A method, device and storage medium for graded forecasting of short-term heavy precipitation

Abstract

This invention discloses a method, device, and storage medium for short-term heavy precipitation grading forecasting, belonging to the field of meteorological forecasting technology. The method includes: inputting preprocessed historical multi-source observation data into a short-term heavy precipitation grading forecasting model; performing multi-scale feature extraction and fusion, temporal coding, cross-modal fusion, multi-scale coding, GRU decoding, and terrain enhancement processing to output a multi-level precipitation prediction probability map; using an ordered classification loss function, calculating the loss based on the multi-level precipitation prediction probability map and the corresponding actual precipitation level label, and updating the parameters of the short-term heavy precipitation grading forecasting model through backpropagation until the short-term heavy precipitation grading forecasting model converges, obtaining a trained and optimized short-term heavy precipitation grading forecasting model, thereby generating short-term heavy precipitation grading forecast results. This invention can more comprehensively capture the spatiotemporal evolution of precipitation systems and effectively alleviate the problems of intensity attenuation and precipitation area shift in traditional extrapolation methods.

Description

Technical Field

[0001] This invention relates to a method, device, and storage medium for short-term heavy precipitation grading forecasting, belonging to the field of meteorological forecasting technology. Background Technology

[0002] Short-duration heavy rainfall is a severe convective weather phenomenon in which precipitation reaches or exceeds a certain threshold within a short period of time. It is a key meteorological factor causing natural disasters such as urban flooding and flash floods, and its accurate prediction is crucial for disaster prevention and mitigation. Current quantitative precipitation extrapolation techniques based on optical flow or convolutional recurrent neural networks suffer from two major bottlenecks: systematic attenuation of intensity and offset of precipitation location, which particularly restricts the timeliness of disaster prevention in complex terrain areas. Summary of the Invention

[0003] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method, device and storage medium for short-term heavy precipitation classification forecasting, which can more comprehensively capture the spatiotemporal evolution of precipitation systems and effectively alleviate the problems of intensity attenuation and precipitation area shift in traditional extrapolation methods.

[0004] To achieve the above objectives, the present invention is implemented using the following technical solution:

[0005] In a first aspect, the present invention provides a method for graded forecasting of short-term heavy precipitation, comprising:

[0006] The current multi-source observation data of the area to be predicted are input into the pre-trained and optimized short-term heavy precipitation classification forecast model to generate short-term heavy precipitation level forecast results.

[0007] The training and optimization methods for the short-term heavy precipitation grading forecast model include:

[0008] The preprocessed historical multi-source observation data is input into a pre-constructed short-term heavy precipitation classification forecast model;

[0009] The short-term heavy precipitation grading forecast model performs multi-scale feature extraction and fusion, temporal coding, cross-modal fusion, multi-scale coding, GRU decoding and terrain enhancement processing to output a multi-level precipitation prediction probability map.

[0010] An ordered classification loss function is used to calculate the loss based on the multi-level precipitation prediction probability map and the corresponding actual precipitation level label. The parameters of the short-term heavy precipitation classification forecast model are then updated through backpropagation until the short-term heavy precipitation classification forecast model converges, resulting in a trained and optimized short-term heavy precipitation classification forecast model.

[0011] Furthermore, the current multi-source observation data of the area to be predicted is input into a pre-trained and optimized short-term heavy precipitation grading forecast model to generate short-term heavy precipitation grading forecast results, including:

[0012] The system acquires and preprocesses multi-source observation data at the current moment to obtain spatially aligned real-time radar echo time-series data, real-time ground precipitation observation time-series data, standardized terrain feature data, and encoded real-time seasonal feature vectors.

[0013] The real-time radar echo time-series data, real-time ground precipitation observation time-series data, standardized terrain feature data, and encoded real-time seasonal feature vectors are input into the pre-trained and optimized short-term heavy precipitation classification forecast model.

[0014] Obtain the multi-level precipitation prediction probability map for future periods output by the short-term heavy precipitation classification forecast model, and generate the final short-term heavy precipitation level forecast result based on the preset threshold.

[0015] Furthermore, the preprocessing method for the historical multi-source observation data includes:

[0016] Receive historical radar echo sequences, historical surface precipitation observation sequences, static topographic data, and seasonal feature vectors;

[0017] The historical radar echo sequences and historical surface precipitation observation sequences are spatially aligned and resampled to be unified to the same grid size and spatial resolution, resulting in spatially aligned radar echo time series data and surface precipitation observation time series data.

[0018] The static terrain data is standardized, and the slope aspect, slope, and roughness features are extracted to obtain standardized terrain feature data.

[0019] The seasonal feature vector is encoded into a numerical vector to obtain the encoded seasonal feature vector.

