A radar echo extrapolation method and device based on a deep learning algorithm

Abstract

The application provides a radar echo extrapolation method and device based on a deep learning algorithm, relates to the meteorological radar prediction technical field, and aims to improve the prediction accuracy, stability and physical interpretability of future radar echo sequences. The method comprises the following steps: acquiring multi-source historical data; the multi-source historical data comprises radar data, satellite data and numerical model data; differentiating and constructing features of the multi-source historical data to obtain multi-source input features; extrapolating the multi-source input features based on a structured radar echo extrapolation model and a residual evolution radar echo extrapolation model respectively; performing consistency analysis on the extrapolation results of the structured radar echo extrapolation model and the residual evolution radar echo extrapolation model, and generating a future radar echo sequence prediction result based on the consistency analysis result.

Description

Technical Field

[0001] This application relates to the field of weather radar forecasting technology, and in particular to a radar echo extrapolation method and apparatus based on deep learning algorithms. Background Technology

[0002] Radar echo extrapolation is used to predict the location, intensity, and morphological evolution of radar echoes over the next 0 to 2 hours based on historical radar composite reflectivity sequences. It is an important technology for near-term warning of severe convective weather such as hail, short-duration heavy precipitation, and thunderstorms. Existing solutions mainly include traditional optical flow or cross-correlation extrapolation methods, as well as radar image sequence prediction methods based on deep learning models.

[0003] Traditional extrapolation methods typically assume that echoes mainly exhibit translation or local topographic changes in the short term. They extrapolate to future times by estimating the echo motion vector, making them suitable for predicting the movement of mature echoes. However, they are insufficient in characterizing nonlinear evolutions such as emergence, enhancement, weakening, and dissipation. Deep learning methods can learn spatiotemporal features from historical radar images, but most schemes still use the reflectivity of a single radar combination as input or simply stitch together external data as ordinary channels. This makes it difficult to fully learn the physical dependencies between existing echo states, cloud development potential, and environmental support conditions.

[0004] Therefore, there is an urgent need for a radar echo extrapolation method based on deep learning algorithms to improve the prediction accuracy, stability, and physical interpretability of future radar echo sequences. Summary of the Invention

[0005] This application provides a radar echo extrapolation method and apparatus based on deep learning algorithms to improve the prediction accuracy, stability and physical interpretability of future radar echo sequences.

[0006] In a first aspect, this application provides a radar echo extrapolation method based on a deep learning algorithm, the method comprising: Acquire multi-source historical data; the multi-source historical data includes radar data, satellite data, and numerical model data; Differential features are constructed from the multi-source historical data to obtain multi-source input features; The multi-source input features are extrapolated based on the structured radar echo extrapolation model and the residual evolution radar echo extrapolation model, respectively. Consistency analysis is performed on the extrapolation results of the structured radar echo extrapolation model and the residual evolution radar echo extrapolation model, and future radar echo sequence prediction results are generated based on the consistency analysis results.

[0007] Optionally, the processing procedure of the residual evolution radar echo extrapolation model includes: Based on the radar feature extractor, the multi-channel data composed of radar combined reflectivity, echo top height and vertical cumulative liquid water content are downsampled and upsampled to extract multi-scale spatial context features. Based on independent feature encoders, radar history sequences, satellite features, numerical mode features, echo top height sequences, and vertical cumulative liquid water content sequences are encoded to obtain their respective independent encoded features. Based on the feature fusion and regulation module, the multi-scale spatial context features and the independent coded features are fused and regulated to generate fused context features, spatial gating weight map and strong echo constraint weight mask. Based on the residual evolution prediction network, the future radar echo sequence is generated frame by frame iteratively using the fused context features, the spatial gating weight map, and the strong echo constraint weight mask as constraints.

[0008] Optionally, the multi-scale spatial context features and the independent encoded features are fused and regulated based on the feature fusion and regulation module to generate fused context features, a spatial gating weight map, and a strong echo constraint weight mask, including: Based on the feature fusion sub-network, the multi-scale spatial context features and the independent encoded features are concatenated to obtain fused features, and the fused features are then subjected to convolutional dimensionality reduction to obtain fused context features; Based on the joint enhanced gating subnetwork, the spatial gating weight map is generated by taking the spliced ​​result of satellite features, numerical model features, echo top height features and vertical cumulative liquid water content features as input and processing it with convolution and Sigmoid activation. Based on the strong echo constraint subnetwork, a strong echo constraint weight mask is generated using the fused context features and multi-source auxiliary coding features as input.

[0009] Optionally, based on the residual evolution prediction network, and constrained by the fused context features, the spatial gating weight map, and the strong echo constraint weight mask, a future radar echo sequence is generated frame by frame iteratively, including: For each prediction time step, the fused context features, the current prediction frame, and the spatial gating weight map are concatenated and then input into the convolutional network to predict the residual relative to the current prediction frame. The residual is scaled according to a preset exponential decay weight, and the scaled residual is spatially modulated using the spatial gating weight map to obtain the modulated residual. The modulation residual is added to the current prediction frame to obtain the next prediction frame, and the variation amplitude of the strong echo core region in the next prediction frame is constrained by the strong echo constraint weight mask.

[0010] Optionally, the composite loss function of the residual evolution radar echo extrapolation model includes time-weighted mean square error loss, high-value region enhancement loss, multi-source consistency loss, smoothness constraint loss, and residual amplitude constraint loss.

[0011] Optionally, the processing procedure for the structured radar echo extrapolation model includes: Based on the encoder, convolution and pooling operations are performed on the multi-source input features to extract multi-scale spatial features; The decoder performs upsampling and skip connections to restore spatial resolution and outputs future radar echo sequences. If the size of the upsampled feature map is inconsistent with the size of the corresponding connection layer, bilinear interpolation is used to smooth the size matching. The loss function of the structured radar echo extrapolation model is calculated in the following way: Based on the horizontal and vertical Sobel operators, the horizontal and vertical gradients of the predicted results and the true labels are calculated. The magnitude of the synthesized gradient is determined based on the horizontal gradient and the vertical gradient; The loss function of the structured radar echo extrapolation model is determined based on the mean square error (MSE) of each pixel and the magnitude of the synthesized gradient.

