Estimation method and device for radar echo quantitative rainfall, computer equipment and storage medium

Abstract

The invention discloses a radar echo quantitative rainfall estimation method comprising the following steps: obtaining historical echo data, and carrying out denoising and normalization processing on the historical echo data to obtain a training sample; constructing a radar echo quantitative rainfall estimation model based on a Z-R empirical formula, a residual network and a Transform algorithm; training the radar echo quantitative rainfall estimation model through the training sample; and inputting echo data to be predicted into the trained radar echo quantitative rainfall estimation model to obtain rainfall estimation data. According to the method, priori knowledge is provided for the model by using a Z-R empirical formula, and the generalization ability of the model is enhanced. Through introduction of the Transform network, the model can fully mine space-time dynamic characteristics in radar echo data, richer information is provided for weather forecast, and the space-time characteristics are effectively captured. By adopting a replaceable data preprocessing method, a feature fusion mode and a network structure, the method has relatively high adaptability, and the flexibility and adaptability are improved.

Description

Technical Field

[0001] The present invention belongs to the field of computer technology, and in particular to a method, device, computer equipment and storage medium for estimating quantitative precipitation using radar echoes. Background Art

[0002] Weather radar is a remote sensing device that uses electromagnetic waves to detect meteorological targets (such as clouds, precipitation, and storms) in the atmosphere. It is widely used in weather forecasting, disaster prediction, and atmospheric research. Radar echo intensity can reflect the size and density of atmospheric targets and is often used for short-term precipitation forecasts (forecast time less than 6 hours). Short-term nowcasting systems based on radar echoes typically consist of two steps: predicting radar echoes in the short future and quantitatively estimating precipitation based on radar echoes. This paper focuses on the second step and proposes a radar echo quantitative precipitation estimation framework that deeply integrates empirical formulas and Transformer technology. The prior art closest to this invention mainly falls into three categories: the first is traditional algorithms based on the ZR empirical formula. These algorithms statistically derive the empirical relationship between the radar reflectivity factor (Z) and the precipitation rate (R) and estimate precipitation based on this empirical formula; the second is machine learning-based algorithms that establish a mapping relationship between radar echoes and precipitation by learning from large amounts of data; and the third is algorithms based on both the ZR empirical formula and deep neural networks.

[0003] The first type of algorithm (ZR empirical formula) offers the advantages of low cost and simple implementation, and can quickly provide precipitation estimates in areas with relatively stable weather conditions and abundant historical data. However, this algorithm has significant regional limitations, as the parameters in its formula are highly dependent on historical precipitation data for a specific region. When applied to other regions or during complex weather conditions (such as severe convective weather or typhoons), the estimation error can be large. Furthermore, this algorithm cannot effectively capture the spatiotemporal dynamics of radar echo data, lacking sufficient analytical capabilities for the evolution of precipitation.

[0004] The second type of algorithm (machine learning) has the ability to handle nonlinear relationships. When sufficient data is available and weather patterns are relatively common, it can perform a certain degree of feature extraction and precipitation estimation on complex radar echo data. However, a large amount of labeled data is required as a training basis. In practice, high-quality radar echo and precipitation data with simultaneous annotation is expensive and difficult to obtain. Furthermore, this type of algorithm has poor model generalization capabilities, significantly reducing the accuracy of the estimation results in scenarios with small sample data or when encountering complex weather patterns not covered in the training data.

[0005] The third type of algorithm (based on empirical formula and deep neural network) is a very popular quantitative precipitation estimation technology, which can give full play to the advantages of empirical formula and deep neural network. However, this type of algorithm only uses the precipitation rate predicted by the empirical formula as the middle layer of the deep neural network, and does not deeply integrate the empirical formula with the deep neural network. At the same time, this type of algorithm is currently in its infancy, and existing related work is relatively limited. The algorithm of this type that is currently closest to the present invention comes from the paper [1] , proposed a fusion model of CNN+LSTM+FC+ZR (corresponding paper), the core of which is that the ZR formula parameters (a, b) are trained through deep learning, but it has limitations such as shallow fusion level, insufficient capture of spatiotemporal features and limited generalization ability. Summary of the Invention

[0006] In order to solve the above technical problems, the embodiments of the present invention provide:

[0007] A method for estimating quantitative precipitation using radar echoes, comprising:

[0008] Acquiring historical echo data, and performing denoising and normalization processing on the historical echo data to obtain training samples;

[0009] A radar echo quantitative precipitation estimation model is constructed based on the ZR empirical formula, residual network and Transformer algorithm;

[0010] Training the radar echo quantitative precipitation estimation model using the training samples;

[0011] The echo data to be predicted is input into the trained radar echo quantitative precipitation estimation model to obtain precipitation estimation data.