[0020] Furthermore, the short-term heavy precipitation grading forecast model performs multi-scale feature extraction and fusion, temporal coding, cross-modal fusion, multi-scale coding, GRU decoding, and terrain enhancement processing to output a multi-level precipitation prediction probability map, including:

[0021] Multi-scale feature extraction is performed on the standardized terrain feature data, and the data is fused with the encoded seasonal feature vector to generate a static feature tensor.

[0022] Time-series coding was performed on spatially aligned radar echo time-series data and ground precipitation observation time-series data to obtain radar stream-coded hidden state sequences and precipitation stream-coded hidden state sequences, respectively.

[0023] The radar stream-coded hidden state sequence, the precipitation stream-coded hidden state sequence, and the static feature tensor are fused across modes to obtain the fused spatiotemporal feature representation.

[0024] The fused spatiotemporal feature representation is encoded at multiple scales to extract hierarchical feature representations at different scales;

[0025] Based on the hierarchical feature representation, GRU decoding is performed, and time-adaptive adjustment is performed in combination with the encoded seasonal feature vector to obtain a preliminary precipitation prediction feature sequence.

[0026] The standardized terrain feature data is used to perform terrain enhancement processing on the preliminary precipitation prediction feature sequence to generate a multi-level precipitation prediction probability map.

[0027] Furthermore, multi-scale feature extraction is performed on the standardized terrain feature data, and it is fused with the encoded seasonal feature vector to generate a static feature tensor, including:

[0028] Multi-scale terrain features are extracted from standardized terrain feature data using a multi-scale convolutional network.

[0029] The terrain features are fused using adaptive pooling and downsampled to the target size to obtain a terrain feature representation;

[0030] The terrain feature representation is fused with the encoded seasonal feature vector to generate a static feature tensor containing terrain and seasonal information.

[0031] Furthermore, time-series coding is performed on the spatially aligned radar echo time-series data and the surface precipitation observation time-series data, including:

[0032] The radar echo timing data is time-series encoded using a residual GRU encoder to obtain a radar stream-coded hidden state sequence.

[0033] The same residual GRU encoder is used to perform time-series encoding on the ground precipitation observation time-series data to obtain the precipitation flow encoded hidden state sequence.

[0034] Further:

[0035] The cross-modal fusion includes:

[0036] Extract the feature representation of the last time step from the radar stream-coded hidden state sequence and the precipitation stream-coded hidden state sequence;

[0037] The static feature tensor is downsampled to the same spatial size as the feature representation;

[0038] The radar flow characteristics, precipitation flow characteristics, and static characteristics are stitched together along the channel dimension;

[0039] The spliced ​​features are fused by residual convolutional blocks to obtain a fused spatiotemporal feature representation;

[0040] The multi-scale coding includes:

[0041] Repeat the aforementioned temporal coding and cross-modal fusion processing flow, and spatially downsample the features in each coding layer;

[0042] Save the features after each layer is fused as intermediate features for subsequent decoding;

[0043] The output contains a set of hierarchical feature representations with features at different scales.

[0044] Further:

[0045] The GRU decoding includes:

[0046] GRU decoding and upsampling are performed layer by layer, starting from the deepest feature in the hierarchical feature representation.

[0047] In each decoding layer, the upsampled features are combined with the fused features of the corresponding scale of the coding layer, and the seasonal feature vectors are used to perform adaptive fusion through a gating mechanism to obtain a preliminary precipitation prediction feature sequence.

[0048] The terrain enhancement process includes:

[0049] The standardized terrain feature data is subjected to convolution processing to extract terrain gradient features, terrain height features, and terrain curvature features;

[0050] By fusing the terrain gradient features, terrain height features, and terrain curvature features with the preliminary precipitation prediction feature sequence through learnable parameters, terrain physical constraints are achieved, and multi-level precipitation prediction probability maps are generated through output convolutional layers.

[0051] Secondly, the present invention provides a short-term heavy precipitation grading forecasting device, comprising:

[0052] Memory, used to store computer programs / instructions;

[0053] A processor for executing the computer program / instructions to implement the steps of any of the methods described above.

[0054] Thirdly, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of any of the methods described above.

[0055] Compared with the prior art, the beneficial effects achieved by the present invention are as follows:

[0056] 1. This invention provides a method, device, and storage medium for short-term heavy precipitation grading forecasting. By fusing multi-source heterogeneous data such as radar echoes, surface precipitation observations, topography, and seasonal characteristics, the model can more comprehensively capture the spatiotemporal evolution of precipitation systems, effectively alleviating the problems of intensity attenuation and precipitation area shift in traditional extrapolation methods. In tests in Hunan Province, the TS score for 1-3 hour short-term heavy precipitation is improved by more than 80% to 160% compared to existing mainstream large model products, and false alarms and missed alarms are significantly reduced.

