corresponding levels of the basic backbone network decoder.

[0022] As a further technical solution, an alignment injection operator is used to inject water vapor regulation features of different levels into the corresponding levels of the basic backbone network decoder. The steps include:

[0023] The alignment injection operator converts the water vapor regulation feature of any level into a feature tensor with the same dimension as the decoding feature of the corresponding level of the basic backbone network decoder, and fuses the feature tensor into the decoding feature of the corresponding level to inject water vapor regulation, thereby obtaining the fused decoding feature after water vapor regulation at that level.

[0024] As a further technical solution, a heavy precipitation sample enhancement strategy is introduced to optimize the heavy precipitation samples. The steps include:

[0025] Based on the precipitation prediction tensor of each training sample, the maximum precipitation intensity, average precipitation intensity, and the percentage of the area with the strongest precipitation in the precipitation prediction target of each training sample are obtained, and the importance score of each training sample is obtained by weighted fusion.

[0026] Based on the importance scores of each training sample, heavy precipitation samples are selected from the training samples, and the sampling probability of each training sample is adjusted so that heavy precipitation samples have a higher sampling probability during the training process.

[0027] According to another aspect of this specification, a fusion rainfall prediction enhancement system based on BeiDou GNSS and rainfall radar is provided, comprising:

[0028] The input module is used to acquire radar data and precipitable water vapor (PWV) data of the target area, and extract radar observation tensors and BeiDou / GNSS water vapor tensors.

[0029] The water vapor dynamics coding module is used to extract the composite features of the water vapor field based on the BeiDou / GNSS water vapor tensor, and to obtain the water vapor dynamic features by nonlinear coding of the composite features of the water vapor field.

[0030] The fusion precipitation prediction module is used to input radar observation tensors and BeiDou / GNSS water vapor tensors into the trained fusion precipitation prediction enhancement model and output the fusion precipitation prediction results for the target area.

[0031] The training steps for the fusion rainfall prediction enhancement model include:

[0032] Obtain training samples for the fusion rainfall prediction enhancement model;

[0033] A fusion-enhanced rainfall prediction model is constructed. The fusion-enhanced rainfall prediction model includes a basic backbone network and a water vapor control module. The encoder of the basic backbone network processes the radar observation tensor to obtain multi-level radar coding features. The water vapor control module processes the water vapor dynamic features to obtain multi-level water vapor regulation features. The multi-level water vapor regulation features are introduced layer by layer during the decoding of the multi-level radar coding features in the decoder of the basic backbone network, and finally the rainfall prediction result is output.

[0034] A heavy precipitation sample enhancement strategy is introduced to optimize the training samples, and the optimized training samples are used to train the fusion precipitation prediction enhancement model, outputting the trained model.

[0035] According to another aspect of this specification, an electronic device is provided, including a memory and a processor, the memory storing program instructions executed by the processor, the processor invoking the program instructions to execute a fusion rainfall prediction enhancement method based on BeiDou GNSS and rain-measuring radar.

[0036] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0037] This method designs a plug-and-play water vapor control module that can seamlessly adapt to and connect to any pre-trained deep learning radar precipitation prediction model with an encoder-decoder structure. It integrates the multi-scale atmospheric water vapor characteristics retrieved from BeiDou / GNSS into the frozen pre-trained backbone network, thereby giving the general model enhanced ability to perceive the influence of water vapor on precipitation evolution without compromising its original powerful radar extrapolation capabilities.

[0038] This invention employs a fusion rainfall prediction enhancement method based on BeiDou GNSS and rain-measuring radar, which improves the performance of quantitative prediction based on indicators and can accurately predict the evolution and enhancement of the core of heavy precipitation. Attached Figure Description

[0039] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the accompanying drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0040] Figure 1 A flowchart illustrating a fusion rainfall prediction enhancement method based on BeiDou GNSS and rain-measuring radar, provided for an embodiment of the present invention;

[0041] Figure 2 This is a visual comparison of precipitation prediction results between the method of the present invention and the baseline model under different forecast lead times in an embodiment of the present invention;

[0042] Figure 3 This is a line graph showing the quantitative evaluation results of the method and baseline model of the present invention under different forecast lead times, as described in this embodiment of the invention.

[0043] Figure 4 A schematic diagram of the structure of a fusion rainfall prediction enhancement system based on BeiDou GNSS and rain measuring radar provided in an embodiment of the present invention;

[0044] Figure 5 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0045] It should be noted that:

[0046] The terms “comprising” and “having”, and any variations thereof, in the specification, claims, and accompanying drawings of this invention are intended to cover a non-exclusive inclusion, such as a process, method, system, product, or apparatus that includes a series of steps or units, not necessarily limited to those explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0047] The block diagrams shown in the accompanying drawings are merely functional entities and do not necessarily correspond to physically independent entities. That is, these functional entities can be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor devices and / or microcontroller devices. The flowcharts shown in the accompanying drawings are merely illustrative and do not necessarily include all content and operations / steps, nor do they necessarily have to be performed in the described order. For example, some operations / steps can be decomposed, while others can be combined or partially combined; therefore, the actual execution order may change depending on the specific circumstances.