[0012] Furthermore, the acquiring of historical echo data and performing denoising and normalization processing on the historical echo data to obtain training samples includes:

[0013] Acquiring the historical echo data within a preset horizontal grid according to a scanning frequency of the weather radar;

[0014] Removing data whose echo reflectivity and precipitation rate do not meet a preset range from the historical echo data to obtain an initial sample;

[0015] The initial samples are normalized according to a data normalization algorithm so that the echo reflectivity and precipitation rate satisfy the standard normal distribution.

[0016] Wherein, the data normalization algorithm is: Among them, x is the initial sample, x′ is the normalized sample, x i is the mean of the data, n is the number of data, and σ is the standard deviation of the data.

[0017] Furthermore, the radar echo quantitative precipitation estimation model based on the ZR empirical formula, residual network and Transformer algorithm is constructed, including:

[0018] Use the ZR empirical formula Z=aR b Multiple sets of preset empirical data and original echo data are calculated to obtain multiple empirical precipitation rates. The empirical formula ZR is: Z=aR b , a, b are empirical values, R is precipitation;

[0019] Extract echo features from the normalized echo data using multiple residual blocks;

[0020] The empirical precipitation rate and the echo feature are reshaped into the same size using a reshaping layer and input into the precipitation-echo decoding layer as input data. The data output by the precipitation-echo decoding layer and the empirical precipitation rate are merged using a merging layer. After multiple repetitions, the final predicted precipitation is obtained using a fully connected feedforward network layer.

[0021] Furthermore, the empirical precipitation rate is: R′=[R′1, R′2, ..., R′6], where Z=10 x / 10,

[0022] x represents the original radar echo reflectivity dBZ, a i and b i Represents the empirical data a and b values ​​of group i.

[0023] Furthermore, the local feature h is:

[0024] h=ResNet4(ResNet3(ResNet2(ResNet1(x′)))),

[0025] Among them, ResNeti represents the i-th residual block, which is the normalized echo data.

[0026] Furthermore, the reshaping layer adopts a masked multi-head attention network and a multi-head attention network, wherein the multi-head attention network is based on the attention mechanism of the echo feature branch (Q) and the empirically estimated precipitation (K and V).

[0027] Multi-head attention network formula definition:

[0028] MultiHead(Q,K,V)=Concat(head1,...,head m )W o , among which, head i =Attention(QWi Q , KW i K , VW i V ) and is a learnable parameter;

[0029] Attention function definition:

[0030] A device for estimating quantitative precipitation using radar echoes, comprising:

[0031] An acquisition module is used to acquire historical echo data, and perform denoising and normalization processing on the historical echo data to obtain training samples;

[0032] The processing module is used to build a radar echo quantitative precipitation estimation model based on the ZR empirical formula, residual network and Transformer algorithm;

[0033] The processing module is further configured to train the radar echo quantitative precipitation estimation model using the training samples;

[0034] The execution module is used to input the echo data to be predicted into the trained radar echo quantitative precipitation estimation model to obtain precipitation estimation data.

[0035] Furthermore, the acquisition module includes:

[0036] A first acquisition submodule is configured to acquire the historical echo data within a preset horizontal grid according to a scanning frequency of the weather radar;

[0037] A first processing submodule is configured to remove data whose echo reflectivity and precipitation rate do not meet a preset range from the historical echo data to obtain an initial sample;

[0038] The first execution submodule is used to perform normalization processing on the initial sample according to the data normalization algorithm so that the echo reflectivity and precipitation rate satisfy the standard normal distribution.

[0039] Wherein, the data normalization algorithm is: Among them, x is the initial sample, x′ is the normalized sample, x i is the mean of the data, n is the number of data, and σ is the standard deviation of the data.

[0040] Furthermore, the processing module includes:

[0041] The second acquisition submodule is used to use the ZR empirical formula Z=aR bA plurality of empirical precipitation rates are calculated from the preset multiple sets of empirical data and the original echo data, and a Z-R empirical formula is Z=aR b , a and b are empirical values, and R is a precipitation amount.

[0042] The second processing submodule is configured to extract echo features from the normalized echo data by using the plurality of residual blocks.

[0043] The second execution submodule is configured to reshape the empirical precipitation rate and the echo features by using a reshaping layer to obtain the same size, input the same size as input data into a precipitation-echo decoding layer at the same time, merge the data output by the precipitation-echo decoding layer and the empirical precipitation rate by using a merging layer, and obtain a finally predicted precipitation amount by using a fully connected feedforward network layer after multiple repetitions.

[0044] Further, the empirical precipitation rate is R'= [R'1, R'2,..., R'6], wherein, Z=10 x / 10,

[0045] x represents the original radar echo reflectivity dBZ, a i and b i represent the empirical data a and b values of the i-th set.