[0057] 2. This invention uses a specially designed terrain enhancement module to explicitly introduce features such as terrain gradient, height and curvature, and integrates physical processes such as terrain uplift and obstruction into the model in a learnable way, which significantly improves the prediction accuracy of precipitation area and intensity in complex terrain, especially its ability to depict heavy precipitation in mountainous areas.

[0058] 3. This invention employs a dual-stream residual GRU to encode radar and precipitation sequences separately, and achieves dynamic interaction between temporal and static features through a cross-modal attention mechanism. Combined with a multi-scale encoding-decoding structure, it realizes hierarchical feature extraction and fusion from local details to the global system, enhancing the model's ability to capture and infer the development and movement of precipitation systems.

[0059] 4. This invention injects seasonal features into the decoding process in a periodic encoding manner, enabling the model to dynamically adjust the prediction features according to the month, better adapt to the statistical characteristics and evolution patterns of precipitation systems in different seasons, and improve the stability of forecasts for all periods throughout the year. Attached Figure Description

[0060] Figure 1 This is a flowchart of a short-term heavy precipitation classification forecasting method provided in an embodiment of the present invention;

[0061] Figure 2 This is a network architecture diagram of the short-term heavy precipitation graded forecasting model provided in this embodiment of the invention;

[0062] Figure 3 This is a schematic diagram of the non-graded TS score of the objective forecast product for short-term heavy precipitation in Hunan Province from April to September 2025 provided in an embodiment of the present invention;

[0063] Figure 4 This is a schematic diagram of the non-graded BIAS score of the objective forecast product for short-term heavy precipitation in Hunan Province from April to September 2025 provided in an embodiment of the present invention;

[0064] Figure 5 This is a schematic diagram of the graded ME score of the objective forecast product for short-term heavy precipitation in Hunan Province from April to September 2025 provided in an embodiment of the present invention;

[0065] Figure 6This is a schematic diagram comparing the short-term heavy rainfall forecast issued starting at 20:00 on June 7, 2025 with the actual situation, provided by an embodiment of the present invention.

[0066] Figure 7 This is a schematic diagram comparing the short-term heavy rainfall forecast issued starting at 23:00 on June 7, 2025 with the actual situation, provided by an embodiment of the present invention.

[0067] Figure 8 This is a schematic diagram comparing the short-term heavy rainfall forecast issued starting at 02:00 on June 8, 2025 with the actual situation, provided by an embodiment of the present invention.

[0068] Figure 9 This is a schematic diagram comparing the short-term heavy rainfall forecast issued starting at 05:00 on June 8, 2025 with the actual situation, provided by an embodiment of the present invention. Detailed Implementation

[0069] The present invention will be further described below with reference to the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and should not be used to limit the scope of protection of the present invention.

[0070] Example 1, such as Figure 1 As shown in the figure, this embodiment introduces a method for graded forecasting of short-term heavy precipitation, including:

[0071] The current multi-source observation data of the area to be predicted are input into the pre-trained and optimized short-term heavy precipitation classification forecast model to generate short-term heavy precipitation level forecast results.

[0072] The training and optimization methods for the short-term heavy precipitation grading forecast model include:

[0073] The preprocessed historical multi-source observation data is input into a pre-constructed short-term heavy precipitation classification forecast model;

[0074] The short-term heavy precipitation grading forecast model performs multi-scale feature extraction and fusion, temporal coding, cross-modal fusion, multi-scale coding, GRU decoding, and terrain enhancement processing to output a multi-level precipitation prediction probability map. The model architecture is as follows: Figure 2 As shown;

[0075] An ordered classification loss function is used to calculate the loss based on the multi-level precipitation prediction probability map and the corresponding actual precipitation level label. The parameters of the short-term heavy precipitation classification forecast model are then updated through backpropagation until the short-term heavy precipitation classification forecast model converges, resulting in a trained and optimized short-term heavy precipitation classification forecast model.