[0048] 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 only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. In addition, the technical features of the various embodiments or individual embodiments provided by the present invention can be arbitrarily combined to form new technical solutions. Such combinations are not bound by the order of steps and / or structural composition patterns, but must be based on the ability of those skilled in the art to implement them. When the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection claimed by the present invention.

[0049] like Figure 1 As shown, a fusion rainfall prediction enhancement method based on BeiDou GNSS and rainfall radar includes the following steps:

[0050] S1. Acquire radar data and precipitable water vapor (PWV) data of the target area, and extract radar observation tensor and BeiDou / GNSS water vapor tensor.

[0051] S2. Extract water vapor field composite features based on BeiDou / GNSS water vapor tensor, and obtain water vapor dynamic features by nonlinear encoding of water vapor field composite features;

[0052] S3. Input the radar observation tensor and the BeiDou / GNSS water vapor tensor into the trained fusion precipitation prediction enhancement model, and output the fusion precipitation prediction results for the target area;

[0053] Specifically, in S1, radar data includes radar quantitative precipitation estimation data or radar echo intensity data for the target area, which are acquired by rain-measuring radar; precipitable water vapor (PWV) data are acquired from BeiDou / GNSS stations and have spatial discrete characteristics.

[0054] Preferably, S1 further includes quality control of the acquired radar data and precipitable water vapor (PWV) data:

[0055] Quality control of radar data: Outliers, missing values ​​and noise points in radar data are processed using physical threshold removal, neighborhood consistency checks, linear interpolation, bilinear interpolation or spatiotemporal proximity averaging methods.

[0056] Quality control of precipitable water vapor (PWV) data: Abnormal PWV values ​​appearing in the reference stations are screened by single-station time series consistency, spatial consistency between adjacent stations, or meteorological and physical thresholds.

[0057] In S1, the steps for extracting the radar observation tensor and the BeiDou / GNSS water vapor tensor include:

[0058] S11. Based on the observation timestamp of the radar data, the precipitable water vapor (PWV) data is time-synchronized, and the time-synchronized PWV data is spatially resampled to generate a gridded PWV water vapor field that is spatiotemporally aligned with the radar data.

[0059] S12. The radar observation tensor is composed of radar data within a preset continuous time step; the gridded PWV water vapor field within the preset continuous time step is selected to form the BeiDou / GNSS water vapor tensor.

[0060] Specifically, in S21, spatial resampling resamples the time-synchronized precipitable water vapor (PWV) data to a two-dimensional plane aligned with the radar grid: interpolation methods are used to generate a gridded PWV water vapor field (two-dimensional) with the same spatial extent, projection method, spatial resolution, and number of rows and columns as the radar data. Optionally, the interpolation method can be Kriging interpolation, inverse range weighted interpolation, radial basis function interpolation, or bilinear interpolation.

[0061] Specifically, in S22, radar data from multiple consecutive historical time steps and the gridded PWV water vapor field organization are used as input tensors, where the radar observation tensor... From history The radar data consists of several consecutive time steps, mathematically represented as:

[0062]

[0063] In the formula, and These represent the height and width of the grid space, respectively.

[0064] BeiDou / GNSS Water Vapor Tensor From the history described The gridded PWV water vapor field consists of several consecutive time steps, mathematically represented as:

[0065]

[0066] After completing the PWV water vapor field grid reconstruction, in order to deeply explore the triggering mechanism of water vapor convergence and divergence on strong convection, this method does not directly use the precipitable water vapor PWV data as a simple superposition of ordinary channels. Instead, it further extracts the water vapor spatial gradient, water vapor temporal change rate and composite water vapor dynamic characteristics in S2, so that the model can perceive the water vapor distribution boundary, water vapor transport enhancement and water vapor convergence process.

[0067] In S2, the composite features of the water vapor field are extracted based on the BeiDou / GNSS water vapor tensor, and the water vapor dynamic features are obtained by nonlinear encoding of the composite features of the water vapor field. The steps include:

[0068] S21. Calculate the water vapor spatial gradient field and water vapor temporal change rate field based on the Beidou / GNSS water vapor tensor, and then stitch the three together in the channel dimension to obtain the composite water vapor feature tensor.

[0069] S22. The composite water vapor feature tensor is nonlinearly encoded using a three-dimensional spatiotemporal convolutional network to obtain the water vapor dynamic features.