[0046] Further, the echo feature h is:

[0047] h=ResNet4(ResNet3(ResNet2(ResNet1(x')))

[0048] wherein ResNeti represents the i-th residual block, and x' is the normalized echo data.

[0049] Further, the reshaping layer adopts a mask multi-head attention network and a multi-head attention network, wherein the multi-head attention network is based on an attention mechanism of an echo feature branch (Q) and an empirical estimated precipitation (K and V),

[0050] The formula definition of the multi-head attention network is:

[0051] MultiHead(Q, K, V)=Concat(head1,...,head m )W o , wherein head i =Attention(QW i Q , KW i K , VW i V ) and is a learnable parameter;

[0052] Attention function definition:

[0053] A computer device includes a memory and a processor, wherein the memory stores computer-readable instructions, and when the computer-readable instructions are executed by the processor, the processor executes the steps of the flight recovery method described above.

[0054] A storage medium storing computer-readable instructions, wherein when the computer-readable instructions are executed by one or more processors, the one or more processors are caused to perform the steps of the flight recovery method described above.

[0055] The beneficial effects of the embodiments of the present invention are as follows: the present invention uses the ZR empirical formula to provide prior knowledge for the model, reduces the model's dependence on a large amount of labeled data, enables the model to maintain good performance in different regions and with different amounts of data, and enhances the generalization ability of the model. In small sample data scenarios, the performance of the present invention is significantly better than traditional neural network algorithms. The introduction of the Transformer network enables the model to fully explore the spatiotemporal dynamic characteristics in radar echo data, conduct a more in-depth analysis and prediction of the evolution of precipitation, provide richer information for weather forecasting, and effectively capture spatiotemporal characteristics. By adopting replaceable data preprocessing methods, feature fusion methods, and network structures, the present invention can be flexibly adjusted according to different application scenarios and data characteristics, has strong adaptability, and improves flexibility and adaptability. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative work.

[0057] Figure 1 A schematic flow chart of a method for estimating quantitative precipitation using radar echoes provided in an embodiment of the present invention;

[0058] Figure 2 A schematic diagram of the network structure of a radar echo quantitative precipitation estimation model that integrates empirical formulas and machine learning, provided in an embodiment of the present invention;

[0059] Figure 3 A schematic diagram of the network structure of a residual block in a radar echo feature extraction module provided in an embodiment of the present invention;

[0060] Figure 4A schematic diagram of a precipitation-echo relationship fitting algorithm flow chart provided in an embodiment of the present invention;

[0061] Figure 5 A schematic diagram showing a network structure of a precipitation-echo decoding layer provided in an embodiment of the present invention;

[0062] Figure 6 A schematic diagram of a computer structure provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0063] In order to enable those skilled in the art to better understand the solutions of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention.

[0064] like Figure 1 As shown, a method for estimating quantitative precipitation using radar echoes includes:

[0065] S1. Acquire historical echo data, and perform denoising and normalization processing on the historical echo data to obtain training samples;

[0066] S2. Build a radar echo quantitative precipitation estimation model based on the ZR empirical formula, residual network and Transformer algorithm;

[0067] S3. Training the radar echo quantitative precipitation estimation model using the training samples;

[0068] S4. Inputting the echo data to be predicted into the trained radar echo quantitative precipitation estimation model to obtain precipitation estimation data.

[0069] The present invention uses the ZR empirical formula to provide prior knowledge for the model, reducing the model's dependence on a large amount of labeled data, so that the model can maintain good performance in different regions and with different data volumes, and enhances the generalization ability of the model. In the small sample data scenario, the performance of the present invention is significantly better than the traditional neural network algorithm. The introduction of the Transformer network enables the model to fully explore the spatiotemporal dynamic characteristics in the radar echo data, conduct a more in-depth analysis and prediction of the evolution process of precipitation, provide richer information for weather forecasting, and effectively capture spatiotemporal characteristics. By adopting replaceable data preprocessing methods, feature fusion methods and network structures, the present invention can be flexibly adjusted according to different application scenarios and data characteristics, has strong adaptability, and improves flexibility and adaptability.

[0070] Wherein, in step S1, historical echo data is obtained, and the historical echo data is subjected to denoising and normalization processing to obtain training samples, including:

[0071] Step 1: Acquire the historical echo data within a preset horizontal grid according to the scanning frequency of the weather radar;

[0072] Step 2: removing data whose echo reflectivity and precipitation rate do not meet a preset range from the historical echo data to obtain an initial sample;

[0073] Step 3: normalize the initial samples according to a data normalization algorithm so that the echo reflectivity and precipitation rate satisfy the standard normal distribution.

