A short-term precipitation prediction method, device, terminal equipment and storage medium

Abstract

The application discloses a short-term and nowcasting precipitation prediction method based on GRU and LSTM space-time fusion, which comprises the following steps: step 1, reading local radar data and performing data preprocessing operation; step 2, dividing the processed data into a training set, a test set and a verification set; step 3, designing an ST-LSGRU basic module, stacking 4 layers to build a radar echo extrapolation model; step 4, inputting the training set, initializing model parameters, training a network model and saving model parameters; step 5, testing model performance by using the verification set, optimizing model hyperparameters, training a network model and saving model parameters; step 6, loading saved model parameters, inputting the test set, obtaining radar extrapolation results and obtaining precipitation in a future period of time through a Z-R relationship. The method can solve the gradient vanishing problem caused by long-time sequence prediction, and effectively improve the definition and stability of the prediction results of the model for strong precipitation weather.

Description

Technical Field

[0001] This invention belongs to the field of short-term meteorological disaster forecasting technology, specifically relating to a short-term precipitation forecasting method, device, terminal equipment, and storage medium. Background Technology

[0002] Radar echo extrapolation is the foundation for meteorological departments to conduct short-term forecasts of potential severe convective weather. Through short-term forecasting, severe weather events such as thunderstorms, heavy precipitation, and hail can be predicted within a short period (0-2 hours), enabling precise prevention and control. Radar echo extrapolation uses existing historical radar echo data to predict the shape, location, intensity, and movement information of future echoes, enabling the tracking and forecasting of severe convective weather systems. This is of great significance for disaster prevention and mitigation, hazardous weather warnings, and protecting the lives and property of the public.

[0003] In recent years, the massive amounts of existing data have attracted widespread attention from data scientists to data-driven deep learning-based nearest-neighbor prediction algorithms. For example, the paper "Wang, Yunbo, et al. Predrnn: Recurrent neural networks for predictive learning using spatiotemporal LSTMs. Advances in neural information processing systems 30 (2017)" reconstructed the LSTM unit, designing a spatiotemporal LSTM (ST-LSTM) unit. The memory state of the spatiotemporal LSTM is no longer restricted to each LSTM unit and can be passed in both vertical and horizontal directions. The paper "Wang, Yunbo, et al. Memory in memory: A predictive neural network for learning higher-order non-stationarity from spatiotemporal dynamics. Proceedings of the IEEE / CVF Conference on Computer Vision and Pattern Recognition. 2019" conducted in-depth research on the saturation of the forget gate in LSTM and introduced the idea of ​​difference, proposing the MIM structure. Compared with traditional numerical weather prediction methods, this can achieve faster and more accurate predictions. However, these methods still have many drawbacks, such as long intervals between radar data and insufficient utilization of the spatiotemporal characteristics of radar data, leading to low prediction accuracy. Furthermore, the complexity of meteorological systems makes predicting radar echo sequences even more difficult. RNNs and their improved neural network models require a long time from training to model convergence, and the large parameter scale of the models results in high hardware requirements, particularly in the inability to accurately predict the generation and dissipation of convection. Additionally, the ambiguity problem of long-tailed predictions, common in spatiotemporal forecasting, especially the ambiguity of the high reflectivity portion of radar extrapolation, cannot be well resolved by current solutions. Summary of the Invention

[0004] To address the technical problems mentioned in the background, this invention proposes a short-term precipitation prediction method based on the spatiotemporal fusion of GRU and LSTM. A novel radar echo extrapolation model is designed, using a residual connection structure and incorporating a coarse feature fast transmission unit to help memorize more spatiotemporal information, thereby improving the accuracy and clarity of long-term precipitation weather prediction results.

[0005] To achieve the above-mentioned technical objectives, the technical solution of the present invention is as follows:

[0006] A short-term precipitation prediction method based on spatiotemporal fusion of GRU and LSTM includes the following steps:

[0007] S1. Read radar data, perform filtering, noise reduction, cleaning and enhancement operations on the data to obtain the processed radar data tensor;

[0008] S2. Divide the radar data tensor obtained in step S1 into training set, validation set and test set, and slice these three datasets into input sequence and prediction sequence. After normalization, unify the data format.

[0009] S3. Construct the basic modules of ST-LSGRU, and stack them optimally to form a radar echo extrapolation model.