[0012] Optionally, the satellite data includes cloud top height, cloud top temperature, cloud cold zone thickness, and cloud optical thickness parameters; Constructing differentiated features from the multi-source historical data includes: Regions that meet the conditions for a thriving cloud region are labeled as the first value, and regions that do not meet the conditions for a thriving cloud region are labeled as the second value, thus obtaining a binary feature mask; the conditions for a thriving cloud region include simultaneously meeting the cloud top height threshold, cloud top temperature threshold, cloud cold zone thickness threshold, and cloud optical thickness threshold. The binary feature mask is mapped onto the radar grid, and the historical radar echo maximum value at the corresponding pixel position is used to form the satellite simulated radar echo feature.

[0013] Optionally, the numerical model data includes total temperature index, atmospheric precipitable water, 700hPa temperature-dew point difference, 0 to 6km vertical wind shear, 700hPa vertical velocity, and -20°C layer height parameter. Constructing differentiated features from the multi-source historical data includes: Regions that meet the environmental potential conditions are labeled as the first value, and regions that do not meet the environmental potential conditions are labeled as the second value, thus obtaining a binary feature mask; the environmental potential conditions include simultaneously meeting the total temperature index threshold, atmospheric precipitable water threshold, 700hPa temperature-dew point difference threshold, 0 to 6km vertical wind shear threshold, 700hPa vertical velocity threshold, and -20℃ layer height threshold. The binary feature mask is mapped onto the radar grid as part of the model-aided input or gated modulation signal.

[0014] Optionally, a consistency analysis is performed on the extrapolation results of the structured radar echo extrapolation model and the residual evolution radar echo extrapolation model, and a prediction result of the future radar echo sequence is generated based on the consistency analysis result, including: Based on at least one of the following factors: data completeness, extrapolation timeliness, echo intensity level, and early warning focus, determine the respective weights of the structured radar echo extrapolation model and the residual evolution radar echo extrapolation model. Based on the weights, the extrapolation results of the structured radar echo extrapolation model and the residual evolution radar echo extrapolation model are weighted and fused to obtain the prediction results of future radar echo sequences.

[0015] Secondly, this application also provides a radar echo extrapolation device based on a deep learning algorithm, the device comprising: The data acquisition module is used to acquire multi-source historical data, including radar data, satellite data, and numerical model data. The feature construction module is used to construct differentiated features from the multi-source historical data to obtain multi-source input features; The model extrapolation module is used to extrapolate the multi-source input features based on the structured radar echo extrapolation model and the residual evolution radar echo extrapolation model, respectively. The prediction result determination module is used to perform consistency analysis on the extrapolation results of the structured radar echo extrapolation model and the residual evolution radar echo extrapolation model, and generate future radar echo sequence prediction results based on the consistency analysis results.

[0016] Thirdly, this application also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the radar echo extrapolation method based on deep learning algorithms as described above.

[0017] Fourthly, this application also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the radar echo extrapolation method based on deep learning algorithms as described above.

[0018] Fifthly, this application also provides a computer program product, including a computer program that, when executed by a processor, implements the radar echo extrapolation method based on deep learning algorithms as described above.

[0019] This application provides a radar echo extrapolation method based on deep learning algorithms. By acquiring and differentially constructing multi-source historical data from radar, satellite, and numerical models, it provides complementary physical information on existing echo states, cloud development potential, and environmental support conditions for echo extrapolation, thus compensating for the shortcomings of pure radar models in predicting the nonlinear evolution of echoes. By employing structured models and residual evolution models for extrapolation, it achieves the complementary advantages of morphological evolution simulation and strong echo intensity maintenance, overcoming the inherent defects of single models in being unable to simultaneously consider morphological reliability and intensity maintenance capabilities. By performing consistency analysis on the extrapolation results of the two models and generating the final prediction results, it realizes mutual verification and dynamic fusion between models, effectively suppressing false alarms while improving prediction accuracy, and significantly improving the reliability and operational applicability of the prediction results. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 A flowchart illustrating a radar echo extrapolation method based on a deep learning algorithm, provided for an embodiment of this application; Figure 2 A schematic diagram of the overall process of a radar echo extrapolation method based on a deep learning algorithm provided in this application embodiment. Figure 1 ; Figure 3 A schematic diagram of a residual evolution radar echo extrapolation model provided for an embodiment of this application; Figure 4 A schematic diagram of the overall process of a radar echo extrapolation method based on a deep learning algorithm provided in this application embodiment. Figure 2 ; Figure 5 A schematic diagram of a radar echo extrapolation device based on a deep learning algorithm provided in an embodiment of this application; Figure 6 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of this application clearer, the application will be further described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0023] All actions involving the acquisition of signal information or data in this application are carried out in accordance with the relevant data protection laws and policies of the country where the application is located, and with the authorization granted by the owner of the relevant device.

[0024] In the embodiments of this application, "multiple" refers to two or more. Terms such as "first" and "second" are used only for descriptive purposes and should not be construed as indicating or implying relative importance or order.

[0025] Figure 1 A flowchart illustrating a radar echo extrapolation method based on a deep learning algorithm, as provided in this application embodiment, is shown below. Figure 1 As shown, the method includes the following steps: Step 110: Obtain multi-source historical data.