[0070] In S21, the water vapor spatial gradient is used to characterize regions with drastic changes in water vapor content. These regions typically correspond to water vapor fronts, wet-dry boundaries, or water vapor transport boundaries, and are indicative of the triggering of severe convective precipitation. Its calculation process includes: for any time step in the BeiDou / GNSS water vapor tensor... The gridded PWV water vapor field was calculated in the east-west direction. (direction) and north-south direction ( The gradient in the direction is calculated, and the spatial gradient of water vapor is further calculated. The calculation process is shown in the following formula:

[0071]

[0072] In the formula, Represents the gridded PWV water vapor field at time step t. direction and The spatial gradient of water vapor in the direction.

[0073] In S21, the water vapor time variation rate is calculated based on the difference between the gridded PWV water vapor fields of adjacent time steps, and the calculation formula is as follows:

[0074]

[0075] In the formula, This represents the rate of change of water vapor over time in the gridded PWV water vapor field at time step t.

[0076] The water vapor temporal variation rate (PWV) is used to characterize the rapid increase or decrease of water vapor over time. When the PWV of a certain region increases continuously, it usually indicates that there is enhanced water vapor transport or water vapor convergence in that region; when the PWV decreases rapidly, it may correspond to precipitation release, the passage of a convective system, or a weakening of water vapor conditions.

[0077] Furthermore, in S21, the gridded PWV water vapor field, water vapor spatial gradient field, and water vapor temporal rate of change field are spliced ​​together along the channel dimension to form a composite water vapor feature tensor. As shown in the following formula:

[0078] ;

[0079] In the formula, Concat represents the channel dimension concatenation operation.

[0080] In S22, the mathematical representation of the water vapor dynamics characteristics is as follows:

[0081] ;

[0082] In the formula, Represents the characteristic tensor of water vapor dynamics; Represents a three-dimensional convolution operation; and These represent the learnable convolutional weights and biases, respectively. This represents a non-linear activation function.

[0083] The obtained water vapor dynamics features not only include absolute PWV content information, but also water vapor spatial gradient, water vapor variation trend and water vapor convergence enhancement information, providing physical constraints for subsequent fusion prediction models.

[0084] In S3, the training steps for the fusion rainfall prediction enhancement model include:

[0085] S31. Obtain training samples for the fusion rainfall prediction enhancement model;

[0086] S32. Construct a fusion-enhanced rainfall prediction model;

[0087] S33. Introduce a heavy precipitation sample enhancement strategy to optimize the training samples, and use the optimized training samples to train the fusion precipitation prediction enhancement model, and output the trained model.

[0088] In step S31, the training samples of the fusion precipitation prediction enhancement model include radar observation tensor, BeiDou / GNSS water vapor tensor and precipitation prediction target tensor; wherein, the precipitation prediction target tensor is used to characterize the precipitation prediction target, i.e., precipitation intensity, within a future preset number of consecutive time steps after the preset number of consecutive time steps.

[0089] Specifically, the fused precipitation prediction enhancement model was trained, validated, and tested using historical radar data, BeiDou / GNSS water vapor data, and real-time future precipitation data. Training samples were divided into training, validation, and test sets according to chronological order. To avoid information leakage caused by adjacent time samples, the sets were further divided by year, month, independent precipitation processes, or independent extreme events.

[0090] The precipitation prediction target tensor is used as the output to characterize the future. The precipitation forecast target at each time step is mathematically represented as:

[0091]

[0092] in, This represents the target tensor for precipitation forecasting; Indicates the predicted number of future time steps; Indicates the number of output channels.

[0093] The training samples thus constructed can be mathematically represented as follows:

[0094] .

[0095] In S32, the fusion rainfall prediction enhancement model includes a basic backbone network and a water vapor control module. The encoder of the basic backbone network processes the radar observation tensor to obtain multi-level radar coding features. The water vapor control module processes the water vapor dynamic features to obtain multi-level water vapor regulation features. During the decoding of the multi-level radar coding features by the decoder of the basic backbone network, the multi-level water vapor regulation features are introduced layer by layer, and finally the rainfall prediction result is output.

[0096] The basic backbone network adopts a pre-trained precipitation prediction deep learning model with an encoder-decoder structure, and the encoder parameters of the basic backbone network are frozen.

[0097] Specifically, the basic backbone network has been pre-trained based on a large amount of single rain-measuring radar data, and already possesses strong capabilities in extrapolating radar echo trajectories and modeling precipitation spatial structures. Its entire encoder parameter weight matrix... The network remains frozen during subsequent training. Optionally, the underlying backbone network can be a U-Net-like model, a ConvLSTM-like model, a PredRNN-like model, a NowcastNet model, a DiffCast model, a diffusion-generative model, or other deep learning models capable of predicting future precipitation fields from historical radar sequences.

[0098] Specifically, the radar observation tensor is input into the basic backbone network encoder of the frozen parameters to obtain multi-level radar coding features, mathematically represented as:

[0099]

[0100] In the formula, Indicates network hierarchy; express Layered radar coding characteristics; These are the parameters for the basic backbone network encoder.

