A method and equipment for forecasting heavy precipitation events using prior diffusion networks based on multi-source meteorological data.

Abstract

This invention provides a method and device for forecasting heavy precipitation events using a priori diffusion network based on multi-source meteorological data. The method includes acquiring time-series multi-source meteorological data and performing data filtering and cleaning; constructing a multi-source meteorological data representation model, which is implemented through a three-layer encoding / decoding network composed of gated units WEGiU guided by the wave equation; inputting the multi-source meteorological data into the joint representation model to generate representation results; inputting the representation results into a potential diffusion network, constructing a representation-diffusion integrated loss with the representation network, and training the network; and inputting the multi-source meteorological data into the trained representation and diffusion networks to forecast heavy precipitation events in the next b hours based on past a hours of multi-source meteorological data. This invention effectively integrates the characteristics of multi-source meteorological data and improves the accuracy of heavy precipitation event prediction.

Description

Technical Field

[0001] This invention belongs to the field of heavy precipitation forecasting technology, and particularly relates to a method and equipment for forecasting heavy precipitation processes using a priori diffusion network based on multi-source meteorological data. Background Technology

[0002] Heavy rainfall has a wide-ranging impact on human society and the natural environment, potentially leading to disasters such as urban flooding, landslides, and flash floods, posing a significant threat to industrial and agricultural production, people's lives and property, and urban development. Therefore, accurate forecasting of heavy rainfall is crucial for disaster prevention and mitigation. However, the diversity and complexity of meteorological conditions make heavy rainfall events difficult to predict, making the improvement of forecast accuracy a key research focus in the meteorological field.

[0003] Traditional heavy rainfall forecasting is accomplished through numerical weather prediction models, typically relying on single data sources such as ground observations and radar data. This approach fails to effectively utilize diverse meteorological data and consequently cannot capture the characteristics of complex meteorological systems. However, with the rapid development of artificial intelligence technology, particularly the outstanding performance of deep learning networks in data representation and prediction, an increasing number of heavy rainfall forecasting methods are based on deep learning. Summary of the Invention

[0004] To address the current issues of insufficient accuracy in heavy precipitation forecasting and the characterization of multi-source meteorological data, this invention provides a priori diffusion network forecasting method for heavy precipitation processes based on multi-source meteorological data. This method can effectively integrate the characteristics of multi-source meteorological data and improve the accuracy of heavy precipitation process forecasting.

[0005] The prior diffusion network forecasting method for heavy precipitation processes based on multi-source meteorological data designed in this invention includes the following steps:

[0006] S1: Acquire time-series multi-source meteorological data and preprocess the data, such as data filtering and data cleaning;

[0007] S2: Construct a multi-source meteorological data representation model, and use the multi-source meteorological data representation model to generate representation results of multi-source meteorological data; the multi-source meteorological data representation model is a three-layer encoding and decoding network structure composed of gated units WEGiU guided by the wave equation;

[0008] S3: Input the characterization results of multi-source meteorological data into the potential diffusion network, construct a characterization-diffusion integrated loss with the characterization network, and train it.

[0009] S4: Input multi-source meteorological data into the trained representation and diffusion network to predict heavy precipitation events in the next b hours based on multi-source meteorological data from the past a hours.

[0010] Preferably, the specific process of S1 is as follows:

[0011] S11: Select precipitation, radar reflectivity, wind speed, temperature, and dew point temperature data from the collected multi-source meteorological data.

[0012] S12: For regions with a few missing or abnormal data, an image interpolation algorithm based on the Navier-Stokes equation is used to estimate the value of pixels in the region based on information about pixels surrounding the missing pixels.

[0013] S13: Process the time-series multi-source meteorological data to a 1-hour time resolution, and filter the multi-source meteorological data according to the standard of more than 16 milliliters of precipitation per hour; trim the multi-source meteorological data to a set size according to the spatial location of heavy precipitation events to obtain multi-source meteorological data of the past a hours and precipitation data of the future b hours, and divide the training set, validation set and test set according to the set ratio.