[0010] S4. Initialize the parameters of the radar echo extrapolation model established in step S3, set the hyperparameters, and train the radar echo extrapolation model by inputting the training set divided in step S2; obtain the model training parameters.

[0011] S5. Input the validation set divided in step S2 into the radar echo extrapolation model for training. After a certain number of training rounds, verify the trend of the radar echo extrapolation model's prediction results. Adjust the hyperparameters to make the radar echo extrapolation model reach a stable gradient update direction. Then, use the training set divided in step S2 to continue training the radar echo extrapolation model. After training, save the radar echo extrapolation model parameters.

[0012] S6. Load the radar echo extrapolation model parameters saved in step S5 and freeze them. Input the test set into the trained radar echo extrapolation model and obtain the precipitation prediction results through ZR transformation.

[0013] Preferably, step S1 specifically includes the following steps:

[0014] S1-1: Read radar data, sum the global information of radar data, calculate the number of radar pixels, distribute the summation result equally to each pixel, retain data samples with an average value greater than 0.01, and satisfy the condition that the average value of 20 consecutive sequences is greater than 0.01;

[0015] S1-2: Clean the filtered data, set the threshold to 0-70, and determine the values ​​outside the threshold as outliers. Set the values ​​less than 0 to 0 and the values ​​greater than 70 to 70.

[0016] S1-3: Denoise the cleaned radar data, identify outlier pixel values, and set them to 0 to ensure data quality and improve the prediction accuracy of data-driven deep learning models.

[0017] S1-4: Data Augmentation: The radar data is cropped around the center and formatted to obtain radar data with a resolution of 400*400. Then, each sequence is randomly flipped, rotated, and chromaticity adjusted to obtain the augmented radar data.

[0018] Preferably, step S2 specifically includes the following steps:

[0019] S2-1: Divide the radar data into training set, validation set, and test set;

[0020] S2-2: The training set, validation set and test set are divided into multiple sequences using the sliding window method. The window size is 20, the length of each sequence is 20, the time interval between units of length is 6 minutes, and each sequence corresponds to 2 hours of radar data. The first ten are used as input sequences and the last ten are used as real labels.

[0021] S2-3: Map the radar data along the channel dimension to grayscale (0-1), then unify the format of the normalized radar data. The normalization calculation method is as follows:

[0022]

[0023] In the formula, X' represents the normalized data, X represents the original radar data, and X... min With X max These represent the maximum and minimum values ​​in the radar data, respectively.

[0024] Preferably, step S3 specifically includes the following steps:

[0025] S3-1: Construct the basic ST-LSGRU module. If this module is located in layer 1, i.e., when l = 1, then the input state X is the tensor converted from the radar echo map at the current time. If this module is not a layer 1 unit, i.e., when l > 1, then the hidden state output at time t is... As the input state of this module; ST-LSGRU first inputs the state and the hidden state output at time t-1 in the same layer as this unit The input is passed through two different convolutional filters to obtain the reset gate R. t and Update Gate Z t The formula is expressed as follows:

[0026]

[0027]

[0028] In the formula, * represents convolution, and σ represents the Sigmoid function.

[0029] Input status and spacetime memory The input is fed into an ST-LSGRU and passed through three different convolutional filters to obtain the forget gate f. t Input gate i t and input modulation gate g t ; through the forgetting gate f t The Hadamard product operation is used to forget non-essential features from past moments in the temporal memory M; the input gate and input modulation gate perform element-wise addition to update the features in the memory, thus obtaining the spatiotemporal memory. The formula is expressed as follows:

[0030]

[0031]

[0032]

[0033]

[0034] In the formula, represents the dot product operation of matrices; tanh represents the hyperbolic tangent activation function, expressed by the following formula:

[0035]

[0036] Then a convolutional filter is used to process the input state. Hidden state And the updated spacetime memory Perform convolution operations to obtain new information. Will By resetting gate R t With Update Gate Z t Update and get The formula is expressed as follows:

[0037]

[0038]

[0039] In the formula, * denotes convolution;

[0040] S3-2: The ST-LSGRU modules are stacked in four layers using a residual method, with a residual connection performed every other layer; multiple coarse feature fast transfer units are embedded between the first and second layers of the stacked structure; a coarse feature fast transfer bus is formed in the time dimension; after passing through two different convolutional filters Wp and Ws, states Pt and St are obtained; then, through Hadamard product and element-wise addition, the coarse feature states are updated to obtain the output B of the current unit. t The calculation method is as follows:

[0041]

[0042]

[0043]

[0044] Preferably, step S4 specifically includes the following steps:

[0045] S4-1: Initialize the parameters of the radar echo extrapolation model, set the input sequence length input_size to 10, the output sequence length to 10, the batch size to 2, the learning rate lr to 0.001, the hidden layer convolution kernel size to 5, the number of convolution kernels n to 128, and the maximum number of training iterations iter to 10000.