[0026] The multi-source historical data includes radar data, satellite data, and numerical model data. After acquiring the multi-source historical data, all types of data undergo quality control, missing data processing, bilinear interpolation, cropping, and normalization to be unified to a uniform latitude and longitude grid, such as a 0.01° uniform latitude and longitude grid. Training samples are matched using the occurrence time of severe convective events as an index. Severe convective types include hail, short-duration heavy precipitation, and thunderstorms with strong winds. Radar data describes the existing echo intensity, morphology, and vertical structure; satellite data describes the cloud development potential, especially cloud top height, cloud top temperature, cloud cold zone thickness, and optical thickness; numerical model data describes the environmental support conditions for convective development, including instability, water vapor, wind shear, and upward motion. These three types of data correspond to the "existing echo state," "cloud development potential," and "environmental support conditions" of the convective system, respectively, and are complementary in physical meaning. Specifically, radar data can be obtained using weather radar mosaic products, satellite data can be obtained using FY-4B geostationary satellite cloud precipitation characteristic parameter products, and numerical model data can be obtained using RMAPS gridded numerical model forecast products.

[0027] Specifically, the radar data includes parameters such as composite reflectivity (CREF), echo top (ET), and vertically integrated liquid water content (VIL). CREF serves as the core input to the model and the true value for future output, while ET and VIL serve as auxiliary inputs to indicate the vertical development height of convection, liquid water content, and strong echo potential. In this embodiment, the temporal resolution of the composite reflectivity is 6 minutes. The model uses 5 frames from the half-hour preceding the extrapolated time as historical input and outputs 20 frames from the next 2 hours.

[0028] Satellite data includes parameters such as cloud top height (CTH), cloud top temperature (CTT), height of supercooled cloud (HSC), and cloud optical thickness (COT). Satellite data has the characteristics of space-based observation and can capture convective cloud development information such as cloud top height, cloud top temperature, supercooled layer thickness, and cloud optical thickness, which is helpful for identifying vigorously developing convective clouds and nascent convective regions.

[0029] Numerical model data includes the Total Totals Index (TT), Precipitable Water (PW), 700 hPa temperature-dew point difference (700T_Td), vertical wind shear from 0 to 6 km (SHR0-6), 700 hPa vertical velocity (w700), and the -20°C layer height (H-20°C). These parameters characterize atmospheric stratification instability, water vapor conditions, wet-dry structure, vertical wind shear, upward motion, and the characteristic layer height associated with hail / strong convection, respectively. To reduce model errors, the most recent forecast data matched to the radar time can be used as input.

[0030] The parameters included in radar data, satellite data, and numerical model data are shown in Table 1.

[0031] Table 1

[0032] Step 120: Construct differentiated features from the multi-source historical data to obtain multi-source input features.

[0033] This application designs different feature construction methods for different data sources. For radar parameters, ET and VIL are normalized and multiplied by learnable weight coefficients, and stacked with the radar echo matrix at the corresponding time to form a radar unit that fuses radar parameters. The weight coefficients are automatically optimized during the training process.

[0034] For satellite parameters, a binary feature mask is constructed using the statistical characteristics of the convective system. Specifically, regions that meet the conditions for a robust cloud region are labeled with a first value, and regions that do not meet the conditions are labeled with a second value, thus obtaining the binary feature mask. The conditions for a robust cloud region include simultaneously meeting cloud top height thresholds, cloud top temperature thresholds, cloud cold zone thickness thresholds, and cloud optical thickness thresholds. The binary feature mask is mapped onto a radar grid, and the satellite simulated radar echo features are formed using the historical radar echo maximum values ​​at the corresponding pixel locations. For example, using the Yellow River Basin convective system feature thresholds CTH>11.69km, CTT<-53.55℃, HSC>7.24km, and COT>34.01 to identify robust cloud regions, regions that meet the conditions are labeled with 1, and the remaining regions are labeled with 0. The binary mask is then mapped onto a radar grid, and the satellite simulated radar echo features are formed using the historical radar echo maximum values ​​at the corresponding pixels. Missing satellite measurements are assigned 0 before processing, and the weighting coefficients of the satellite simulated radar echo are automatically learned as model parameters.

[0035] For numerical model parameters, regions meeting the environmental potential conditions are labeled as first values, and regions not meeting the environmental potential conditions are labeled as second values, resulting in a binary feature mask. The environmental potential conditions include simultaneously satisfying the total temperature index threshold, atmospheric precipitable water threshold, 700 hPa temperature-dew point difference threshold, 0 to 6 km vertical wind shear threshold, 700 hPa vertical velocity threshold, and -20℃ layer height threshold. The binary feature mask is mapped onto the radar grid and used as part of the model auxiliary input or gated modulation signal. For example, using thresholds such as TT>44.13, 700T_Td<2.0℃, PW>21.54mm, SHR0-6>7.8m / s, and w700<0.0Pa / s, regions with unstable atmospheric stratification, good lower-level humidity conditions, abundant water vapor, strong environmental wind shear, and upward motion are identified, and a binary environmental potential mask is generated. This mask is mapped back to the radar grid and used as part of the model auxiliary input or gated modulation signal.

[0036] The final multi-source fusion input may include: 5 frames of combined radar reflectivity, 10 frames of radar parametric data (5 frames each of ET and VIL), 5 frames of satellite data (1 frame each of CTH, CTT, HSC, COT and ΔCOT), and 6 frames of numerical model data (1 frame each of TT, H-20℃, PW, SHR0-6, 700T_Td and w700), for a total of 26 frames of input. The output is 20 frames of combined radar reflectivity for the next 2 hours.

[0037] Step 130: Extrapolate the multi-source input features based on the structured radar echo extrapolation model and the residual evolution radar echo extrapolation model, respectively.

[0038] This application constructs a structured radar echo extrapolation model and a residual evolution radar echo extrapolation model to extrapolate multi-source input features. The structured radar echo extrapolation model improves the ability to characterize echo morphology, boundaries, and evolution through a shape adaptation mechanism and a structured loss function. The residual evolution radar echo extrapolation model improves the maintenance of strong echoes and the stability of long-term extrapolation through multi-source feature fusion, joint enhancement gating, strong echo constraints, and frame-by-frame residual evolution mechanisms.