[0076] The short-term heavy precipitation grading forecasting method provided in this embodiment involves the following steps in its application:

[0077] Step 1: Data preprocessing and alignment;

[0078] The data preprocessing module receives historical radar echo sequences, historical surface precipitation observation sequences, static topographic data, and seasonal feature vectors as input. The radar echo data has a 12-minute time interval and a spatial resolution of 1 km. The historical radar echo sequence input required by the short-term heavy precipitation classification forecast model consists of 5 frames of echo data from the past hour. The surface precipitation observation data comprises hourly precipitation observations from all precipitation observation stations within the spatial range. The station data is interpolated into 1 km resolution grid data using an inverse distance weighted interpolation method. The historical precipitation observation sequence input required by the short-term heavy precipitation classification forecast model consists of 2 frames of precipitation data from the past 2 hours. The static topographic data includes topographic features such as elevation, aspect, slope, and roughness. The seasonal features are monthly information, ranging from 0 to 11, representing the 12 months of the year. The data preprocessing module spatially aligns historical radar echo sequences and historical surface precipitation observation sequences, resampling data from different sources to the same grid size. It standardizes static topographic data and extracts derived features such as aspect, slope, and roughness. Seasonal features are encoded into numerical vectors, resulting in encoded seasonal feature vectors. After these processes, the data preprocessing module outputs spatially aligned radar echo time-series data, spatially aligned surface precipitation observation time-series data, standardized topographic feature data, and encoded seasonal feature vectors. All output data have a unified spatial resolution of 1 km, serving as input for subsequent short-term heavy precipitation grading forecasting models.

[0079] Step 2: Terrain feature extraction and static feature encoding;

[0080] The static feature encoder receives the terrain feature representation and the encoded seasonal feature vector output from step one as input. The terrain feature representation uses a multi-scale convolutional network to extract terrain features at three scales: high resolution, medium resolution, and low resolution. The multi-scale features are fused and downsampled to the target size through adaptive pooling. The seasonal features are encoded using a periodic encoding method, mapping the monthly information to sine and cosine periodic features. The static feature encoder outputs a static feature tensor that fuses the terrain and seasonal information.

[0081] Step 3: Dual-stream residual GRU coding;

[0082] The radar encoder receives the radar echo time-series data output from step one as input and uses a residual GRU encoder to perform time-series encoding on the sequence. The residual GRU encoder updates the hidden state step-by-step through update gates, reset gates, and output gates, and uses residual connections to directly superimpose the input features onto the output features, maintaining gradient stability and outputting a radar stream-coded hidden state sequence. The precipitation encoder receives the surface precipitation observation time-series data output from step one as input and uses the same residual GRU encoder to perform time-series encoding on the sequence, outputting a precipitation stream-coded hidden state sequence. The two residual GRU encoders process in parallel, extracting the time-series evolution features of the radar echo and surface precipitation respectively.

[0083] Step 4: Cross-modal spatiotemporal feature fusion;

[0084] The feature fusion module receives the radar stream-coded hidden state sequence and precipitation stream-coded hidden state sequence output from step three, as well as the static feature tensor output from step two, as input. The feature fusion module extracts the feature representations of the last time step from both the radar stream-coded hidden state sequence and the precipitation stream-coded hidden state sequence. It then spatially downsamples the static feature tensor to match the feature size of the current coding layer. Next, it concatenates the radar stream features, precipitation stream features, and static features along the channel dimension. Finally, it uses residual convolutional blocks to fuse the concatenated multi-source features. The fusion process is represented as follows:

[0085] ;

[0086] in, This represents the spatiotemporal features after fusion. Indicates radar flow characteristics, Indicates precipitation flow characteristics, [⋅] represents static features, [⋅] represents channel dimension concatenation, and ResConv represents residual convolutional blocks. The feature fusion module achieves cross-modal feature interaction through convolution operations, captures the spatiotemporal correlation between radar echoes and ground precipitation, and outputs a fused spatiotemporal feature representation that simultaneously includes radar echoes, ground precipitation, and topographic seasonal information.

[0087] Step 5: Multi-scale coding and feature extraction;

[0088] The multi-scale encoding module repeats steps three and four, but in each encoding layer, it spatially downsamples the features, extracting feature representations at different scales layer by layer. After each layer of encoding is completed, the multi-scale encoding module saves the fused features as intermediate features for subsequent decoding stages. Through multi-scale encoding, hierarchical feature representations from local details to global structure are extracted step by step. The multi-scale encoding module outputs a set of multi-scale fused feature representations, containing features at different scales from fine-grained to coarse-grained.

[0089] Step Six: GRU Decoding and Time Adaptation;

[0090] The GRU decoder starts with the deepest features and decodes layer by layer. In each decoding layer, the GRU decoder receives the initial state feature sequence output from step six and the fused features of the corresponding encoding layer output from step five as input, and generates feature representations for future time steps through a gating mechanism. The upsampling process receives the features output from the GRU decoder as input and performs bilinear interpolation upsampling on the decoded features to restore spatial resolution. The time-adaptive module receives the seasonal feature vector output from step one and the upsampled features as input, and encodes the seasonal feature vector into a seasonal feature representation.