[0014] Preferably, the encoding and decoding process of the multi-source meteorological data representation model includes:

[0015] S21: Perform the first layer of the encoding network, and convert the hidden state H... 1,0 Meteorological data is input into the WEGiU module for state updates based on the wave equation;

[0016] S22: After the state update calculation guided by the wave equation is completed, input A according to the calculation process of the Convolutional Gated Recurrent Unit (ConvGRU). l and the hidden state H″ updated based on the wave equation l-1 The new hidden state H is calculated. l After multiple iterations, representation extraction is achieved, resulting in S obtained from the first layer of encoding. 1,t-a+1 ,…,S 1,t ;

[0017] S23: Perform the second layer of the encoding network, and transfer the hidden state H 2,0 S obtained from the first layer encoding 1,t-a+1 ,…,S 1,t The data is passed to the WEGiU module, where the same wave equation-based state update and representation extraction operations as in steps S21 and S22 are performed to obtain the second-layer encoding result S. 2,t-a+1 ,…,S 2,t ;

[0018] S24: Perform the third layer of the encoding network, and transfer the hidden state H 3,0 S obtained from the second layer of encoding 2,t-a+1 ,…,S 2,t The data is passed to the WEGiU module, where the same wave equation-based state update and representation extraction operations as in steps S21 and S22 are performed to obtain the third-layer encoding result S. 3,t-a+1,…,S 3,t , denoted as E3;

[0019] S25: Perform decoding operation; the first layer of the decoding network will decode the S in the third encoding network. 3,t As the initial hidden state input to the WEGiU module, S 3,t Deconstructed into The second layer of the decoding network will As input, S 2,t As the initial hidden state, the reconstruction yields... The third layer of the decoding network will As input, S 1,t As the initial hidden state, multi-source meteorological data were reconstructed.

[0020] S26: Construct a multi-level reconstruction loss and use three mean squared error losses to ensure that the reconstruction results of three different levels are consistent with the input or features in terms of time, space and attributes.

[0021] Preferably, the specific processing steps of the diffusion model include:

[0022] Add noise to the precipitation data for the next b hours;

[0023] The noisy data and the multi-source meteorological data representation results are concatenated according to the first dimension and input into the UNet network for denoising.

[0024] Preferably, a representation-diffusion integrated loss is constructed and trained, with the loss function as follows:

[0025] Loss=w1L rec +w2L diffu

[0026] Where w1 and w2 are the weights controlling the two losses, L rec L is the multi-level reconstruction loss function in the multi-source meteorological data representation model. diffu This is the loss function in the diffusion model.

[0027] Preferably, the specific implementation process of S4 is as follows:

[0028] Based on the precipitation data of the past hour 'a' to be predicted, a preliminary precipitation data for the next hour 'b' is predicted using a pre-trained convolutional long short-term memory (convLSTM) network. Will By inputting the trained multi-source meteorological data representation network and diffusion network, predicted b-hour precipitation data are obtained.

[0029] Based on the same inventive concept, this invention also designs a system for implementing a priori diffusion network forecasting method for heavy precipitation processes based on multi-source meteorological data, including a data preprocessing module and a precipitation prediction module;

[0030] The data preprocessing module filters and cleans the acquired time-series multi-source meteorological data;

[0031] The precipitation prediction module includes a data representation module and a diffusion network module. It uses multi-source meteorological data inputs to train a multi-source meteorological data representation network and a diffusion network to predict the heavy precipitation process in the next b hours based on multi-source meteorological data from the past a hours.

[0032] The data representation module is a three-layer encoding and decoding network structure composed of gated units WEGiU guided by the wave equation. It uses the multi-source meteorological data representation model to generate representation results of multi-source meteorological data. The representation results of multi-source meteorological data are input into the potential diffusion network, and a representation-diffusion integrated loss is constructed with the representation network and trained to obtain the diffusion network.

[0033] Based on the same inventive concept, the present invention also designs an electronic device, comprising:

[0034] One or more processors;

[0035] Storage device for storing one or more programs;

[0036] When one or more programs are executed by the one or more processors, the one or more processors implement a method such as a priori diffusion network forecasting method for heavy precipitation processes based on multi-source meteorological data.