[0074] In one embodiment of the present invention, data can be acquired from regions with high rainfall, such as South China. Historical echo data uses radar live-stripping spatiotemporal feature data. Specifically, 7×7×10 vector data is used as input, with a 7×7 horizontal grid at an altitude of 2.5 km (with a horizontal resolution of 1 km), and 10 echo images per hour (the weather radar scans every 6 minutes). The output data is a numerical value (the amount of precipitation at the center grid point within that hour).

[0075] Denoising is performed during radar echo data processing. Specifically, samples with echo reflectivity (dBZ < 10), precipitation rate (R <= 0 or R > 120 mm / h), and average dBZ > 60 and R = 0 are removed to ensure data rationality and quality, obtaining the initial sample pair (x, y). A data normalization algorithm is used to normalize the processed data so that it conforms to a standard normal distribution (mean 0, standard deviation 1), resulting in the normalized sample pair (x′, y′). The normalization calculation method is shown in the formula:

[0076] Among them, x is the initial sample, x′ is the normalized sample, x i is the mean of the data, n is the number of data points, and σ is the standard deviation of the data. Precipitation is also processed from y to y′ using the same formula. This processing method ensures that the radar reflectivity and precipitation data conform to a standard normal distribution, facilitating subsequent model training and calculations.

[0077] This paper mainly builds a radar echo quantitative precipitation estimation model based on the ZR empirical formula, residual network, and Transformer technology. The model mainly consists of three modules: empirical precipitation rate generation, echo feature extraction, and precipitation-echo relationship fitting module.

[0078] In step S2, a radar echo quantitative precipitation estimation model is constructed based on the ZR empirical formula, residual network, and Transformer algorithm, including:

[0079] Step 1: Use the ZR empirical formula Z=aR bMultiple sets of preset empirical data and original echo data are calculated to obtain multiple empirical precipitation rates. The empirical formula ZR is: Z=aR b , a, b are empirical values, R is precipitation;

[0080] Step 2: Extract echo features from the normalized echo data using multiple residual blocks;

[0081] Step 3: Use the reshaping layer to reshape the empirical precipitation rate and the echo feature to the same size, and input them into the precipitation-echo decoding layer as input data. Use the merging layer to merge the data output by the precipitation-echo decoding layer and the empirical precipitation rate. After multiple repetitions, use the fully connected feedforward network layer to obtain the final predicted precipitation.

[0082] In one embodiment, Figure 2 The figure shows the flow chart of empirical precipitation generation, where the empirical precipitation rate algorithm can be calculated based on the original radar echo using the ZR empirical formula and six sets of (a, b) parameters, specifically:

[0083] Using the traditional ZR empirical formula (Z = aR b ) to obtain the preliminary empirical precipitation rate. The input is 7×7×10 size data. In order to give full play to the potential role of the empirical formula, this algorithm uses the commonly used 6 sets of a, b values ​​and the radar echo data based on the original size to calculate multiple empirical precipitation rates. The generation module of R′ obtained from the radar echo reflectivity dBZ through the ZR empirical formula and 6 sets of a and b values ​​is as follows:

[0084] R′=[R′1,R′2,...,R′6]

[0085] in, Z=10 x / 10, x represents the original radar echo reflectivity dBZ (not normalized), a i and b i The a and b values ​​of the 6 groups are as shown in Table 1.

[0086] Table 1. Six sets of parameters used in the ZR empirical formula

[0087] Group 1 Group 2 Group 3 Group 4 Group 5 Group 6 a 303 405 289 219 67.66 66.5 b 1.7 1.49 1.59 1.41 1.94 1.92

[0088] In one embodiment, Figure 3 Figure 2 shows the network structure of a residual block in the radar echo feature extraction module. The algorithm used for echo feature extraction can be implemented based on normalized radar echoes using four convolutional neural networks + residual structures. Specifically:

[0089] The algorithm's feature extraction module is based on the Residual Network (ResNet) structure and Convolutional Neural Network (CNN) technology. Residual networks allow the network to learn residual mappings rather than raw input-output mappings. This allows the gradient of the residual network to be directly transferred back to shallow layers via skip connections during backpropagation, thus avoiding gradient attenuation caused by deeper layers. Convolutional neural networks are neural networks designed specifically for processing grid-like data. They achieve hierarchical feature extraction through operations such as local connections and weight sharing in the convolutional layer and downsampling in the pooling layer.