[0091] The intensity weights of seasonal features are adjusted according to the decoding layer depth, and a gating mechanism is used to generate gating weights. Then, the seasonal features and the decoded features are adaptively fused. The fusion process is represented as follows:

[0092] ;

[0093] ;

[0094] in This represents the decoded features after upsampling. This represents the encoded seasonal characteristics. This represents the seasonal feature intensity weights adjusted based on the decoding layer depth. Indicates the gating weight, Indicates feature splicing, Represents the gate function. This represents the decoded features after seasonal adaptive enhancement, enabling dynamic adjustment of predicted features by seasonal factors. The residual refinement module receives the features after time-adaptive processing. As input, the upsampled features are convolutionally processed, and feature details are preserved through residual connections to obtain a preliminary precipitation prediction feature sequence. ,in This represents the calculation of the residual module.

[0095] Step 7: Terrain Enhancement and Final Prediction Output;

[0096] The terrain enhancement module receives the preliminary precipitation prediction feature sequence output from step six and the standardized terrain feature data output from step one as input. The module first performs convolution processing on the standardized terrain feature data to extract terrain gradient features, terrain height features, and terrain curvature features. The terrain gradient reflects slope changes, terrain height reflects elevation information, and terrain curvature reflects the degree of terrain undulation. The module then fuses the extracted terrain features with the preliminary prediction features using learnable parameters, spatially modulating the precipitation prediction using terrain gradient, terrain height, and terrain curvature to constrain physical processes such as terrain uplift and blocking. The calculation formula for terrain enhancement is:

[0097] ;

[0098] ;

[0099] ;

[0100] ;

[0101] Where Z represents terrain data. , , These represent terrain gradient features, terrain height features, and terrain curvature features, respectively. , , These represent the convolution operations used to extract terrain gradient, height, and curvature, respectively. This is the preliminary precipitation prediction feature sequence output from step six, where α, β, and γ represent learnable terrain modulation parameters. Indicates the output convolutional layer. This represents the final output. The terrain-enhanced features are passed through an output convolutional layer to generate the final precipitation intensity prediction result, outputting a multi-level precipitation prediction probability map for the future time period. During inference, a threshold judgment is applied to the prediction probability map for each precipitation level. Levels with a probability greater than 0.5 are judged as "there is precipitation of this level". If multiple levels are listed as "there is", they are stacked from low to high to obtain the final level judgment result. Thus, the final precipitation level output can be obtained from the multi-level probability map.

[0102] Step 8: Loss Measurement and Parameter Update;

[0103] The loss function receives the multi-level precipitation prediction probability map and corresponding actual precipitation level labels output from step eight as input. Due to the extremely high data imbalance in severe convective event data, the loss calculation module uses the ORDiceLoss loss function of ordered classification to calculate the loss value between the prediction results and the actual observations. The ORDiceLoss loss function calculates the loss independently for each category, considering only the positive event impact of that category while ignoring the impact of a large number of negative events. The loss of each category has an equal contribution, so there is no need to assign corresponding weights to each category for balancing. The formula for calculating the ORDiceLoss loss function is:

[0104] ;

[0105] in, Represents the ORDiceLoss loss function. The total number of levels. This represents the predicted probability value of the i-th pixel at the c-th level. This represents the true label value of the i-th pixel at the c-th level. The gradient of the loss function with respect to the model parameters is calculated using the backpropagation algorithm, and the model parameters are updated accordingly. Steps one through eight are repeated during the training process until the model converges. This method significantly outperforms loss functions such as MAE (Mean Absolute Error), MSE (Mean Squared Error), and Weighted Cross Entropy for training on imbalanced data, effectively suppressing the increase in false alarm rate while improving the hit rate. After training, the final short-term heavy precipitation classification prediction model is obtained.

[0106] The following description, in conjunction with a preferred embodiment, illustrates the content involved in the above embodiments.

[0107] The forecasts for hourly heavy precipitation products (0-3 hours) generated based on FTNet (represented by FTNet in the result images) in Hunan Province were verified using indicators such as ThreatScore (TS), Bias score (BIAS), and graded short-term heavy precipitation score (ME). The precipitation products of the China Meteorological Administration's wind and thunder model (represented by Fenglei in the result images) were mainly used as a reference.

[0108] The TS score and BIAS score for ungraded short-duration heavy rainfall of 20 mm or more are calculated as follows:

[0109] Table 1 shows the evaluation table for short-duration heavy rainfall:

[0110] Table 1. Evaluation Form for Short-Term Heavy Rainfall;

[0111]

[0112] In the formula To forecast the correct number of stations (times), Number of empty station reports (times) This represents the number of stations (times) that were missed in reporting. This represents the number of stations (times) where no heavy precipitation forecast was correct.

[0113] Calculate the TS score for ungraded short-term heavy precipitation forecasts: ;

[0114] Special Circumstances Agreement: When + + =0, at this time TS is recorded as the missing value 999999.