[0046] S4-2: Read the training set and perform initial training. Save the radar echo extrapolation model training parameters once every 500 iterations. After 2500 training iterations, save the radar echo extrapolation model training parameters 5 times.

[0047] Preferably, adjusting the hyperparameters in step S5 specifically includes the following steps:

[0048] S5-1: Read the validation set, load the radar echo extrapolation model training parameters saved in step S4, obtain the prediction results, and save the results to the local disk;

[0049] S5-2: Based on the change in mean squared error loss, determine whether the model is predicting in the direction of decreasing mean squared error loss, and adjust the hyperparameters according to the prediction results;

[0050] S5-3: Repeat steps S4 and S5 to adjust the model hyperparameters so that the mean squared error loss function tends to converge during training; read the validation set, train the radar extrapolation model, and save the model parameters after training.

[0051] Preferably, in step S6, obtaining the future rainfall through ZR transformation specifically involves: loading the model parameters saved in step S5, reading the test set, and obtaining the rainfall amount through ZR transformation of the prediction results, as expressed by the following formula:

[0052] Z = aR b

[0053] In the formula, a is the first characteristic constant, b is the second characteristic constant, Z is the radar reflectivity, and R is the rainfall intensity.

[0054] A short-term precipitation forecasting device, comprising:

[0055] The radar data acquisition and processing module is used to acquire radar data, perform filtering, noise reduction, cleaning and enhancement operations on the data, and then divide it into training set, validation set and test set. The three datasets are then sliced ​​into input sequence and prediction sequence, and normalized to unify the data format.

[0056] The radar echo extrapolation module is used to receive data collected by the radar data acquisition and processing module, and obtain precipitation prediction results through ZR transformation.

[0057] A short-term precipitation forecasting terminal device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the method described above.

[0058] A computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method.

[0059] The beneficial effects of adopting the above technical solution are as follows:

[0060] 1. During the construction of the dataset, the dataset was filtered for severe convective weather, which improved the quality of the data;

[0061] 2: A brand-new ST-LSGTU basic module was designed to support the construction of radar echo extrapolation models, which can effectively improve training and prediction efficiency for high-resolution radar data.

[0062] 3: In the process of constructing the radar echo extrapolation model, a residual connection structure is adopted, and a coarse feature fast transmission unit (FMU) is added to form a coarse feature fast transmission bus in the time dimension, which can alleviate the problems of fuzzy and distorted prediction results caused by information attenuation and gradient vanishing.

[0063] 4. During the training process, using a validation set to adjust the model's hyperparameters helps the model converge in a stable direction, improving training speed and prediction quality. Attached Figure Description

[0064] Figure 1 This is a flowchart of the method of the present invention;

[0065] Figure 2 This is a basic module diagram of the ST-LSGRU in this invention;

[0066] Figure 3 This is a diagram of the radar echo extrapolation model with four stacked layers in this invention;

[0067] Figure 4 This is a diagram showing the extrapolation experiment results of this invention in a certain region of eastern China. Detailed Implementation

[0068] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings.

[0069] like Figure 1 As shown, this invention discloses a short-term precipitation prediction method based on the spatiotemporal fusion of GRU and LSTM, comprising the following steps:

[0070] Read radar data, sum the global information of the radar data, calculate the number of radar pixels, distribute the summation result evenly to each pixel, and retain data samples with an average value greater than 0.01. In order to ensure the continuity of the samples, the above method needs to satisfy that the average value of 20 consecutive sequences is greater than 0.01.

[0071] The filtered data is cleaned by setting a threshold of 0-70. Values ​​outside the threshold are considered outliers. Values ​​less than 0 are set to 0, and values ​​greater than 70 are set to 70.