[0101] Secondly, the water vapor control module is configured as a plug-and-play module, including a homogeneous feature mapping layer and a cross-layer regulation output layer. The homogeneous feature mapping layer adopts a coding branch that is homogeneous with the basic backbone network encoder, receives water vapor dynamic features, and outputs water vapor regulation features that match different levels of the basic backbone network. The cross-layer regulation output layer uses an alignment injection operator to inject water vapor regulation features of different levels into the corresponding levels of the basic backbone network decoder.

[0102] Specifically, the water vapor control module includes the following key mechanisms:

[0103] Firstly, the isomorphic feature mapping layer: Structurally, this branch instantiates a network structure with the exact same hyperparameter levels (such as network depth and number of channels) as the encoder of the frozen backbone network, but its parameter matrix is ​​different. Independently initialized and fully differentiable. The water vapor characteristics obtained in the second step. Input this branch to extract multi-level water vapor regulation features. .

[0104] After the water vapor dynamics characteristics are input into the isomorphic feature mapping layer of the water vapor control module, multi-scale water vapor regulation characteristics are obtained, as shown in the following equation:

[0105] = ;

[0106] In the formula, express Layer water vapor regulation characteristics; These are the parameters for the water vapor control module.

[0107] Secondly, in order to achieve lossless injection of water vapor regulation features into the backbone decoder, this method designs an alignment injection operator in the cross-layer regulation output layer. .

[0108] The mathematical representation of the alignment injection operator is:

[0109] ;

[0110] In the formula, Indicates the level depth; Indicates the first Hierarchical water vapor regulation characteristics; Indicates the first The feature tensor after transformation by the hierarchical alignment injection operator; This represents the convolution operation; , For the first The parameters for the hierarchical alignment injection operator are the weight parameter matrix and the bias parameter matrix, respectively.

[0111] The process involves using an alignment injection operator to inject water vapor regulation features from different levels into the corresponding levels of the basic backbone network decoder. The steps include:

[0112] The alignment injection operator converts the water vapor regulation feature of any level into a feature tensor with the same dimension as the decoding feature of the corresponding level of the basic backbone network decoder, and fuses the feature tensor into the decoding feature of the corresponding level to inject water vapor regulation, thereby obtaining the fused decoding feature after water vapor regulation at that level.

[0113] Specifically, for the Lth level, the alignment injection operator of the Lth level converts the water vapor regulation features of the Lth level into a feature tensor with the same dimension as the decoded features of the Lth level of the decoder, mathematically represented as:

[0114]

[0115] In the formula, Indicates the first Layer alignment injection operator; and They represent the first Learnable weights and biases of the alignment injection operator for the layer.

[0116] The feature tensor obtained by transforming the water vapor regulation features of the Lth layer is fused into the original decoded features to inject water vapor regulation. The mathematical representation of the fused decoded features is as follows:

[0117]

[0118] in, Indicates the first Decoding characteristics of layered backbone network decoders; Indicates the first Fusion decoding features after layer fusion and water vapor regulation.

[0119] Through the alignment injection mechanism provided by the alignment injection operator, BeiDou / GNSS water vapor information can be injected into the pre-trained radar prediction model in a progressive and controllable manner, so that the model can enhance its ability to characterize the growth, enhancement, weakening and dissipation of precipitation while maintaining the original radar motion extrapolation capability.

[0120] In S33, to address the issue that the proportion of samples of heavy rain, short-duration heavy precipitation, and severe convection is relatively low in all samples, this method introduces a heavy precipitation sample enhancement strategy. This strategy enables the model to learn more about heavy rain, severe convection, and local heavy precipitation processes, thereby improving the predictive ability for extreme precipitation events.

[0121] A heavy precipitation sample enhancement strategy is introduced to optimize the heavy precipitation samples. The steps include:

[0122] Based on the precipitation prediction tensor of each training sample, the maximum precipitation intensity, average precipitation intensity, and the percentage of the area with the strongest precipitation in the precipitation prediction target of each training sample are obtained, and the importance score of each training sample is obtained by weighted fusion.

[0123] Based on the importance scores of each training sample, heavy precipitation samples are selected from the training samples, and the sampling probability of each training sample is adjusted so that heavy precipitation samples have a higher sampling probability during the training process.

[0124] The weighted fusion yields the importance score for each training sample, mathematically represented as:

[0125]

[0126] In the formula, Indicates the first The importance score of each training sample; This indicates the maximum precipitation intensity in the precipitation prediction target of this training sample; This represents the average precipitation intensity in the precipitation prediction target of the training sample. This indicates the percentage of the area of ​​the strongest precipitation grid point in the precipitation prediction target of the training sample; , , These are adjustable weighting coefficients.