[0037] Based on the same inventive concept, the present invention also designs a computer-readable medium having a computer program stored thereon, which, when executed by a processor, implements a method for forecasting heavy precipitation processes using a priori diffusion network based on multi-source meteorological data.

[0038] Compared with the prior art, the advantages of the present invention are as follows:

[0039] 1) Multi-source meteorological data were used to improve the robustness of the model and enhance its generalization ability.

[0040] 2) A gated convolutional network based on the wave equation for physical state updates was designed, which effectively integrates the characteristics of multi-source meteorological data and improves the model's ability to characterize atmospheric dynamic changes and the spatiotemporal correlation of heavy precipitation.

[0041] 3) The diffusion model based on the characterization results can effectively capture the spatiotemporal dynamics of heavy precipitation processes, and improve the precision and accuracy of heavy precipitation prediction. Attached Figure Description

[0042] Figure 1 This is a flowchart illustrating the prior diffusion network forecasting method for heavy precipitation processes based on multi-source meteorological data according to the present invention.

[0043] Figure 2 This is a schematic diagram of the characterization network structure of the present invention.

[0044] Figure 3 This is a schematic diagram of the diffusion network of the present invention. Detailed Implementation

[0045] The technical solutions of the present invention will be described in detail below with reference to the accompanying drawings.

[0046] Example 1

[0047] The purpose of this invention is to provide a priori diffusion network forecasting method for heavy precipitation processes based on multi-source meteorological data, such as... Figure 1 As shown, it includes the following steps:

[0048] S1: Acquire time-series multi-source meteorological data and perform data filtering and cleaning;

[0049] S2: Construct a multi-source meteorological data representation model. The representation model is implemented through a three-layer encoding and decoding network composed of wave equation-guided gated convolutional units (WEGiU). Multi-source meteorological data is input into the joint representation model to generate the representation results of the multi-source meteorological data.

[0050] S3: Input the characterization results of multi-source meteorological data into the potential diffusion network, construct a characterization-diffusion integrated loss with the characterization network, and train it.

[0051] S4: Input the multi-source meteorological data into the trained representation and diffusion network to predict the heavy precipitation process in the next b hours based on the multi-source meteorological data from the past a hours. In this embodiment, a is 6 and b is 3.

[0052] The specific implementation process of step S1 is as follows:

[0053] S11: Select precipitation, radar reflectivity, wind speed, temperature, and dew point temperature data from the collected multi-source meteorological data.

[0054] S12: For regions with a few missing or anomalous data, an image interpolation algorithm based on the Navier-Stokes equation is used to estimate the values ​​of pixels in these regions based on information about the pixels surrounding the missing pixels.

[0055] S13: In this embodiment, the time-series multi-source meteorological data is processed to a 1-hour time resolution, and the multi-source meteorological data is screened according to the standard of more than 16 milliliters of precipitation per hour; the multi-source meteorological data is cropped to a set size according to the spatial location of the heavy precipitation event, such as a 128×128 grid with a spatial resolution of 1km, to obtain the multi-source meteorological data of the past a hours and the precipitation data of the future b hours, and the training set, validation set and test set are divided according to the set ratio.

[0056] Step S2 is as follows Figure 2 As shown, the specific implementation process includes:

[0057] S21: Perform the first layer coding network, such as Figure 2 Hidden state H 1,0 Meteorological data is input into the WEGiU module to obtain S 1,t The process. Specifically, it involves processing n from the past a hours. w ×n h k-dimensional time-series multi-source meteorological data from a grid And the hidden state H initialized as an all-zero matrix 1,0 Input the WEGiU module to perform a state update based on the wave equation. The state update is calculated using the following formula.