[0090] This algorithm extracts echo features (h) from the normalized radar echo (x′) by connecting four residual blocks in series. Each residual block consists of two convolutional layers with 3×3 convolution kernels and a residual connection. The mathematical expression of the feature extraction module of this algorithm is as follows:

[0091] h=ResNet4(ResNet3(ResNet2(ResNet1(x))))

[0092] h1=ResNet1(x′)=ReLU(x+h′1)

[0093] h′1=BN(Conv2d(ReLU(BN(Conv2d(x′)))))

[0094] Among them, ResNeti represents the i-th residual block, h i It represents the output of the ith residual block (and the input of the i+1th residual block, except for the last one, h1 is the output of the first residual block and the input of the second residual block), and the final extracted echo feature is represented by h = h4; h′ i is the output of the two concatenated convolutional layers in the i-th residual block (h′1 is the output of the two concatenated convolutional layers in the first residual block); Conv2d represents the two-dimensional convolution operation, BN (Batch Normalization) is batch normalization, and ReLU (Rectified Linear Unit) is the activation function used in the network. Each two-dimensional convolution operation in the feature extraction module of this algorithm uses 10 filters of size 3×3.

[0095] The precipitation-echo relationship fitting module uses the decoding layers in three Transformers to deeply fuse the empirical precipitation rate and echo features to obtain the final predicted precipitation.

[0096] like Figure 4The figure below is a flow chart of the precipitation-echo relationship fitting algorithm. The relationship fitting module of this algorithm is based on both the empirically estimated precipitation (R′) and echo features (h) mentioned above, and is implemented using the Decoder structure in the Transformer. The settings of the hyperparameters for each layer of the network are shown in Table 2. In this module, the radar echo features and the empirically estimated precipitation are first reshaped through two layers to obtain data of the same size, which are then passed as two input data to the precipitation-echo decoding layer. Secondly, the output of the decoding layer and the empirically estimated precipitation are respectively passed through two merging layers to obtain more condensed features, which serve as the input of the next layer of precipitation-echo decoding layer. Finally, after three cycles, the fully connected feedforward network layer is used to obtain the final predicted precipitation amount. The reshaping layer, precipitation-echo decoding layer, merging layer, and fully connected feedforward network layer involved in this module are described as follows:

[0097] The reshape layer primarily uses the reshape function and linear transformation to ensure that the empirically estimated precipitation and echo features have the same size. First, the reshape function transforms R′ / h to a size of 49×60 / 10. Second, a learnable parameter matrix of size 60 / 10×40 is used to linearly transform the empirically estimated precipitation and echo features to 49×40. The linear transformation matrix involved is automatically learned by the network through model training.

[0098] The precipitation-echo decoding layer is built based on the Decoder module in Transformer to achieve deep fusion of empirically estimated precipitation and echo features and fit the relationship between the two. Figure 5 Figure 2 shows the network structure of the precipitation-echo decoding layer. Two types of attention mechanisms are used in the precipitation-echo decoding layer: a masked multi-head attention network and a multi-head attention network. The masked multi-head attention network is a self-attention mechanism based on the echo feature branch, with a mask superimposed on it. The multi-head attention network used in this algorithm is based on the attention mechanism of the echo feature branch (Q) and empirically estimated precipitation (K and V). The multi-head attention network formula is defined as follows:

[0099] MultiHead(Q,K,V)=Concat(head1,....,head m )W O , among which, head i =Attention(QW i Q , KW i K , VWV i V ),in and is a learnable parameter; the attention function is defined as: In this algorithm, three precipitation-echo decoding layers are connected in series to fit the relationship between the two. Each precipitation-echo decoding layer adopts a 5-head attention mechanism, that is, m = 5; the D of the three precipitation-echo decoding layers are 40, 80 and 160 respectively. k They are 8, 10, and 20 respectively. Masking is to add a mechanism to the Attention function so that the current feature in a sample is only related to the feature before it, that is, the connection between the current feature and the subsequent feature is set to negative infinity. The masking formula is as follows:

[0100]

[0101] The Add+Norm layer normalizes the output of the network's intermediate layers and is defined as: layernorm(x+sublayer(x)). Both the precipitation-echo decoding layer and the precipitation-echo relationship fitting module utilize a fully connected feed-forward network (FFN).

[0102] The merging layer utilizes padding, reshape functions, and linear transformations to summarize and abstract the local information of the current feature, preserving the relatively important data features. Initially, the first dimension of the data is padded with zeros to a size divisible by 4. Second, the first dimension is divided by 4 and the second dimension is multiplied by 4 (d·4) using the reshape function. Finally, the linear transformation of the data is performed using a learnable weight matrix of size d·4 × d·2. In the precipitation-echo relationship fitting module network of this algorithm, the parameters d involved in the three merging layers are set to 40, 80, and 160, respectively.