[0115] Calculate the BIAS bias of ungraded short-term heavy precipitation forecasts: ;

[0116] Special Circumstances Agreement:

[0117] Scenario 1 When + =0, + When B = 0, B is denoted as 0.

[0118] Scenario 2 When + =0, + When the value is greater than 0, B is recorded as 999999.

[0119] For short-duration heavy precipitation exceeding 50 mm / h, calculate the scores for graded short-duration heavy precipitation forecasts:

[0120] ;

[0121] in, The score is based on the skill of comparing the short-term heavy precipitation station forecasts with the actual conditions obtained by looking up the data in Table 2. This represents the total number of samples in Table 2, excluding those with a score of Nan. The average score is obtained by summing the scores of each station and dividing by the total number of samples. Special case stipulation: When the total sample size is 0, at this time... Recorded as the missing value 999999.

[0122] Table 2. Graded Short-Term Heavy Precipitation Forecast Score Table (SE);

[0123]

[0124] like Figure 3 As shown, for short-duration heavy precipitation in Hunan Province from April to September 2025, the statistical scoring results show that the TS scores for the first 1-3 hours of precipitation above 20 mm / h in the FTNet precipitation forecast for 2025 are 0.359, 0.213, 0.14, and 0.06, respectively, which are higher than the 0.197, 0.096, and 0.056 for wind and thunder precipitation forecasts (an increase of 82.2%, 122.9%, and 167.6%, respectively).

[0125] like Figure 4 As shown, the BIAS scores of the FTNet precipitation forecast without grading are 0.06, 0.069, and 0.096, respectively, which are lower than those of the large-scale wind and thunder model (0.459, 0.586, and 0.651) (reduced by 86.9% and 85.3%, respectively). The BIAS of the FTNet precipitation forecast is closer to 0 than that of the wind and thunder model, indicating that its forecast and the number of actual observation stations have reached a balance, while the wind and thunder precipitation forecast has obvious false alarms or missed reports.

[0126] like Figure 5 As shown, for heavy precipitation events with wind speeds exceeding 50 mm / h, the FTNet precipitation forecast scores were 0.102, 0.045, and 0.022, respectively, which are higher than the wind and thunder precipitation forecast scores of 0.028, 0.012, and 0.007 (improvements of 264.3%, 275.0%, and 214.3%, respectively). This score result indicates that for heavy precipitation events with wind speeds exceeding 50 mm / h, the FTNet precipitation forecast significantly outperforms the wind and thunder precipitation forecast, demonstrating a superior ability to capture extreme heavy precipitation events.

[0127] like Figures 6-9 As shown, the short-term heavy precipitation forecasts issued at 20:00 and 23:00 on June 7, 2025, and at 02:00 and 05:00 on June 8, 2025, were compared and verified with the actual Ground_Truth data.

[0128] like Figure 6 As shown in AC, the area of ​​heavy rainfall in the next 1-3 hours from 20:00 on the 7th will mainly be distributed in central and northern Hunan Province, with many areas experiencing short-term heavy rainfall of more than 50mm. Among them, Changsha area recorded more than 70mm in the first time period. Figure 6 As shown in df, FTNet precipitation forecasts accurately captured the range and intensity of short-duration heavy rainfall. Specifically, the 3-hour ungraded TS scores were 0.479, 0.309, and 0.224, respectively, while the scores for graded precipitation exceeding 50mm were 0.129, 0.254, and 0.150, higher than those for wind and thunder precipitation forecasts. Figure 6 gi).

[0129] like Figure 7As shown in AC, in the next 1-3 hours from 23:00, the areas with heavy rainfall exceeding 20mm were mainly located in Changsha, Yiyang, Loudi, Huaihua, and Shaoyang, with multiple stations recording rainfall of 50mm or more. The center of the rainfall gradually shifted southward over time. Figure 7 As shown in df, the TS of 1-3 hour lead time for FTNet precipitation forecasts reached 0.471, 0.264 and 0.328 respectively, which are significantly higher than those for wind and thunder precipitation forecasts. The scores of FTNet precipitation forecasts for graded precipitation of 50 mm or more were 0.129, 0.254 and 0.150, which are significantly higher than those for wind and thunder precipitation forecasts.

[0130] From 03:00 to 08:00 on the 8th, areas experiencing heavy rainfall exceeding 20mm were mainly located in Loudi, Huaihua, and Shaoyang, with multiple stations recording rainfall of 100mm or more. The center of the rainfall gradually shifted southward over time, and the intensity of the rainfall weakened around 08:00. Figure 8 As shown in the data, the ungraded TS scores for the FTNet precipitation forecasts issued starting at 02:00 on the 8th reached 0.691, 0.631, and 0.608 respectively, indicating good agreement between the center of the heavy precipitation area and the actual situation. The graded heavy precipitation scores for FTNet precipitation forecasts exceeding 50mm were as high as 0.303, 0.369, and 0.201, demonstrating outstanding ability to characterize the center of precipitation exceeding 80mm. Figure 9 As shown in df, the intensity of the FTNet precipitation forecast issued at 05:00 weakened over time, which is close to the actual situation.