[0072] The cleaned radar data is denoised, outlier pixel values ​​are identified and set to 0 to ensure data quality and improve the prediction accuracy of data-driven deep learning models.

[0073] Data augmentation: The radar data is cropped around the center and formatted to obtain radar data with a resolution of 400*400. Then, each sequence is randomly flipped, rotated, and chromaticity adjusted to obtain the augmented radar data.

[0074] The radar data is divided into a training set, a validation set, and a test set, accounting for 80%, 10%, and 10% respectively. The training set is used for model training; the validation set is used to adjust model parameters; and the test set can be replaced with online data to obtain prediction results for a future period of time.

[0075] The training set, validation set, and test set are divided into multiple sequences using a sliding window method. The window size is 20, the length of each sequence is 20, the time interval between units is 6 minutes, and each sequence corresponds to 2 hours of radar data. The first ten sequences are used as input sequences, and the last ten sequences are used as real labels.

[0076] The radar data is mapped along the channel dimension and normalized to grayscale data of 0-1. The normalization calculation method is as follows:

[0077]

[0078] Where X' represents the normalized data, X represents the original radar data, and X... min With X max These represent the maximum and minimum values ​​in the radar data, respectively, and standardize the normalized radar data to facilitate subsequent training and testing.

[0079] ST-LSGRU Figure 2 As shown.

[0080] Construct a basic ST-LSGRU (Spatio-temporal Long Short-term Gate Recurrent Unit) module. If this module is located in Layer 1 (i.e., when l = 1), then the input state X is the tensor converted from the radar echo map at the current time. If this module is not a Layer 1 unit (i.e., l > 1), then the hidden state output at time t is... This serves as the input state for the module. The ST-LSGRU first sets the input state... and the hidden state output at time t-1 in the same layer as this unit The input is passed through two different convolutional filters to obtain the reset gate R. t Update Gate Z t :

[0081]

[0082]

[0083] Where * denotes convolution, and σ denotes the Sigmoid function.

[0084] Input status and spacetime memory The input is fed into an ST-LSGRU and passed through three different convolutional filters to obtain the forget gate f. t Input gate i t and input modulation gate g t Through the forgetting gate f t The Hadamard product operation is used to forget non-essential features from past moments in the temporal memory M. Element-wise addition between the input gate and the input modulation gate updates the features in the memory, resulting in the spatiotemporal memory. The calculation method is as follows:

[0085]

[0086]

[0087]

[0088]

[0089] Where ⊙ denotes the dot product of matrices, and tanh denotes the hyperbolic tangent activation function.

[0090]

[0091] Then a convolutional filter is used to process the input state. Hidden state And the updated spacetime memory Perform convolution operations to obtain new information. Will By resetting gate R t With Update Gate Z t Update and get

[0092]

[0093]

[0094] Where * denotes convolution;

[0095] The four-layer stacked ST-LSGRU model is as follows Figure 3 As shown.

[0096] The ST-LSGRU modules are stacked in four layers using a residual method, with a residual connection performed every other layer. Multiple coarse feature fast transfer units (FMUs) are embedded between the first and second layers of the stack, forming a coarse feature fast transfer bus (FMB) in the time dimension. After passing through two different convolutional filters Wp and Ws, the states Pt and St are obtained, similar to the input gate and input modulation gate in an LSTM unit. These states contain the current time step. And the previous moment B t-1 The coarse state is obtained, and then the coarse feature state is updated through Hadamard product and element-wise addition to obtain the output B of the current unit. t The calculation method is as follows:

[0097]

[0098]

[0099]

[0100] Before the radar data is input into the multi-layered stacked ST-LSGRU, it goes through two convolutional layers for downsampling;

[0101] The output is passed through two convolutional layers to restore the original resolution; each downsampling step reduces the side length of the original image by half, and the calculation method is as follows:

[0102] N = (M - F + 2P) / S + 1

[0103] Where M is the side length of the input image, F is the kernel size, P is the padding pixels, S is the stride, and N is the side length of the downsampled image;

[0104] Initialize the parameters of the radar echo extrapolation model, set the input sequence length input_size to 10, the output sequence length to 10, the batch size to 2, the learning rate lr to 0.001, the hidden layer convolution kernel size to 5, the number of convolution kernels n to 128, and the maximum number of training iterations iter to 10000.