[0103] The fully connected feedforward network layer uses two fully connected layers with a ReLU activation function added in the middle. The specific formula is: FFN(x) = ReLU(W1x+b1)W2+b2. The output size of the middle layer of the fully connected feedforward network of the three series-connected precipitation-echo decoding layers is 2048 (W1 and The sizes of the three precipitation-echo decoding layers are 40×2048, 80×2048, and 160×2048, respectively. The b1 in the fully connected feedforward networks of the three series-connected precipitation-echo decoding layers is 2048, and the b2 are 40, 80, and 160, respectively. The precipitation-echo relationship fitting module finally uses a fully connected feedforward network to obtain the final precipitation amount: in and

[0104] In step S3, the radar echo quantitative precipitation estimation model is trained using the training samples, specifically:

[0105] In the model training phase, the present invention uses 500,000 samples to train the radar echo quantitative precipitation estimation model. The detailed settings of the training model are: Adam optimization algorithm, piecewise learning rate (learning rate, lr; before 98 training rounds: lr = 1e-4 and after: lr = 1e-5), least squares method is the loss function of the model, 248 training rounds, and the error measurement standard is Mean Squared Error (MSE). During the training phase, 80% of the training samples are used to solve the weights of the model, and the remaining 20% ​​are used to evaluate the performance of the model. For the loss values ​​of the first 98 training rounds (lr=1e-4) and the 150 training rounds after 98 training rounds (lr=1e-5), after the segmented 248 training rounds, although the training loss value still has a downward trend, the verification loss value has become tight and flat. Therefore, in order to prevent the model from overfitting and ensure the generalization ability of the model, the present invention determines the final weight of the model after 248 training rounds.

[0106] Regarding the processing of the predicted data in the rainfall prediction in step S5, please refer to step S1 and will not be described in detail here.

[0107] The present invention can be applied in the following technical fields:

[0108] Weather Forecasting: This algorithm can be used in meteorological departments' real-time precipitation forecasting systems. By processing and analyzing radar echo data in real time, it can quickly and accurately estimate precipitation intensity and distribution, providing a key basis for weather forecasting and disaster warning. For example, before the arrival of typhoons, rainstorms, and other disastrous weather, using this algorithm to accurately estimate precipitation can help formulate response measures and reduce disaster losses.

[0109] Hydrological Monitoring: Combined with hydrological monitoring equipment, this system provides real-time estimates of precipitation within a watershed, supporting water resource management, flood forecasting, and water conservancy project scheduling. Accurate precipitation estimates enable more effective predictions of hydrological parameters such as river flows and reservoir levels, ensuring the safe operation of water conservancy facilities.

[0110] Agricultural disaster warning: In agricultural production, it can be used to monitor precipitation in farmland areas, providing timely warnings of agricultural disasters such as floods and droughts, and providing decision support for agricultural production. For example, by accurately estimating precipitation, farmers can rationally arrange irrigation, drainage, and other agricultural activities, reducing the impact of disasters on crops.

[0111] Urban Flooding Prediction: This algorithm combines urban geographic information with drainage system data to estimate precipitation distribution within urban areas and predict urban flooding risks, providing a reference for urban planning and drainage system design. By understanding flooding risk areas in advance, urban management departments can take appropriate measures, such as strengthening drainage infrastructure and developing emergency evacuation plans, to improve the city's flood control and drainage capabilities.

[0112] Aerospace: In aerospace activities, it can be used to monitor and predict precipitation along the flight path to ensure flight safety. For example, during takeoff and landing, accurate precipitation estimation can help pilots judge weather conditions and choose appropriate takeoff and landing times and routes.

[0113] The present invention provides a device for estimating quantitative precipitation using radar echoes, comprising:

[0114] An acquisition module is used to acquire historical echo data and perform denoising and normalization on the historical echo data to obtain training samples; a processing module is used to construct a radar echo quantitative precipitation estimation model based on the ZR empirical formula, the residual network, and the Transformer algorithm; the processing module is also used to train the radar echo quantitative precipitation estimation model using the training samples; and an execution module is used to input the echo data to be predicted into the trained radar echo quantitative precipitation estimation model to obtain precipitation estimation data.

[0115] The present invention uses the ZR empirical formula to provide prior knowledge for the model, reducing the model's dependence on a large amount of labeled data, so that the model can maintain good performance in different regions and with different data volumes, and enhances the generalization ability of the model. In the small sample data scenario, the performance of the present invention is significantly better than the traditional neural network algorithm. The introduction of the Transformer network enables the model to fully explore the spatiotemporal dynamic characteristics in the radar echo data, conduct a more in-depth analysis and prediction of the evolution process of precipitation, provide richer information for weather forecasting, and effectively capture spatiotemporal characteristics. By adopting replaceable data preprocessing methods, feature fusion methods and network structures, the present invention can be flexibly adjusted according to different application scenarios and data characteristics, has strong adaptability, and improves flexibility and adaptability.