[0127] Specifically, during the training of S33, the parameters of the backbone network are frozen first, and only the water vapor control module and the alignment injection operator are trained. After the water vapor control module converges, the backbone network is kept frozen, or only the back layers of the backbone network decoder are unfrozen, and joint fine-tuning is performed with a small learning rate.

[0128] During the network training initialization phase, the parameters of the alignment injection operator are initialized with zero: constraining the parameters of the alignment operator. and This design allows the feature tensors injected into the backbone decoder in the initial stage. The zero matrix ensures that the initial forward inference state of the fusion system is equivalent to a pure radar pre-trained model. As the training process progresses, the parameters are guided to evolve smoothly during fine-tuning, gradually learning the adjustment weight of water vapor dynamic characteristics on precipitation prediction results.

[0129] To enable the model to accurately capture the peak values ​​of extreme precipitation and maintain the coherence of the physical structure, this method employs a spatiotemporal intensity-aware joint loss function with strong generalization capabilities. Total loss function. Due to intensity-sensitive loss Structural similarity constraints composition:

[0130]

[0131] To address the issue of limited sample sizes but significant impact from extreme rainfall events, a method combining dynamic precipitation weighting masks was designed. The intensity-sensitive loss calculation method possesses both the smoothness of L2 loss near zero and the robustness of L1 loss to outliers (extreme heavy precipitation):

[0132]

[0133] In the formula, This represents the actual rainfall value of the i-th training sample; This represents the predicted rainfall value output by the fusion rainfall prediction enhancement model; The function for calculating the dynamic precipitation weight mask is: [Equation omitted]. This causes the penalty weight in areas of high-intensity precipitation to increase exponentially. , These represent the basic weight gain coefficient and the exponential growth coefficient, respectively.

[0134] Spatiotemporal structural similarity constraints ( ): Introducing a structural similarity index to evaluate the predicted precipitation field at the feature block level. With real scene Covariance consistency in brightness, contrast, and structure preserves the texture edges of storm cells.

[0135] Preferably, S3 further includes: evaluating the fused rainfall prediction enhancement model. Evaluation metrics include mean squared error, mean absolute error, hit rate, false alarm rate, critical success index, neighborhood critical success index, and fractional skill score.

[0136] The present invention also provides a comparative example, which compares the method of the present invention with the traditional optical flow method, a single radar deep learning model, and a pre-trained model without the introduction of a water vapor control module, to verify the improvement effect of BeiDou / GNSS inverted water vapor information and water vapor control module on precipitation prediction performance.

[0137] In actual online operations, the system automatically collects the latest radar sequences and BeiDou / GNSS water vapor fields, inputting them into the end-to-end fused precipitation prediction enhancement model. Through collaborative forward computation between the backbone network and the water vapor control module, the decoder outputs the future... High-resolution precipitation forecast map at each time step This provides high-fidelity data-driven information for subsequent disaster early warning.

[0138] During the evaluation process, group analyses can be performed according to different precipitation intensity thresholds, different forecast lead times, and different precipitation types. For example, the forecast results for the next 30 minutes, 1 hour, 2 hours, and 3 hours can be evaluated separately; forecast indicators can be calculated for light rain, moderate rain, heavy rain, and short-duration heavy precipitation thresholds to verify the applicability of the method of the present invention under different precipitation intensities and different forecast lead times.

[0139] To verify the effectiveness and advancement of the plug-and-play water vapor control module in this method, the method, in a specific embodiment, selected NowcastNet, a leading generative deep learning model in this technical field, as the basic backbone network for adaptation. This backbone network comprises a 9-layer cascaded encoder-decoder structure, with the number of hidden layer feature channels increasing layer by layer from an initial 32 dimensions to 512 dimensions. To achieve isomorphic mapping, the isomorphic feature mapping layer designed in this method also instantiates a 9-layer deep network structure, and a 3×3×3 three-dimensional spatiotemporal convolutional kernel is used in the spatiotemporal water vapor feature extraction stage to expand the receptive field. Experiments were conducted on 100 independent extreme precipitation events using meteorological radar and BeiDou / GNSS measured datasets spanning several years, performing rigorous statistical evaluation. The specific implementation effects are as follows:

[0140] First, at the tracking level of precipitation process modeling, this method completely overcomes the common problem of traditional neural network models losing details in long-term predictions. In the evolution of actual extreme events, ordinary deep learning networks (such as ConvLSTM and PredRNN) will produce a serious "smoothing effect" over time, causing the core area of ​​heavy precipitation to become blurred; while optical flow methods, due to their simple motion translation assumptions, cannot predict the generation and dissipation of convective cells, resulting in huge spatial displacement errors and serious mismatches in precipitation intensity in the middle and late stages. WAVE-NowcastNet, after being adapted to this method module, introduces the physical constraints of BeiDou / GNSS inversion PWV reflecting atmospheric information, which not only maintains clear spatial structure details of precipitation throughout the long prediction period of 2 to 3 hours, but also accurately predicts the evolution and intensification of the heavy precipitation core.