[0039] In one possible implementation, the processing of the structured radar echo extrapolation model includes: performing convolution and pooling operations on multi-source input features based on the encoder to extract multi-scale spatial features; performing upsampling and skip connections based on the decoder to restore spatial resolution and output the future radar echo sequence. Wherein, if the size of the upsampled feature map is inconsistent with the corresponding connection layer, bilinear interpolation is used for smooth size matching.

[0040] Specifically, the structured radar echo extrapolation model can be a U-Net-based structured radar echo extrapolation model. Based on the standard U-Net, it includes an encoder, a bottleneck layer, a decoder, and skip connections. The encoder extracts multi-scale features through convolution and pooling, while the decoder restores spatial resolution through upsampling and skip connections, outputting future radar echo sequences. The number of basic feature channels can be set to 64, 128, 256, and 512, the convolution kernel size is 3×3, the pooling kernel window size is 2×2, and the activation function is ReLU. To accommodate non-power-of-two input sizes that may occur after region clipping and meshing of radar, satellite, and model data, this model adds a shape adaptation mechanism before concatenating the upsampled features with the encoder skip connection features. If the size of the upsampled feature map is inconsistent with the corresponding connection layer, bilinear interpolation is used for smooth size matching, avoiding information loss and concatenation failure due to boundary padding or size rounding.

[0041] The loss function of the structured radar echo extrapolation model is calculated in the following way: Step a: Based on the horizontal and vertical Sobel operators, calculate the horizontal and vertical gradients of the predicted results and the true labels.

[0042] To overcome the image smoothing and boundary blurring problems caused by traditional MSE loss, the structured radar echo extrapolation model introduces a structured loss function based on the Sobel operator, defining two Sobel operators for horizontal and vertical directions respectively: Horizontal Sobel operator (Sobel_x): [[ 1.0, 0.0, -1.0], [ 2.0, 0.0, -2.0], (1) [ 1.0, 0.0, -1.0]] Vertical Sobel operator (Sobel_y): [[ 1.0, 2.0, 1.0], [ 0.0, 0.0, 0.0], (2) [-1.0, -2.0, -1.0]] The horizontal and vertical gradients of the predicted results are calculated based on the horizontal and vertical Sobel operators and compared with the true labels. / x and / y.

[0043] Step b: Determine the magnitude of the synthesized gradient based on the horizontal and vertical gradients.

[0044] The magnitude of the synthesized gradient is determined based on the horizontal and vertical gradients, and this synthesized gradient magnitude is used as part of the structured loss function. ; (3) Step c: Determine the loss function of the structured radar echo extrapolation model based on the mean square error (MSE) of each pixel and the synthetic gradient magnitude.

[0045] As a structural feature, the gradient is added to the loss function, resulting in the loss function: ; (4) Here, Lpixel is the pixel-based MSE, and wpixel and wstruct are the weight coefficients of the two loss functions, respectively. For example, wpixel = 0.7 and wstruct = 0.3. Through this loss function, the model not only learns that the reflectivity value is close to the true value, but also learns that the echo boundary, morphology, and spatial gradient are close to the true structure.

[0046] The advantage of the structured radar echo extrapolation model lies in its sensitivity to multi-source data. It can learn spatial clues about the evolution of the system from satellite cloud top information, model environmental potential, and radar parameters. When integrating multi-source data, it has a good ability to express echo morphology, splitting and merging, and development and dissipation trends.

[0047] Figure 3This is a schematic diagram of a residual evolution radar echo extrapolation model provided in an embodiment of this application. The residual evolution radar echo extrapolation model can be a U-Net-based residual evolution radar echo extrapolation model, which includes a U-Net radar feature extractor, an independent feature encoder, a feature fusion and control module, a residual evolution prediction network, and a composite loss function.

[0048] Specifically, the processing steps of the residual evolution radar echo extrapolation model include the following steps: Step 210: Based on the radar feature extractor, downsample and upsample the multi-channel data composed of radar combined reflectivity, echo top height and vertical cumulative liquid water content to extract multi-scale spatial context features.

[0049] Specifically, the radar feature extractor takes 15-channel data consisting of 5 frames of combined radar reflectivity, 5 frames of ET, and 5 frames of VIL as input, and extracts multi-scale spatial context features through three-layer downsampling and two-layer upsampling, outputting a 64-channel feature map.

[0050] Step 220: Based on the independent feature encoder, the radar history sequence, satellite features, numerical mode features, echo top height sequence and vertical cumulative liquid water content sequence are encoded respectively to obtain their respective independent encoded features.

[0051] The independent feature encoder encodes radar history sequences, satellite features, mode features, ET sequences, and VIL sequences respectively. The radar history sequence can be encoded into 64 channels, while satellite, mode, ET, and VIL sequences are each encoded into 32 channels.

[0052] It should be noted that this application does not specify the order of steps 210 and 220.

[0053] Step 230: Based on the feature fusion and regulation module, the multi-scale spatial context features and independent coding features are fused and regulated to generate fused context features, spatial gating weight map and strong echo constraint weight mask.

[0054] The feature fusion and control module includes three parallel sub-networks: a feature fusion-based sub-network, a joint enhancement gating-based sub-network, and a strong echo constraint-based sub-network.

[0055] Based on the feature fusion subnetwork, multi-scale spatial context features are concatenated with independently encoded features to obtain fused features, and then the fused features are subjected to convolutional dimensionality reduction to obtain fused context features. Specifically, the feature fusion subnetwork concatenates the 64-channel multi-scale spatial features extracted by the radar feature extractor with the 64-channel radar, 32-channel satellite, 32-channel mode, 32-channel ET, and 32-channel VIL independently encoded features to form a 256-channel fused feature, which is then integrated and dimensionality reduced to 64-channel context features through multi-layer convolution.

[0056] Based on the joint enhancement gating subnetwork, the splicing results of satellite features, numerical model features, echo top height features and vertical cumulative liquid water content features are used as input. Through convolution and Sigmoid activation processing, a spatial gating weight map is generated. The value range of the spatial gating weight map is [0,1], which is used to enhance the convection development potential region indicated by the consistency of multi-source data.