[0116] In some embodiments, the acquisition module includes: a first acquisition submodule, configured to acquire the historical echo data within a preset horizontal grid according to a scanning frequency of a weather radar; a first processing submodule, configured to remove data in the historical echo data whose echo reflectivity and precipitation rate do not meet a preset range to obtain an initial sample; and a first execution submodule, configured to normalize the initial sample according to a data normalization algorithm so that the echo reflectivity and precipitation rate meet a standard normal distribution, wherein the data normalization algorithm is: Among them, x is the initial sample, x′ is the normalized sample, x i is the mean of the data, n is the number of data, and σ is the standard deviation of the data.

[0117] In some embodiments, the processing module includes: a second acquisition submodule for using the ZR empirical formula Z=aR b Multiple sets of preset empirical data and original echo data are calculated to obtain multiple empirical precipitation rates. The empirical formula ZR is: Z=aR b , a, b are empirical values, and R is precipitation; the second processing submodule is used to extract echo features from the normalized echo data using multiple residual blocks; the second execution submodule is used to use the reshaping layer to reshape the empirical precipitation rate and the echo features to the same size, and input them as input data to the precipitation-echo decoding layer at the same time, and use the merging layer to merge the data output by the precipitation-echo decoding layer and the empirical precipitation rate. After multiple repetitions, the fully connected feedforward network layer is used to obtain the final predicted precipitation.

[0118] In some embodiments, the empirical precipitation rate is: R′=[R′1, R′2, ..., R′6], where Z=10 x / 10, x represents the original radar echo reflectivity dBZ, a i and b i Represents the empirical data a and b values ​​of group i.

[0119] In some embodiments, the echo characteristic h is:

[0120] h=ResNet4(ResNet3(ResNet2(ResNet1(x′)))), where ResNeti represents the i-th residual block, which is the normalized echo data.

[0121] Furthermore, the reshaping layer uses a masked multi-head attention network and a multi-head attention network. The multi-head attention network is based on the attention mechanism of the echo feature branch (Q) and the empirically estimated precipitation (K and V). The multi-head attention network formula is defined as:

[0122] MultiHead(Q,K,V)=Concat(head1,...,head m )W o ,in,

[0123] head i =Attention(QW i Q , KW i K , VWi V ), and is a learnable parameter;

[0124] Attention function definition:

[0125] In order to solve the above technical problems, the embodiment of the present invention also provides a computer device, such as Figure 6 As shown, the computer device includes a processor, a non-volatile storage medium, a memory, and a network interface connected via a system bus. The non-volatile storage medium of the computer device stores an operating system, a database, and computer-readable instructions. The database may store a control information sequence. When executed by the processor, the computer-readable instructions enable the processor to implement a method for estimating quantitative precipitation using radar echoes. The processor of the computer device is used to provide computing and control capabilities, supporting the operation of the entire computer device. The memory of the computer device may store computer-readable instructions. When executed by the processor, the computer-readable instructions enable the processor to implement a method for estimating quantitative precipitation using radar echoes. The network interface of the computer device is used to connect and communicate with a terminal.

[0126] In this embodiment, the processor is used to execute the specific contents of the acquisition module, processing module, and execution module. The memory stores the program code and various data required to execute these modules. The network interface is used to transmit data between user terminals and servers. The memory in this embodiment stores the program code and data required to execute all submodules of the flight recovery fee method. The server can call the server's program code and data to execute the functions of all submodules.

[0127] The present invention also provides a storage medium storing computer-readable instructions. When the computer-readable instructions are executed by one or more processors, the one or more processors execute the steps of the method for estimating quantitative precipitation based on radar echoes described in any of the above embodiments.

[0128] Those skilled in the art will appreciate that all or part of the processes in the above-described method embodiments can be implemented by instructing the relevant hardware through a computer program. The computer program can be stored in a computer-readable storage medium. When executed, the program can include the processes in the above-described method embodiments. The aforementioned storage medium can be a non-volatile storage medium such as a magnetic disk, an optical disk, a read-only memory (ROM), or a random access memory (RAM).

[0129] It should be understood that although the steps in the flowcharts of the accompanying drawings are shown in sequence as indicated by the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some of the steps in the flowcharts of the accompanying drawings may include multiple sub-steps or multiple stages, and these sub-steps or stages are not necessarily executed at the same time, but can be executed at different times, and their execution order is not necessarily sequential, but can be executed in turn or alternately with other steps or at least a portion of the sub-steps or stages of other steps.

[0130] The above descriptions are only partial embodiments of the present invention. It should be pointed out that ordinary technicians in this technical field can make several improvements and modifications without departing from the principles of the present invention. These improvements and modifications should also be regarded as within the scope of protection of the present invention.