[0105] Read the training set and perform initial training. Save the model training parameters once every 500 iterations. After 2500 training iterations, save the model training parameters 5 times.

[0106] Read the validation set, load the five model training parameters saved in step 4-2 respectively, obtain the prediction results, and save the results to the local disk;

[0107] Compare five different prediction results to see if they are predicting in a better direction. Based on the prediction results, adjust the hyperparameters to achieve the optimal effect.

[0108] Repeat steps 4 and 5 to adjust the model hyperparameters until the model can make stable predictions; read the test set, train the radar extrapolation model, and save the model parameters after training.

[0109] Load the model parameters saved in step 5-3, read the test set, and the prediction results are as follows: Figure 4 As shown, precipitation is obtained through Zr transformation:

[0110] Future precipitation can be obtained through Zr transformation:

[0111] Z = aR b

[0112] Here, a and b are characteristic constants, which may vary depending on the region and radar station.

[0113] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product implemented 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. The solutions in the embodiments of this application can be implemented in various computer languages, such as the object-oriented programming language Java and the interpreted scripting language JavaScript.

[0114] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0115] 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.

[0116] 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.

[0117] Although preferred embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application.

[0118] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.

Claims

1. A short-term precipitation prediction method based on GRU and LSTM spatiotemporal fusion, characterized in that, The method comprises the following steps: S1, reading radar data, filtering, denoising, cleaning and enhancing the data, and obtaining the processed radar data tensor; S2, dividing the radar data tensor obtained in step S1 into a training set, a validation set and a test set, and performing a slicing operation on the three data sets to divide them into input sequences and prediction sequences, and after normalization, the data format is unified; S3, constructing an ST-LSGRU basic module, stacking it optimally to form a radar echo extrapolation model; step S3 specifically comprises the following steps: S3-1: Constructing the ST-LSGRU basic module, if the module is located in the first layer, that is, when l = 1, the input state x is the tensor converted by the radar echo map input at the current time; if the module is not the first layer unit, that is, l > 1, the hidden state output at t time is transmitted to the input state of the module; the ST-LSGRU first transmits the input state and the hidden state output at t-1 time by the last layer of the unit to the reset gate and the update gate through two different convolution filters, and the formula is as follows: and the update gate , the formula is as follows: wherein denotes a convolution, denotes a Sigmoid function, ; Input state and spatiotemporal memory The input ST-LSGRU is obtained by three different convolution filters , the forget gate , the input gate and the input modulation gate ; the unimportant features of the past time in the time memory M are forgotten through the forget gate and the Hadamard product operation; the input gate and the input modulation gate are added element by element to realize the update of the features in the memory, and the spatiotemporal memory is obtained, and the formula is as follows: wherein denotes a dot product operation of matrices, denotes a hyperbolic tangent activation function, expressed by the formula ； Then a convolutional filter is used to process the input state. Hidden state And the updated spacetime memory Perform convolution operations to obtain new information. ;Will By resetting the door With the update gate Update and get The formula is expressed as follows: wherein denotes a convolution; S3-2: stack the ST-LSGRU module in a residual manner for 4 layers, and make a residual connection every other layer; embed multiple rough feature fast transmission units between the 1st layer and the 2nd layer of the stacked structure; form a rough feature fast transmission bus in the time dimension; after the action of two different convolution filters W p and W s , the state P t and S t are obtained; then the update of the rough feature state is realized through Hadamard product and element-wise addition operation, and the output of the current unit at the current moment is obtained; the calculation mode is as follows: ； S4, initializing the parameters of the radar echo extrapolation model established in step S3, setting the hyperparameters, and inputting the training set divided in step S2 to train the radar echo extrapolation model; obtaining the model training parameters; S5, inputting the validation set divided in step S2 into the radar echo extrapolation model for training, verifying the prediction result trend of the radar echo extrapolation model after a certain number of training, adjusting the hyperparameters to make the radar echo extrapolation model reach a stable gradient update direction, then using the training set divided in step S2 to continue training the radar echo extrapolation model, and after training, saving the radar echo extrapolation model parameters; S6, loading the radar echo extrapolation model parameters saved in step S5 and freezing them, inputting the test set into the trained radar echo extrapolation model, and obtaining the precipitation prediction result through Z-R transformation.