[0141] Secondly, performance was improved in terms of the accuracy of quantitative prediction based on indicators. This technology evaluation introduced the neighborhood-based critical success index (... Strict mathematical indicators such as Skill Score (FSS) and Fractional Skill Score (FSS) are used. The calculation formula is defined as follows:

[0142] ;

[0143] ;

[0144] In the formula, This refers to the number of hits. False Alarms (empty alarms) Misses; and These represent the fractional probability densities of prediction and observation within the local sliding window, respectively. Statistical results show that, under the heavy rainfall threshold (32 mm / h), the model constructed in this invention achieves a significant improvement in average prediction performance over a long timeframe (i.e., within the prediction interval of the next 2 to 3 hours).

[0145] like Figure 2 The image shows a visualization comparison of precipitation prediction results between our method and the baseline model at different forecast lead times. The columns represent different forecast lead times, including frames 1, 6, 12, and 18, corresponding to precipitation prediction results at different future times. The rows represent prediction results from different methods, including observed values, our method, NowcastNet, optical flow, ConvLSTM, and PredRNN. Colors in the image represent precipitation intensity, with darker colors indicating greater intensity. Figure 2It can be seen that, compared with baseline models such as optical flow, ConvLSTM, PredRNN and NowcastNet, the method of this invention can better maintain the spatial structure of precipitation bands and the core location of heavy precipitation over a longer forecast lead time, and reduce problems such as excessive smoothing of precipitation areas, spatial displacement and underestimation of intensity, demonstrating the enhancing effect of BeiDou / GNSS water vapor information on radar precipitation prediction models.

[0146] like Figure 3 Line graph showing the quantitative evaluation results of this method and the baseline model under different forecast lead times. Figure 3 The left side shows a comparison of the Critical Success Index (CSI) score as a function of forecast lead time. Figure 3 The right side shows a comparison of the FSS (Final Skill Score) as a function of forecast lead time, with the evaluation threshold being a precipitation intensity greater than or equal to 16 mm / hour. In the figure, the horizontal axis represents forecast lead time, and the vertical axis represents the corresponding score value. Different colored curves represent different prediction methods, including the method described in this invention, NowcastNet, ConvLSTM, optical flow, and PredRNN. Figure 3 It can be seen that as the forecast lead time increases, the scores of each method generally show a downward trend. However, the method of this invention maintains higher CSI and FSS scores in most forecast lead times, indicating that the method of this invention has better performance in terms of heavy precipitation prediction accuracy, spatial structure preservation ability and long-term forecast stability.

[0147] Therefore, this invention employs the aforementioned fusion rainfall prediction enhancement method based on BeiDou GNSS and rain-measuring radar. Based on the implementation of various embodiments of this invention, it is achieved through programmed processing by a system with processor functionality. Therefore, in practical engineering, the technical solutions and functions of various embodiments of this invention are encapsulated into various modules. Based on this reality, and building upon the above embodiments, this invention provides a fusion rainfall prediction enhancement system based on BeiDou GNSS and rain-measuring radar. This system is used to execute one of the fusion rainfall prediction enhancement methods based on BeiDou GNSS and rain-measuring radar described in the above method embodiments.

[0148] See Figure 4 The system includes:

[0149] The input module is used to acquire radar data and precipitable water vapor (PWV) data of the target area, and extract radar observation tensors and BeiDou / GNSS water vapor tensors.

[0150] The water vapor dynamics coding module is used to extract the composite features of the water vapor field based on the BeiDou / GNSS water vapor tensor, and to obtain the water vapor dynamic features by nonlinear coding of the composite features of the water vapor field.

[0151] The fusion precipitation prediction module is used to input radar observation tensors and BeiDou / GNSS water vapor tensors into the trained fusion precipitation prediction enhancement model and output the fusion precipitation prediction results for the target area.

[0152] The training steps for the fusion rainfall prediction enhancement model include:

[0153] Obtain training samples for the fusion rainfall prediction enhancement model;

[0154] A fusion-enhanced rainfall prediction model is constructed. The fusion-enhanced rainfall prediction model includes a basic backbone network and a water vapor control module. The encoder of the basic backbone network processes the radar observation tensor to obtain multi-level radar coding features. The water vapor control module processes the water vapor dynamic features to obtain multi-level water vapor regulation features. The multi-level water vapor regulation features are introduced layer by layer during the decoding of the multi-level radar coding features in the decoder of the basic backbone network, and finally the rainfall prediction result is output.

[0155] A heavy precipitation sample enhancement strategy is introduced to optimize the training samples, and the optimized training samples are used to train the fusion precipitation prediction enhancement model, outputting the trained model.