[0058] H″ l-1 =H l-1 +Φ(H l-1 )

[0059] Where l belongs to 0~a, and is a certain time step in the time series multi-source meteorological data. H″ l-1 H represents the hidden state after a physical update guided by the wave equation. l-1 This represents the hidden state at time step l-1. The initial hidden state H0 is an all-zero matrix, Φ(H l-1 The wave equation (☐) represents a commonly used linear partial differential equation. The formula for calculating the wave equation is as follows:

[0060]

[0061] Among them, c ij Let x and y represent the coefficients of a partial derivative of order q, where x and y represent H. l-1 The x and y directions of the grid points. In this invention, q is 1, and c... ij Taking 1, the wave equation simplifies to... The remaining partial derivatives of the wave equation are expressed as residual terms Res.

[0062] S22: After the state update calculation guided by the wave equation is completed, you can input A according to the calculation process of the Convolutional Gated Recurrent Unit (ConvGRU). l and the hidden state H″ updated based on the wave equation l-1 The new hidden state H is calculated. l After multiple iterations, representation extraction is achieved, resulting in S obtained from the first layer of encoding. 1，t-a+1 S 1，t The calculation process of ConvGRU is as follows:

[0063] Z l =σ(conv) az (A l )+conv hz (H) l-1 ))

[0064] R l =σ(conv) az (A l )+conv hz (H) l-1 ))

[0065]

[0066] Among them, H l This indicates the hidden state after the WEGiU update, A l This represents multi-source meteorological data at time step l. H″ l-1 It is the hidden state after physical update based on the wave equation, H′ l-1 It is in H″ l-1 The hidden state is updated regularly based on the previous state. l It is an update gate that determines which previous memories to retain in the current time step. R l It's a reset gate; it determines how new inputs are incorporated into the current memory. The subscripts a, z, r, h of conv(·) correspond to A, Z, R, H, and also indicate the input and output types of conv(·), for example, conv az This indicates that the input of conv(·) is A. l The output terminal is Z. l σ is the Sigmoid function, and ι is the leakyReLU function. It represents the Hadamardi (or Hadama) stack.

[0067] S23: Perform the second layer coding network, such as Figure 2 H 2，0 Passing the S data to the WEGiU module 2，tThe process is an approximate repetition of S21 and S22. The S21 obtained from the first encoding network is then used... 1，t-a+1 S 1，t As input, replace A in S21 t-a+1 A t The remaining parts are calculated according to the S2l and S22 calculation process. The second-layer encoding result S is obtained. 2，t-a+1 S 2，t .

[0068] S24: Perform the third-layer coding network, such as Figure 2 H 3,0 Passing the S data to the WEGiU module 3，t The process is an approximate repetition of S21 and S22. The S obtained from the second encoding network... 2，t-a+1 S 2，t As input, replace A in S21 t-a+1 A t The remaining parts are calculated according to the S2l and S22 calculation process. The third-layer encoding result S is obtained. 3，t-a+1 S 3，t The S obtained from the third encoding network 3，t It is denoted as E3 and can be used as input to subsequent potential diffusion networks.

[0069] S25: Decode the network. To construct the multi-layer reconstruction loss in the representation model, based on the obtained representation results S... 3，t Based on this, the representation results also need to be decoded. The decoding network is based on S 1，t S 2t S 3，t As input, meteorological data is reconstructed through a three-layer network. The process. The first layer of the decoding network will decode the S in the third encoding network. 3，t As the initial hidden state input to the WEGiU module, S 3，t Deconstructed into The second layer of the decoding network will As input, S 2,t As the initial hidden state, the reconstruction yields... The third layer of the decoding network will As input, S 1,t As the initial hidden state, multi-source meteorological data were reconstructed.

[0070] S26: Constructing Multi-Level Reconstruction Loss. In this embodiment, three Mean Square Error (MSE) losses are used to ensure that the reconstruction results at three different levels maintain spatiotemporal and attribute consistency with the input or features. The formula for calculating the multi-level reconstruction loss is as follows.

[0071]

[0072] Where w r0 ,w r1 ,w r2 By controlling the weights of reconstruction loss at different levels and comparing model training results, the optimal weights are selected in this implementation example. r0 =w r1 =w r2 =1 / 3.

[0073] Step S3 further involves the following process: Figure 3 As shown.