[0057] Based on the strong echo constraint subnetwork, a strong echo constraint weight mask is generated by taking fused context features and multi-source auxiliary coding features as input. This mask is used to stabilize mature strong echo cores and suppress unreasonable mutations.

[0058] Step 240: Based on the residual evolution prediction network, and with the constraints of fused context features, spatial gating weight map and strong echo constraint weight mask, generate future radar echo sequences frame by frame.

[0059] The residual evolution prediction network does not directly output the future sequence all at once, but generates the future radar echo sequence by using a frame-by-frame iterative method of "current frame + residual = next frame". Specifically, it includes the following steps: Step 241: For each prediction time step, the fused context features, the current prediction frame, and the spatial gating weight map are concatenated and input into the convolutional network to predict the residual relative to the current prediction frame.

[0060] Step 242: Scale the residuals according to the preset exponential decay weights, and use the spatial gating weight map to spatially modulate the scaled residuals to obtain the modulated residuals.

[0061] For example, the residuals are scaled with an exponentially decaying weight of 0.15 × (0.92^t) and spatially modulated using gating weights.

[0062] Step 243: Add the modulation residual to the current prediction frame to obtain the next prediction frame, and constrain the change amplitude of the strong echo core region in the next prediction frame through the strong echo constraint weight mask.

[0063] The composite loss function of the residual evolution radar echo extrapolation model includes time-weighted mean square error loss, high-value region enhancement loss, multi-source consistency loss, smoothness constraint loss, and residual amplitude constraint loss.

[0064] The composite loss function includes time-weighted mean squared error loss, high-value region enhancement loss, multi-source consistency loss, smoothness constraint loss, and residual amplitude constraint loss. Time-weighted loss emphasizes near-term prediction accuracy; high-value region enhancement loss increases the learning weight for regions with strong echoes; multi-source consistency loss prompts the model to learn reasonable development trends in potential regions indicated by multi-source data; smoothness loss constrains non-physical abrupt changes between adjacent frames; and residual amplitude constraint loss prevents overall drift in the prediction sequence. The total loss function is the sum of all weighted loss terms, and its mathematical expression is as follows: (5) in, , , , and These represent the time-weighted mean squared error loss, high-value region enhancement loss, multi-source consistency loss, smoothness constraint loss, and residual magnitude constraint loss, respectively, with corresponding weighting coefficients set to... =0.5, =0.1, =0.1, =1.0. The specific design and physical meaning of each loss term are as follows: Time-weighted mean square error loss This term forms the basis of the loss function and measures the overall deviation between the predicted and actual sequences. To reflect the higher reliability of short-term forecasts compared to long-term operational prior knowledge in weather forecasting, this model introduces a linearly decaying time weighting coefficient. , where t To predict the time step index, this design assigns higher optimization weights to recent prediction frames, guiding the model to prioritize short-term prediction accuracy while allowing for a relatively large error tolerance in long-term predictions. Its mathematical form is: ; (6) in, and Let represent the t-th predicted frame and the actual frame, respectively.

[0065] High-value area enhancement loss Strong convective echoes typically correspond to heavy precipitation, and accurate prediction of strong convective echoes is crucial for disaster prevention and mitigation. However, traditional mean square error loss is not sensitive enough to such strong signal areas that constitute a relatively small proportion. Therefore, this loss function specifically enhances the prediction capability for high-value echo areas with reflectivity factors exceeding a certain threshold. It combines mean square error and mean absolute error to more strictly penalize prediction biases in high-value areas. ; (7) in, The set of pixels representing the high-value region. Its quantity.

[0066] Multi-source consistency loss To effectively integrate the physical information provided by satellite, model, ET, and VIL data into the model training process, this loss introduces a physical consistency constraint. It first identifies regions with strong convection development potential based on ET and VIL features. Then, if the actual radar echo does show significant development in these regions (i.e., the difference from the current frame exceeds a threshold), the model's prediction is encouraged to show a corresponding development trend. This loss makes the model's learning not only data-driven but also guided by physical mechanisms, enhancing the physical interpretability of the prediction results.

[0067] ; (8) in, This indicates an area identified as a potential region by satellite / model and where the actual echo shows significant development.

[0068] Smoothness constraint loss To ensure the naturalness and continuity of the predicted sequence's evolution over time and to avoid non-physical abrupt changes between adjacent frames, this model introduces a smoothness constraint. This loss minimizes the difference between the predicted and actual sequences in the frame-to-frame variation, i.e., the difference in the first-order difference, ensuring that the model's predicted convective system movement and development speed matches actual observations.

[0069] ; (9) Residual amplitude constraint loss This loss term directly serves the core mechanism of residual evolution in the model. It constrains the average change of the predicted sequence relative to the starting point of evolution, encourages the model to predict small and reasonable residual changes, conforms to the basic assumptions of residual evolution, and prevents unreasonable overall shifts in the predicted sequence.

[0070] ; (10) in, The radar frame that marks the starting point of the evolution.

[0071] In summary, this composite loss function, through multi-objective collaborative optimization, aims to improve overall prediction accuracy. At the same time, the ability to issue early warnings for severe weather was given priority. Introducing physical constraints And ensured the smoothness of the prediction results over time. Stability of evolutionary mechanisms This fully supports the realization of the design objectives of the residual evolution radar echo extrapolation model.

[0072] The advantage of this model lies in its ability to maintain strong echo intensity and structural integrity, especially in strong echoes above 30 dBZ and long extrapolation aging of more than 1 hour, where it has a high hit rate.

[0073] Step 140: Perform consistency analysis on the extrapolation results of the structured radar echo extrapolation model and the residual evolution radar echo extrapolation model, and generate future radar echo sequence prediction results based on the consistency analysis results.