[0156] It should be noted that the system embodiments provided by the present invention are used not only to implement the methods in the above method embodiments, but also to implement the methods in other method embodiments provided by the present invention. The only difference is that corresponding functional modules are set. The principle is basically the same as that of the above system embodiments provided by the present invention. As long as those skilled in the art can improve the modules in the above system embodiments by referring to the specific technical solutions in other method embodiments and combining technical features to obtain corresponding technical means and technical solutions composed of these technical means, on the basis of the above system embodiments, and on the premise of ensuring the practicality of the technical solutions, they can obtain corresponding system-like embodiments for implementing the methods in other method-like embodiments.

[0157] The method described in this invention is implemented using an electronic device; therefore, it is necessary to introduce the relevant electronic device. For this purpose, embodiments of this invention provide an electronic device, such as... Figure 5 As shown, the electronic device includes: at least one processor, a communication interface, at least one memory, and a communication bus, wherein the at least one processor, the communication interface, and the at least one memory communicate with each other via the communication bus. The at least one processor invokes logical instructions stored in the at least one memory to execute all or part of the steps of the methods provided in the foregoing method embodiments.

[0158] Furthermore, when the logical instructions in at least one of the aforementioned memories are implemented as software functional units and sold or used as independent products, they are stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, is embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (a personal computer, server, or network device) to execute all or part of the steps of the methods described in the various method embodiments of the present invention. The aforementioned storage medium includes: USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks—various media for storing program code.

[0159] The system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, located in one place, or distributed across multiple network units. The purpose of this embodiment is achieved by selecting some or all of the modules according to actual needs. Those skilled in the art will understand and implement this without any inventive effort.

[0160] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0161] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0162] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0163] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0164] In summary, this invention discloses a fusion precipitation prediction enhancement method based on BeiDou GNSS and rain-measuring radar, belonging to the field of precipitation prediction technology. The method includes: acquiring radar data of the target area and obtaining precipitable water vapor quantity (PWV) data from BeiDou / GNSS stations; performing quality control on the data acquired in step S1, constructing historical radar observation tensors, water vapor tensors, and future precipitation prediction target samples; obtaining water vapor dynamic characteristics; constructing a fusion precipitation prediction enhancement model and extracting multi-scale water vapor regulation characteristics; training and optimizing the water vapor control module and alignment injection operator; and evaluating the fusion precipitation prediction enhancement model. This invention employs the aforementioned fusion precipitation prediction enhancement method based on BeiDou GNSS and rain-measuring radar, achieving performance improvement in quantitative prediction accuracy based on indicators, and is capable of accurately predicting the evolution and enhancement of heavy precipitation cores.

[0165] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; 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 or all of the technical features therein; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the technical solutions of the embodiments of the present invention.

Claims

1. A fusion rainfall prediction enhancement method based on BeiDou GNSS and rainfall radar, characterized in that, Includes the following steps: Acquire radar data and precipitable water vapor (PWV) data for the target area, and extract radar observation tensor and BeiDou / GNSS water vapor tensor; Based on the water vapor tensor of Beidou / GNSS, the composite features of the water vapor field are extracted, and the water vapor dynamic features are obtained by nonlinear encoding of the composite features of the water vapor field. The radar observation tensor and the BeiDou / GNSS water vapor tensor are input into the trained fusion precipitation prediction enhancement model, and the fusion precipitation prediction results for the target area are output. The training steps for the fusion rainfall prediction enhancement model include: Obtain training samples for the fusion rainfall prediction enhancement model; A fusion-enhanced rainfall prediction model is constructed. The fusion-enhanced rainfall prediction model includes a basic backbone network and a water vapor control module. The encoder of the basic backbone network processes the radar observation tensor to obtain multi-level radar coding features. The water vapor control module processes the water vapor dynamic features to obtain multi-level water vapor regulation features. The multi-level water vapor regulation features are introduced layer by layer during the decoding of the multi-level radar coding features in the decoder of the basic backbone network, and finally the rainfall prediction result is output. A heavy precipitation sample enhancement strategy is introduced to optimize the training samples, and the optimized training samples are used to train the fusion precipitation prediction enhancement model, outputting the trained model.

2. The fusion rainfall prediction enhancement method based on BeiDou GNSS and rain-measuring radar as described in claim 1, characterized in that, The steps for extracting the radar observation tensor and the BeiDou / GNSS water vapor tensor include: Based on the observation timestamp of radar data, the precipitable water vapor (PWV) data is time-synchronized, and the time-synchronized PWV data is spatially resampled to generate a gridded PWV water vapor field that is spatiotemporally aligned with the radar data. The radar observation tensor is composed of radar data within a preset continuous time step; the gridded PWV water vapor field within the preset continuous time step is selected to form the BeiDou / GNSS water vapor tensor.