Claims

1. A method for estimating quantitative precipitation using radar echoes, characterized in that: include: Acquiring historical echo data, and performing denoising and normalization processing on the historical echo data to obtain training samples; A radar echo quantitative precipitation estimation model is constructed based on the ZR empirical formula, residual network and Transformer algorithm; Training the radar echo quantitative precipitation estimation model using the training samples; The echo data to be predicted is input into the trained radar echo quantitative precipitation estimation model to obtain precipitation estimation data.

2. The estimation method according to claim 1, wherein: The acquiring of historical echo data and performing denoising and normalization processing on the historical echo data to obtain training samples includes: Acquiring the historical echo data within a preset horizontal grid according to a scanning frequency of the weather radar; Removing data whose echo reflectivity and precipitation rate do not meet a preset range from the historical echo data to obtain an initial sample; The initial samples are normalized according to a data normalization algorithm so that the echo reflectivity and precipitation rate satisfy the standard normal distribution. Wherein, the data normalization algorithm is: Among them, x is the initial sample, x′ is the normalized sample, x i is the mean of the data, n is the number of data, and σ is the standard deviation of the data.

3. The estimation method according to claim 1, wherein: The radar echo quantitative precipitation estimation model based on the ZR empirical formula, residual network and Transformer algorithm includes: Use the ZR empirical formula Z=aR b Multiple sets of preset empirical data and original echo data are calculated to obtain multiple empirical precipitation rates. The empirical formula ZR is: Z = aR b , a, b are empirical values, R is precipitation; Extract echo features from the normalized echo data using multiple residual blocks; The empirical precipitation rate and the echo feature are reshaped into the same size using a reshaping layer and input into the precipitation-echo decoding layer as input data. The data output by the precipitation-echo decoding layer and the empirical precipitation rate are merged using a merging layer. After multiple repetitions, the final predicted precipitation is obtained using a fully connected feedforward network layer.

4. The estimation method according to claim 3, wherein: The empirical precipitation rate is: R′=[R′1, R′2, ..., R′6], where Z=10 x / 10 , x represents the original radar echo reflectivity dBZ, a i and b i Represents the empirical data a and b values ​​of group i.

5. The estimation method according to claim 3, wherein: The echo feature h is: h = ResNet4(ResNet3(ResNet2(ResNet1(x′)))), Among them, ResNeti represents the i-th residual block, which is the normalized echo data.

6. The estimation method according to claim 3, wherein: The reshaping layer uses a masked multi-head attention network and a multi-head attention network, where the multi-head attention network is based on the attention mechanism of the echo feature branch (Q) and the empirically estimated precipitation (K and V). Multi-head attention network formula definition: MultiHead(Q,K,V)=Concat(head1,...,head m )W o Among them, head i =Attention(QW i Q , KW i k , VW i V ), and are learnable parameters, where MultiHead(Q, K, V) represents a multi-head attention function. The Concat operation concatenates the outputs of m heads along the feature dimension into a complete tensor. Q, K, and V represent query, key, and value, respectively. The attention function is defined as: D represents the dimension of the input feature, d k is the dimension of each attention head, D = m × d k , d out Represents the dimension of the final output of multi-head attention.

7. A device for estimating quantitative precipitation using radar echoes, characterized in that: include: An acquisition module is used to acquire historical echo data, and perform denoising and normalization processing on the historical echo data to obtain training samples; The processing module is used to build a radar echo quantitative precipitation estimation model based on the ZR empirical formula, residual network and Transformer algorithm; The processing module is further configured to train the radar echo quantitative precipitation estimation model using the training samples; The execution module is used to input the echo data to be predicted into the trained radar echo quantitative precipitation estimation model to obtain precipitation estimation data.

8. The estimation device according to claim 7, wherein: The acquisition module includes: A first acquisition submodule is configured to acquire the historical echo data within a preset horizontal grid according to a scanning frequency of the weather radar; A first processing submodule is configured to remove data whose echo reflectivity and precipitation rate do not meet a preset range from the historical echo data to obtain an initial sample; The first execution submodule is used to perform normalization processing on the initial sample according to the data normalization algorithm so that the echo reflectivity and precipitation rate satisfy the standard normal distribution. Wherein, the data normalization algorithm is: Among them, x is the initial sample, x′ is the normalized sample, x i is the mean of the data, n is the number of data, and σ is the standard deviation of the data.

9. A computer device comprising a memory and a processor, wherein the memory stores computer-readable instructions, and when the processor executes the computer-readable instructions, the processor performs the steps of the method for estimating quantitative precipitation based on radar echoes according to any one of claims 1 to 6.

10. A storage medium storing computer-readable instructions, wherein when the computer-readable instructions are executed by one or more processors, the one or more processors are caused to perform the steps of the method for estimating quantitative precipitation based on radar echoes according to any one of claims 1 to 6.

Images