[0074] The structured radar echo extrapolation model and the residual evolution radar echo extrapolation model are not simple substitutes, but rather complementary in operational extrapolation. The structured radar echo extrapolation model is more suitable for integrating multi-source data to grasp system morphology, splitting and merging trends, and dissipation trends, reducing over-response to single environmental signals. The residual evolution radar echo extrapolation model is better suited for forecasting strong echo persistence, longer extrapolation lead times, and higher reflectivity thresholds. In practical applications, the weights of the structured radar echo extrapolation model and the residual evolution radar echo extrapolation model can be determined based on at least one of the following: data completeness, extrapolation lead time, echo intensity level, and the target of early warning. The extrapolation results from the structured radar echo extrapolation model and the residual evolution radar echo extrapolation model are then weighted and fused to obtain the predicted future radar echo sequence.

[0075] The complementarity between the structured radar echo extrapolation model U-Net_struct and the residual evolution radar echo extrapolation model U-Net_RENet is reflected in the following aspects: (1) Complementary morphology and intensity: U-Net_struct focuses on spatial morphology, boundary structure and system evolution trend, while U-Net_RENet focuses on maintaining strong echo intensity and long-term continuous evolution.

[0076] (2) Complementary false alarms and false alarms: U-Net_struct can improve the judgment of morphology and development dissipation after fusing multi-source data, but the strong echo may be low; U-Net_RENet can improve the hit rate of strong echo, but false alarms need to be controlled through multi-source consistency, strong echo constraints and model consistency checks.

[0077] (3) Complementary data sensitivity: U-Net_struct is more sensitive to the addition of data such as satellite, mode and radar parameters, and is suitable for use when multi-source data is complete; U-Net_RENet can still maintain good strong echo details and intensity extrapolation ability when only radar echoes are available.

[0078] (4) Complementary weather systems: For organized multi-convective single cells, satellite data significantly improves the morphology of U-Net_struct; for independent mesoscale convective systems, radar parameters and numerical model background fields can provide more valuable intensity and environmental constraints for U-Net_RENet and U-Net_struct.

[0079] Due to the above characteristics, refer to Figure 2 The business integration and decision-making strategy steps can be configured with the following strategies when outputting business data: (1) When the multi-source data is complete and the focus is on the system morphology, splitting, merging, weakening or emerging trends, increase the output weight of the structured radar echo extrapolation model and use its structured loss and data sensitivity to obtain a more reliable morphological evolution.

[0080] (2) When focusing on radar echoes with intensity exceeding the set threshold, such as focusing on strong echoes above 30 dBZ, or when the extrapolation time exceeds the preset time, such as the extrapolation time exceeding 1 hour, or when only radar data is available, increase the output weight of the residual evolution extrapolation model and use its residual evolution mechanism to maintain the core of strong echoes.

[0081] (3) When both the structured radar echo extrapolation model and the residual evolution radar echo extrapolation model predict the presence of strong echoes in the same region, and are consistent with the satellite potential mask and the model potential mask, the early warning confidence of the region is increased; when the residual evolution radar echo extrapolation model predicts the presence of strong echoes, but the structured radar echo extrapolation model does not predict the presence of strong echoes and the satellite potential mask or the model potential mask does not support them, the early warning confidence of the region is reduced or false alarm suppression is performed.

[0082] (4) When the structured radar echo extrapolation model shows a trend of system dissipation, splitting or emergence and is supported by satellite potential mask or mode potential mask, the prediction results of the residual evolution radar echo extrapolation model are morphologically corrected.

[0083] After business integration and decision-making, the system performs multi-dimensional confidence scoring on the integration results or each candidate region, and classifies the warning level accordingly.

[0084] Specifically, refer to Figure 2 The confidence assessment and early warning levels include a confidence score: location score S. pos Intensity rating S int Morphological score S shape and multi-source consistency score S multi Among them, the location score S pos The intensity score S can be calculated based on the degree of overlap between the predicted echo and the location of the multi-source potential mask and historical data, reflecting the reliability of the echo location prediction; intIt can be calculated based on the consistency between predicted echo intensity and measured or multi-source indications, reflecting the reliability of strong echo core maintenance; morphological score S shape The consistency score can be calculated based on the degree of agreement between the predicted echo morphology and structure, such as system boundaries, splitting / merging / new generation trends, and the physical trends supported by the multi-source mask. This consistency can be measured using metrics such as Intersection over Union (IoU) or Structural Similarity SSIM. multi It can be calculated based on the support of the prediction results and the satellite potential mask and the numerical model environmental potential mask, reflecting the consistency between the forecast and multi-source physical information.

[0085] Based on the above sub-item scores, calculate the comprehensive score S: Among them, w1, w2, w3, and w4 are the weight coefficients of each scoring item, which can be configured according to business needs.

[0086] Furthermore, the warning level is divided into low, medium, relatively high, and high levels based on the numerical range of the comprehensive score S. The level can also be raised separately for areas with strong echoes. This warning level can be used to guide subsequent product output and release strategies.

[0087] refer to Figure 4 Based on the aforementioned future radar echo sequence prediction results, confidence scores, and warning levels, the system generates and outputs various business products, including but not limited to: Future radar combined reflectivity: Outputs a 20-frame radar combined reflectivity prediction sequence for the next 0-2 hours, and generates a product sequence diagram to intuitively show the generation, dissipation, movement and evolution trends of the echo system.

[0088] Early Warning Confidence Product: Outputs a confidence distribution map corresponding to the prediction results of future radar echo sequences, displaying the credibility of forecasts for different regions in the form of a spatial field.

[0089] Threshold warning product: Generates corresponding threshold warning maps for reflectivity factor thresholds such as 10 / 20 / 30 / 40 / 50 dBZ, and identifies areas that reach each threshold standard.

[0090] Key Area Warnings: For high-risk areas of severe weather such as hail, short-term heavy rainfall, thunderstorms and strong winds, warning maps or tables are generated, marking the strong echo center, potential disaster type and confidence level.