[0074] S31: Add noise to the precipitation data R0 for the next b hours to obtain the data R after t noise additions. t The noise addition process is as follows:

[0075]

[0076] Among them, R t R represents the data after adding noise t times. t-1 This represents the data after adding noise t-1 times. Let I represent the standard Gaussian distribution, where I is the identity matrix and β is the standard Gaussian distribution. t This represents the diffusion rate of the t-th noise addition.

[0077] It can also be expressed as Based on this derivation, we can obtain R. t It can be directly calculated from R0, as shown in the following formula.

[0078]

[0079] Where, α t =1-β t ,

[0080] S32: Add noise to the data R t The multi-source meteorological data representation result E3 obtained from S24 is concatenated along the first dimension and then denoised. The denoising process involves calculating q(R). t-1 |R t The process is as follows, and the calculation formula is as follows:

[0081]

[0082] in

[0083] S33: In S32, It cannot be calculated directly and needs to be predicted using a neural network. The noisy data R... tThe multi-source meteorological data representation result E3 obtained from S24 is concatenated and input into the UNet network for denoising. The loss function is as follows:

[0084]

[0085] ∈ θ (R t ,t) is from the noisy result R t Predict a standard Gaussian noise, and This constitutes the loss function.

[0086] S34: Construct the representation-diffusion integrated loss and train it. The loss function is as follows, where w1 and w2 are the weights that control the two losses.

[0087] Loss=w1L rec +w2L diffu

[0088] Step S4 further includes:

[0089] S41: Based on the precipitation data of the past 'a' hours to be predicted, a preliminary precipitation data for the next 'b' hours is predicted using a pre-trained Convolutional Long Short-Term Memory (ConvLSTM) network. Will After T rounds of noise addition, the noise is obtained. The multi-source meteorological data from the past 'a' hours to be predicted is input into the trained representation network to obtain representation E3. After being concatenated with E3, it is input into the trained UNet network and subjected to T reverse diffusion processes to obtain noise N. Subtract N to obtain the predicted hourly precipitation data for hour b. Example 2

[0090] Based on the same inventive concept, this embodiment also designs a system for implementing a priori diffusion network forecasting method for heavy precipitation processes based on multi-source meteorological data, including a data preprocessing module and a precipitation prediction module;

[0091] The data preprocessing module filters and cleans the acquired time-series multi-source meteorological data;

[0092] The precipitation prediction module includes a data representation module and a diffusion network module. It uses multi-source meteorological data inputs to train a multi-source meteorological data representation network and a diffusion network to predict the heavy precipitation process in the next b hours based on multi-source meteorological data from the past a hours.

[0093] The data representation module is a three-layer encoding and decoding network structure composed of gated units WEGiU guided by the wave equation. It uses the multi-source meteorological data representation model to generate representation results of multi-source meteorological data. The representation results of multi-source meteorological data are input into the potential diffusion network, and a representation-diffusion integrated loss is constructed with the representation network and trained to obtain the diffusion network.

[0094] Since the device described in Embodiment 2 of the present invention is the system used to implement the prior diffusion network forecasting method for heavy precipitation processes based on multi-source meteorological data in Embodiment 1 of the present invention, and its specific process is the same as that in Embodiment 1 of the present invention, it will not be repeated here.

[0095] Example 3

[0096] Based on the same inventive concept, the present invention also provides an electronic device, including one or more processors; a storage device for storing one or more programs; and when the one or more programs are executed by the one or more processors, the one or more processors implement the method described in Embodiment 1.

[0097] Since the device described in Embodiment 3 of this invention is the electronic device used in implementing the prior diffusion network forecasting method for heavy precipitation processes based on multi-source meteorological data in Embodiment 1 of this invention, those skilled in the art can understand the specific structure and variations of this electronic device based on the method described in Embodiment 1 of this invention, and therefore will not be described again here. All electronic devices used in any method of this invention fall within the scope of protection of this invention.

[0098] Example 4

[0099] Based on the same inventive concept, the present invention also provides a computer-readable medium having a computer program stored thereon, which, when executed by a processor, implements the method described in Embodiment 1.