[0131] For this heavy rainfall event, the FTNet precipitation forecast showed a three-hour average TS of 0.436 and an average ME of 0.183, which are excellent in terms of both the area affected and the intensity.

[0132] Example 2: This example provides a short-term heavy precipitation grading forecasting device, comprising:

[0133] Memory, used to store computer programs / instructions;

[0134] A processor for executing the computer program / instructions to implement the steps of any of the methods described in Embodiment 1.

[0135] Example 3: This example provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of any of the methods described in Example 1.

[0136] The terminology is defined as follows:

[0137] Radar echo data: Reflectivity intensity gridded data acquired by meteorological radar, used to characterize the distribution and intensity of precipitation clouds.

[0138] Surface precipitation observation data: Surface automatic weather station or gridded precipitation observation data, used to characterize the actual precipitation intensity.

[0139] Topographic data: Static geographic information data from digital elevation models.

[0140] Seasonal characteristics: Time markers such as months or solar terms are used to characterize the seasonal variation patterns of precipitation systems.

[0141] The residual GRU encoder (GRU stands for Gated Recurrent Unit) is a gated recurrent unit network with residual connections. It updates the hidden state step by step through update gates, reset gates, and output gates, and uses residual connections to maintain gradient stability for extracting temporal features.

[0142] Feature fusion module: a cross-modal attention or gating mechanism used to fuse features from multiple sources.

[0143] Terrain Enhancement Module: A module that uses terrain gradient information to physically constrain precipitation forecasts.

[0144] Time-adaptive module: A module that dynamically adjusts predicted features based on seasonal characteristics.

[0145] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

[0146] Those skilled in the art will understand that embodiments of this disclosure can be provided as methods, systems, or computer program products. Therefore, this disclosure can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this disclosure 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.

[0147] This disclosure is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this disclosure. 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, create a machine for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1A device that provides the functions specified in one or more boxes.

[0148] 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.

[0149] 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.

[0150] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this disclosure and not to limit its protection scope. Although this disclosure has been described in detail with reference to the above embodiments, those skilled in the art should understand that after reading this disclosure, they can still make various changes, modifications or equivalent substitutions to the specific implementation of the invention, but these changes, modifications or equivalent substitutions are all within the protection scope of the pending claims.

Claims

1. A method for graded forecasting of short-term heavy precipitation, characterized in that, include: The current multi-source observation data of the area to be predicted are input into the pre-trained and optimized short-term heavy precipitation classification forecast model to generate short-term heavy precipitation level forecast results. The training and optimization methods for the short-term heavy precipitation grading forecast model include: The preprocessed historical multi-source observation data is input into a pre-constructed short-term heavy precipitation classification forecast model; The short-term heavy precipitation grading forecast model performs multi-scale feature extraction and fusion, temporal coding, cross-modal fusion, multi-scale coding, GRU decoding and terrain enhancement processing to output a multi-level precipitation prediction probability map. An ordered classification loss function is used to calculate the loss based on the multi-level precipitation prediction probability map and the corresponding actual precipitation level label. The parameters of the short-term heavy precipitation classification forecast model are then updated through backpropagation until the short-term heavy precipitation classification forecast model converges, resulting in a trained and optimized short-term heavy precipitation classification forecast model.

2. The short-term heavy precipitation classification forecasting method according to claim 1, characterized in that, The process involves inputting the current multi-source observation data of the area to be predicted into a pre-trained and optimized short-term heavy precipitation classification forecast model to generate short-term heavy precipitation level forecast results, including: The system acquires and preprocesses multi-source observation data at the current moment to obtain spatially aligned real-time radar echo time-series data, real-time ground precipitation observation time-series data, standardized terrain feature data, and encoded real-time seasonal feature vectors. The real-time radar echo time-series data, real-time ground precipitation observation time-series data, standardized terrain feature data, and encoded real-time seasonal feature vectors are input into the pre-trained and optimized short-term heavy precipitation classification forecast model. Obtain the multi-level precipitation prediction probability map for future periods output by the short-term heavy precipitation classification forecast model, and generate the final short-term heavy precipitation level forecast result based on the preset threshold.