3. The fusion rainfall prediction enhancement method based on BeiDou GNSS and rain-measuring radar as described in claim 1, characterized in that, The composite features of the water vapor field are extracted based on the BeiDou / GNSS water vapor tensor, and the water vapor dynamic features are obtained by nonlinear encoding of the composite features of the water vapor field. The steps include: The spatial gradient field and the temporal rate of change field of water vapor are calculated based on the BeiDou / GNSS water vapor tensor, and the three are spliced ​​together in the channel dimension to obtain the composite water vapor feature tensor. The water vapor dynamics characteristics are obtained by nonlinearly encoding the composite water vapor feature tensor using a three-dimensional spatiotemporal convolutional network.

4. The fusion rainfall prediction enhancement method based on BeiDou GNSS and rainfall radar as described in claim 1, characterized in that, The training samples of the fusion precipitation prediction enhancement model include radar observation tensor, BeiDou / GNSS water vapor tensor and precipitation prediction target tensor; wherein, the precipitation prediction target tensor is used to characterize the precipitation prediction target in the future preset number of consecutive time steps after the preset number of consecutive time steps, including precipitation intensity and precipitation grid area ratio.

5. The fusion rainfall prediction enhancement method based on BeiDou GNSS and rain-measuring radar as described in claim 1, characterized in that, The basic backbone network adopts a pre-trained deep learning model for precipitation prediction with an encoder-decoder structure, and the encoder parameters of the basic backbone network are frozen.

6. The fusion rainfall prediction enhancement method based on BeiDou GNSS and rain-measuring radar as described in claim 5, characterized in that, The water vapor control module is configured as a plug-and-play module, including a homogeneous feature mapping layer and a cross-layer regulation output layer. The homogeneous feature mapping layer adopts a coding branch that is homogeneous with the encoder of the basic backbone network, receives water vapor dynamic features, and outputs water vapor regulation features that match different levels of the basic backbone network. The cross-layer regulation output layer uses an alignment injection operator to inject water vapor regulation features from different layers into the corresponding layers of the basic backbone network decoder.

7. The fusion rainfall prediction enhancement method based on BeiDou GNSS and rain-measuring radar as described in claim 6, characterized in that, The alignment injection operator is used to inject water vapor regulation features at different levels into the corresponding levels of the basic backbone network decoder. The steps include: The alignment injection operator converts the water vapor regulation feature of any level into a feature tensor with the same dimension as the decoding feature of the corresponding level of the basic backbone network decoder, and fuses the feature tensor into the decoding feature of the corresponding level to inject water vapor regulation, thereby obtaining the fused decoding feature after water vapor regulation at that level.

8. The fusion rainfall prediction enhancement method based on BeiDou GNSS and rain-measuring radar as described in claim 1, characterized in that, A heavy precipitation sample enhancement strategy is introduced to optimize the heavy precipitation samples. The steps include: Based on the precipitation prediction tensor of each training sample, the maximum precipitation intensity, average precipitation intensity, and the percentage of the area with the strongest precipitation in the precipitation prediction target of each training sample are obtained, and the importance score of each training sample is obtained by weighted fusion. Based on the importance scores of each training sample, heavy precipitation samples are selected from the training samples, and the sampling probability of each training sample is adjusted so that heavy precipitation samples have a higher sampling probability during the training process.

9. A fusion rainfall prediction enhancement system based on BeiDou GNSS and rainfall radar, characterized in that, include: The input module is used to acquire radar data and precipitable water vapor (PWV) data of the target area, and extract radar observation tensors and BeiDou / GNSS water vapor tensors. The water vapor dynamics coding module is used to extract the composite features of the water vapor field based on the BeiDou / GNSS water vapor tensor, and to obtain the water vapor dynamic features by nonlinear coding of the composite features of the water vapor field. The fusion precipitation prediction module is used to input radar observation tensors and BeiDou / GNSS water vapor tensors into the trained fusion precipitation prediction enhancement model and output the fusion precipitation prediction results for the target area. The training steps for the fusion rainfall prediction enhancement model include: Obtain training samples for the fusion rainfall prediction enhancement model; A fusion-enhanced rainfall prediction model is constructed. The fusion-enhanced rainfall prediction model includes a basic backbone network and a water vapor control module. The encoder of the basic backbone network processes the radar observation tensor to obtain multi-level radar coding features. The water vapor control module processes the water vapor dynamic features to obtain multi-level water vapor regulation features. The multi-level water vapor regulation features are introduced layer by layer during the decoding of the multi-level radar coding features in the decoder of the basic backbone network, and finally the rainfall prediction result is output. A heavy precipitation sample enhancement strategy is introduced to optimize the training samples, and the optimized training samples are used to train the fusion precipitation prediction enhancement model, outputting the trained model.

10. An electronic device, characterized in that, The method includes a memory and a processor, the memory storing program instructions that are executed by the processor, the processor invoking the program instructions to perform the method according to any one of claims 1 to 9.

Images