2. The short-term precipitation prediction method based on GRU and LSTM spatiotemporal fusion according to claim 1, characterized in that, Step S1 specifically comprises the following steps: S1-1: reading radar data, summing the global information of the radar data, calculating the number of radar pixels, evenly distributing the sum to each pixel, retaining data samples with an average value greater than 0.01, and satisfying the condition that the average value of 20 consecutive sequences is greater than 0.01; S1-2: cleaning the filtered data, setting the threshold to 0-70, determining the abnormal value as not between the threshold, setting the part less than 0 to 0, and setting the part greater than 70 to 70; S1-3: denoising the cleaned radar data, identifying outlier pixel values, setting them to 0 to ensure data quality and improve the prediction accuracy of data-driven deep learning models; S1-4: data enhancement: center cropping the radar data, performing format unification to obtain radar data with a resolution of 400*400, then performing random flipping, rotation and chroma adjustment operations on each sequence to obtain enhanced radar data.

3. The short-term precipitation prediction method based on GRU and LSTM spatiotemporal fusion according to claim 1, characterized in that, Step S2 specifically comprises the following steps: S2-1: dividing the radar data into a training set, a validation set and a test set; S2-2: dividing the training set, the validation set and the test set into multiple sequences using a sliding window method, with a window size of 20, a sequence length of 20, and a unit length time interval of 6 minutes, each sequence corresponding to 2 hours of radar data, the first ten being input sequences and the last ten being true labels; S2-3: mapping the radar data in the channel dimension to 1, normalizing it to 0-1 grayscale data, and then unifying the format of the normalized radar data, the normalization calculation method being: wherein, denotes the normalized data, denotes the original radar data, and denotes the maximum and minimum value in the radar data, respectively.

4. The short-term precipitation prediction method based on GRU and LSTM spatiotemporal fusion according to claim 1, characterized in that, Step S4 specifically comprises the following steps: S4-1: initialize the parameters of the radar echo extrapolation model, set the input sequence length input_size to 10, the output sequence length to 10, the batch size batch_size to 2, the learning rate lr to 0.001, the convolution kernel size of the hidden layer to 5, the number of convolution kernels n to 128, and the maximum number of iterations iter for training to 10000; S4-2: read the training set and perform preliminary training, save the radar echo extrapolation model training parameters once every 500 iterations, and save the radar echo extrapolation model training parameters 5 times after 2500 training.

5. The short-term precipitation prediction method based on GRU and LSTM spatiotemporal fusion according to claim 1, characterized in that, The adjustment of the hyperparameters in step S5 specifically includes the following steps: S5-1: read the validation set, load the radar echo extrapolation model training parameters saved in step S4 respectively, obtain the prediction results, and save the results in the local disk; S5-2: based on the change of mean square error loss, determine whether the model predicts in the direction of reducing the mean square error loss, and adjust the hyperparameters according to the prediction results; S5-3: repeat steps S4 and S5, adjust the model hyperparameters to make the model reach a state where the mean square error loss function tends to converge during training; read the validation set, train the radar extrapolation model, and save the model parameters after training.

6. The short-term precipitation prediction method based on GRU and LSTM spatiotemporal fusion according to claim 1, characterized in that, In step S6, the future rainfall is obtained through Z-R transformation, which is specifically: loading the model parameters saved in step S5, reading the test set, and predicting the results through Z-R transformation to obtain the precipitation, which can be expressed by the following formula: In the formula, a is the first characteristic constant, b is the second characteristic constant, Z is the radar reflectivity, and R is the rainfall intensity.

7. A short-term precipitation forecasting device implemented using the method according to any one of claims 1 to 6, characterized in that, The short-term precipitation prediction device comprises: A radar data acquisition and processing module is configured to acquire radar data, filter, denoise, clean and enhance the data, and then divide the data into a training set, a validation set and a test set, and perform slicing operation on the three data sets to divide them into input sequences and prediction sequences, and normalize and unify the data format; A radar echo extrapolation module is configured to receive the data acquired by the radar data acquisition and processing module, and obtain the precipitation prediction result through Z-R transformation.

8. A short-term precipitation prediction terminal device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, The processor executes the computer program to implement the method of any one of claims 1 to 6.

9. A computer-readable storage medium storing a computer program, the computer program comprising instructions that, when executed by a computer, cause the computer to perform the method of any one of claims 1 to 8. The computer program is executed by the processor to implement the method of any one of claims 1 to 6.

Images