[0091] refer to Figure 4The business applications and releases generated through this system can be visualized, alerted, and integrated through multiple channels, achieving a feedback loop: Visualization includes displaying forecast sequence diagrams, confidence distribution diagrams, and early warning diagrams / tables in the form of animations and layer overlays on business terminals or web interfaces; Early warning prompts include triggering alarm mechanisms such as sound, pop-up windows, SMS, or APP push when the comprehensive score or strong echo threshold reaches a preset standard; FTP / Web / API interfaces include automatically pushing products to external systems through standard file transfer protocols (FTP), web services, or application programming interfaces (APIs); Integration with early warning platforms does not include data integration with existing early warning release platforms and short-term forecast business systems of meteorological departments to achieve automated data entry and product sharing; Feedback and iterative optimization include collecting forecast effect evaluations, missed / false reporting cases, and user feedback in business applications for subsequent model updates, threshold optimization, and effect evaluation, continuously improving the accuracy and practicality of the complementary business output of the dual models.

[0092] This application transforms the prediction results of deep learning models into quantitative products and decision-making basis that can directly support the monitoring and early warning of severe convective weather, taking into account both algorithm accuracy and operational interpretability.

[0093] Compared to extrapolation models based solely on combined radar reflectivity, this application provides more complete information on strong convection conditions by fusing radar, satellite, and numerical model data. Radar data describes the current echo location and fine structure, satellite data describes cloud development and initial convection, and numerical model data describes the large-scale background field and dynamic-thermal conditions. This multi-source data collectively compensates for the limitations of pure radar models in predicting echo generation, dissipation, and intensity changes. Experimental results show that the addition of multi-source data improves the overall performance of both improved models. For the residual evolution radar echo extrapolation model, multi-source data significantly increases the POD by approximately 0.1 at various time points, reduces the FAR by approximately 0.1, and improves the TS score by approximately 0.1. For the structured radar echo extrapolation model, except for the 10 dBZ threshold, the POD at all thresholds is significantly improved, and TS is improved at the 20, 30, 40, and 50 dBZ thresholds.

[0094] From the perspective of the complementary effects of the two models, the structured radar echo extrapolation model has advantages in simulating the evolution and development of the system. After integrating multi-source data, it can better reproduce the echo morphology, system splitting and dissipation process, but the intensity of strong echoes may be too low. The residual evolution radar echo extrapolation model has obvious advantages in maintaining the integrity of the echo structure and controlling the attenuation of strong echo intensity, especially in longer extrapolation times and stronger echo regions. However, attention should be paid to the control of false echoes. This application, through the complementary application of the two models, can balance the reliability of the morphology and the early warning capability of strong echoes in operation.

[0095] From the perspective of data contribution, the improvement of a single data source is related to the type of weather system. For organized multi-convective cells, satellite data can significantly improve model performance; for independent mesoscale convective systems, radar parameters and numerical model data can bring more improvements. When multiple data sources are added simultaneously, the model output is not a simple summation of single data results, but can automatically learn the physical dependencies between multiple data sources, improving model performance in both index scoring and visualization case evaluation.

[0096] Therefore, this application can improve the accuracy, stability and interpretability of radar echo extrapolation in severe convective nowcasting, and is especially suitable for 0 to 2-hour monitoring and early warning services for severe weather such as hail, short-duration heavy precipitation, thunderstorms and strong winds.

[0097] Figure 5 A schematic diagram of a radar echo extrapolation device based on a deep learning algorithm provided in this application embodiment is shown below. Figure 5 As shown, the radar echo extrapolation device 500 based on deep learning algorithms includes: The data acquisition module 510 is used to acquire multi-source historical data, including radar data, satellite data, and numerical model data. The feature construction module 520 is used to construct differential features from the multi-source historical data to obtain multi-source input features; The model extrapolation module 530 is used to extrapolate the multi-source input features based on the structured radar echo extrapolation model and the residual evolution radar echo extrapolation model, respectively. The prediction result determination module 540 is used to perform consistency analysis on the extrapolation results of the structured radar echo extrapolation model and the residual evolution radar echo extrapolation model, and generate future radar echo sequence prediction results based on the consistency analysis results.

[0098] It should be noted that the radar echo extrapolation device based on deep learning algorithm provided in this application embodiment can realize all the method steps implemented in the above-mentioned radar echo extrapolation method embodiment based on deep learning algorithm, and can achieve the same technical effect. Here, the parts that are the same as those in the method embodiment and the beneficial effects will not be described in detail.

[0099] Figure 6 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 6As shown, the electronic device may include a processor 610, a communications interface 620, a memory 630, and a communication bus 640. The processor 610, communications interface 620, and memory 630 communicate with each other via the communication bus 640. The processor 610 can call logical instructions from the memory 630 to execute the aforementioned radar echo extrapolation method based on deep learning algorithms.

[0100] Furthermore, the logical instructions in the aforementioned memory 630 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be 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 (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0101] On the other hand, this application also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer is able to execute the above-mentioned radar echo extrapolation method based on deep learning algorithms.

[0102] In another aspect, this application also provides a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the aforementioned radar echo extrapolation method based on deep learning algorithms.

[0103] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0104] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence 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 ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0105] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A radar echo extrapolation method based on deep learning algorithms, characterized in that, The method includes: Acquire multi-source historical data; the multi-source historical data includes radar data, satellite data, and numerical model data; Differential features are constructed from the multi-source historical data to obtain multi-source input features; The multi-source input features are extrapolated based on the structured radar echo extrapolation model and the residual evolution radar echo extrapolation model, respectively. Consistency analysis is performed on the extrapolation results of the structured radar echo extrapolation model and the residual evolution radar echo extrapolation model, and future radar echo sequence prediction results are generated based on the consistency analysis results.