3. The method for graded forecasting of short-term heavy precipitation according to claim 1, characterized in that, The preprocessing method for the historical multi-source observation data includes: Receive historical radar echo sequences, historical surface precipitation observation sequences, static topographic data, and seasonal feature vectors; The historical radar echo sequences and historical surface precipitation observation sequences are spatially aligned and resampled to be unified to the same grid size and spatial resolution, resulting in spatially aligned radar echo time series data and surface precipitation observation time series data. The static terrain data is standardized, and the slope aspect, slope, and roughness features are extracted to obtain standardized terrain feature data. The seasonal feature vector is encoded into a numerical vector to obtain the encoded seasonal feature vector.

4. The short-term heavy precipitation classification forecasting method according to claim 3, characterized in that, The process involves performing multi-scale feature extraction and fusion, temporal coding, cross-modal fusion, multi-scale coding, GRU decoding, and terrain enhancement processing through a short-term heavy precipitation grading forecast model, outputting a multi-level precipitation prediction probability map, including: Multi-scale feature extraction is performed on the standardized terrain feature data, and the data is fused with the encoded seasonal feature vector to generate a static feature tensor. Time-series coding was performed on spatially aligned radar echo time-series data and ground precipitation observation time-series data to obtain radar stream-coded hidden state sequences and precipitation stream-coded hidden state sequences, respectively. The radar stream-coded hidden state sequence, the precipitation stream-coded hidden state sequence, and the static feature tensor are fused across modes to obtain the fused spatiotemporal feature representation. The fused spatiotemporal feature representation is encoded at multiple scales to extract hierarchical feature representations at different scales; Based on the hierarchical feature representation, GRU decoding is performed, and time-adaptive adjustment is performed in combination with the encoded seasonal feature vector to obtain a preliminary precipitation prediction feature sequence. The standardized terrain feature data is used to perform terrain enhancement processing on the preliminary precipitation prediction feature sequence to generate a multi-level precipitation prediction probability map.

5. The short-term heavy precipitation classification forecasting method according to claim 4, characterized in that, Multi-scale feature extraction is performed on the standardized terrain feature data, and the data is fused with the encoded seasonal feature vector to generate a static feature tensor, including: Multi-scale terrain features are extracted from standardized terrain feature data using a multi-scale convolutional network. The terrain features are fused using adaptive pooling and downsampled to the target size to obtain a terrain feature representation; The terrain feature representation is fused with the encoded seasonal feature vector to generate a static feature tensor containing terrain and seasonal information.

6. The method for graded forecasting of short-term heavy precipitation according to claim 1, characterized in that, Time-series coding was performed on spatially aligned radar echo time-series data and surface precipitation observation time-series data, including: The radar echo timing data is time-series encoded using a residual GRU encoder to obtain a radar stream-coded hidden state sequence. The same residual GRU encoder is used to perform time-series encoding on the ground precipitation observation time-series data to obtain the precipitation flow encoded hidden state sequence.

7. The method for graded forecasting of short-term heavy precipitation according to claim 1, characterized in that: The cross-modal fusion includes: Extract the feature representation of the last time step from the radar stream-coded hidden state sequence and the precipitation stream-coded hidden state sequence; The static feature tensor is downsampled to the same spatial size as the feature representation; The radar flow characteristics, precipitation flow characteristics, and static characteristics are stitched together along the channel dimension; The spliced ​​features are fused by residual convolutional blocks to obtain a fused spatiotemporal feature representation; The multi-scale coding includes: Repeat the aforementioned temporal coding and cross-modal fusion processing flow, and spatially downsample the features in each coding layer; Save the features after each layer is fused as intermediate features for subsequent decoding; The output contains a set of hierarchical feature representations with features at different scales.

8. The short-term heavy precipitation graded forecasting method according to claim 4, characterized in that: The GRU decoding includes: GRU decoding and upsampling are performed layer by layer, starting from the deepest feature in the hierarchical feature representation. In each decoding layer, the upsampled features are combined with the fused features of the corresponding scale of the coding layer, and the seasonal feature vectors are used to perform adaptive fusion through a gating mechanism to obtain a preliminary precipitation prediction feature sequence. The terrain enhancement process includes: The standardized terrain feature data is subjected to convolution processing to extract terrain gradient features, terrain height features, and terrain curvature features; By fusing the terrain gradient features, terrain height features, and terrain curvature features with the preliminary precipitation prediction feature sequence through learnable parameters, terrain physical constraints are achieved, and multi-level precipitation prediction probability maps are generated through output convolutional layers.

9. A short-term heavy precipitation grading forecasting device, characterized in that, include: Memory, used to store computer programs / instructions; A processor for executing the computer program / instructions to implement the steps of the method according to any one of claims 1-8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When executed by a processor, the computer program implements the steps of the method according to any one of claims 1-8.

Images