2. The method according to claim 1, characterized in that, The processing steps of the residual evolution radar echo extrapolation model include: Based on the radar feature extractor, the multi-channel data composed of radar combined reflectivity, echo top height and vertical cumulative liquid water content are downsampled and upsampled to extract multi-scale spatial context features. Based on independent feature encoders, radar history sequences, satellite features, numerical mode features, echo top height sequences, and vertical cumulative liquid water content sequences are encoded to obtain their respective independent encoded features. Based on the feature fusion and regulation module, the multi-scale spatial context features and the independent coded features are fused and regulated to generate fused context features, spatial gating weight map and strong echo constraint weight mask. Based on the residual evolution prediction network, the future radar echo sequence is generated frame by frame iteratively using the fused context features, the spatial gating weight map, and the strong echo constraint weight mask as constraints.

3. The method according to claim 2, characterized in that, The feature fusion and modulation module fuses and modulates the multi-scale spatial context features and the independent encoded features to generate fused context features, a spatial gating weight map, and a strong echo constraint weight mask, including: Based on the feature fusion sub-network, the multi-scale spatial context features and the independent encoded features are concatenated to obtain fused features, and the fused features are then subjected to convolutional dimensionality reduction to obtain fused context features; Based on the joint enhanced gating subnetwork, the spatial gating weight map is generated by taking the spliced ​​result of satellite features, numerical model features, echo top height features and vertical cumulative liquid water content features as input and processing it with convolution and Sigmoid activation. Based on the strong echo constraint subnetwork, a strong echo constraint weight mask is generated using the fused context features and multi-source auxiliary coding features as input.

4. The method according to claim 2, characterized in that, Based on the residual evolution prediction network, and constrained by the fused context features, the spatial gating weight map, and the strong echo constraint weight mask, future radar echo sequences are generated iteratively frame by frame, including: For each prediction time step, the fused context features, the current prediction frame, and the spatial gating weight map are concatenated and then input into the convolutional network to predict the residual relative to the current prediction frame. The residual is scaled according to a preset exponential decay weight, and the scaled residual is spatially modulated using the spatial gating weight map to obtain the modulated residual. The modulation residual is added to the current prediction frame to obtain the next prediction frame, and the variation amplitude of the strong echo core region in the next prediction frame is constrained by the strong echo constraint weight mask.

5. The method according to claim 2, characterized in that, The composite loss function of the residual evolution radar echo extrapolation model includes time-weighted mean square error loss, high-value region enhancement loss, multi-source consistency loss, smoothness constraint loss, and residual amplitude constraint loss.

6. The method according to claim 1, characterized in that, The processing steps of the structured radar echo extrapolation model include: Based on the encoder, convolution and pooling operations are performed on the multi-source input features to extract multi-scale spatial features; The decoder performs upsampling and skip connections to restore spatial resolution and outputs future radar echo sequences. If the size of the upsampled feature map is inconsistent with the size of the corresponding connection layer, bilinear interpolation is used to smooth the size matching. The loss function of the structured radar echo extrapolation model is calculated in the following way: Based on the horizontal and vertical Sobel operators, the horizontal and vertical gradients of the predicted results and the true labels are calculated. The magnitude of the synthesized gradient is determined based on the horizontal gradient and the vertical gradient; The loss function of the structured radar echo extrapolation model is determined based on the mean square error of the pixels and the magnitude of the synthesized gradient.

7. The method according to claim 1, characterized in that, The satellite data includes cloud top height, cloud top temperature, cloud cold zone thickness, and cloud optical thickness parameters; Constructing differentiated features from the multi-source historical data includes: Regions that meet the conditions for a thriving cloud region are labeled as the first value, and regions that do not meet the conditions for a thriving cloud region are labeled as the second value, thus obtaining a binary feature mask; the conditions for a thriving cloud region include simultaneously meeting the cloud top height threshold, cloud top temperature threshold, cloud cold zone thickness threshold, and cloud optical thickness threshold. The binary feature mask is mapped onto the radar grid, and the historical radar echo maximum value at the corresponding pixel position is used to form the satellite simulated radar echo feature.

8. The method according to claim 1, characterized in that, The numerical model data includes total temperature index, atmospheric precipitable water, 700 hPa temperature-dew point difference, 0 to 6 km vertical wind shear, 700 hPa vertical velocity, and -20°C layer height parameters. Constructing differentiated features from the multi-source historical data includes: Regions that meet the environmental potential conditions are labeled as the first value, and regions that do not meet the environmental potential conditions are labeled as the second value, thus obtaining a binary feature mask; The environmental potential conditions include simultaneously satisfying the total temperature index threshold, the atmospheric precipitable water threshold, the 700hPa temperature-dew point difference threshold, the 0 to 6km vertical wind shear threshold, the 700hPa vertical velocity threshold, and the -20℃ layer height threshold. The binary feature mask is mapped onto the radar grid as part of the model-aided input or gated modulation signal.

9. The method according to claim 1, characterized in that, A consistency analysis is performed on the extrapolation results of the structured radar echo extrapolation model and the residual evolution radar echo extrapolation model. Based on the consistency analysis results, prediction results of future radar echo sequences are generated, including: Based on at least one of the following factors: data completeness, extrapolation timeliness, echo intensity level, and early warning focus, determine the respective weights of the structured radar echo extrapolation model and the residual evolution radar echo extrapolation model. Based on the weights, the extrapolation results of the structured radar echo extrapolation model and the residual evolution radar echo extrapolation model are weighted and fused to obtain the prediction results of future radar echo sequences.

10. A radar echo extrapolation device based on a deep learning algorithm, characterized in that, The device includes: The data acquisition module is used to acquire multi-source historical data, including radar data, satellite data, and numerical model data. The feature construction module is used to construct differentiated features from the multi-source historical data to obtain multi-source input features; The model extrapolation module is used to extrapolate the multi-source input features based on the structured radar echo extrapolation model and the residual evolution radar echo extrapolation model, respectively. The prediction result determination module is used to perform consistency analysis on the extrapolation results of the structured radar echo extrapolation model and the residual evolution radar echo extrapolation model, and generate future radar echo sequence prediction results based on the consistency analysis results.

Images