[0100] Since the device described in Embodiment 4 of this invention is a computer-readable medium used to implement the a priori diffusion network forecasting method for heavy precipitation processes based on multi-source meteorological data in Embodiment 1 of this invention, those skilled in the art can understand the specific structure and variations of this electronic device based on the method described in Embodiment 1 of this invention, and therefore will not be described again here. All electronic devices used in any method of this invention fall within the scope of protection of this invention.

[0101] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for forecasting heavy precipitation processes using a priori diffusion network based on multi-source meteorological data, characterized in that, Includes the following steps: S1: Acquire time-series multi-source meteorological data and preprocess the data; S2: Construct a multi-source meteorological data representation model, and use the multi-source meteorological data representation model to generate representation results of multi-source meteorological data; the multi-source meteorological data representation model is a three-layer encoding and decoding network structure composed of gated units WEGiU guided by the wave equation; The encoding and decoding process of the multi-source meteorological data representation model includes: S21: Perform the first layer of the encoding network, and hide the state. Meteorological data is input into the WEGiU module for state updates based on the wave equation; S22: After the state update calculation guided by the wave equation is completed, the input is processed according to the calculation process of the Convolutional Gated Recurrent Unit (ConvGRU). and hidden states updated based on the wave equation The new hidden state is calculated. After multiple iterations, representation extraction is achieved, resulting in the first layer of encoding. ; S23: Perform the second layer of the encoding network to hide the state. The first layer of encoding The data is passed to the WEGiU module, where the same wave equation-based state update and representation extraction operations as in steps S21 and S22 are performed to obtain the second-layer encoding result. ; S24: Perform the third layer of the encoding network to hide the state. The second layer of encoding is obtained The data is passed to the WEGiU module, where the same wave equation-based state update and representation extraction operations as in steps S21 and S22 are performed to obtain the third-layer encoding result. , recorded as ; S25: Perform decoding operation; the first layer of the decoding network will decode the data from the third encoding network. As the initial hidden state input to the WEGiU module, Deconstructed into The second layer of the decoding network will As input, As the initial hidden state, the reconstruction yields... The third layer of the decoding network will As input, As the initial hidden state, multi-source meteorological data were reconstructed. ; S26: Construct a multi-layer reconstruction loss to ensure that the reconstruction results of 3 different levels are consistent with the input or features in terms of time, space and attributes; S3: Input the characterization results of multi-source meteorological data into the potential diffusion network, construct a characterization-diffusion integrated loss with the characterization network, and train it. S4: Input multi-source meteorological data into the trained multi-source meteorological data representation network and diffusion network to predict the heavy precipitation process in the next b hours based on the multi-source meteorological data of the past a hours.

2. The method for forecasting heavy precipitation processes using a priori diffusion network based on multi-source meteorological data according to claim 1, characterized in that: The specific process of S1 is as follows: S11: Select precipitation, radar reflectivity, wind speed, temperature, and dew point temperature data from the collected multi-source meteorological data. S12: For regions with a few missing or abnormal data, an image interpolation algorithm based on the Navier-Stokes equation is used to estimate the value of pixels in the region based on information about pixels surrounding the missing pixels. S13: Process the time-series multi-source meteorological data to a 1-hour time resolution, and filter the multi-source meteorological data according to the standard of more than 16 milliliters of precipitation per hour; trim the multi-source meteorological data to a set size according to the spatial location of heavy precipitation events to obtain multi-source meteorological data of the past a hours and precipitation data of the future b hours, and divide the training set, validation set and test set according to the set ratio.

3. The method for forecasting heavy precipitation processes using a priori diffusion network based on multi-source meteorological data according to claim 1, characterized in that: Using three mean squared error losses ensures that the reconstruction results at three different levels remain consistent with the input or features in terms of time, space, and attributes.

4. The method for forecasting heavy precipitation processes using a priori diffusion network based on multi-source meteorological data according to claim 1, characterized in that: The specific processing steps of the diffusion model include: Add noise to the precipitation data for the next b hours; The noisy data and the multi-source meteorological data representation results are concatenated according to the first dimension and input into the UNet network for denoising.

5. The method for forecasting heavy precipitation processes using a priori diffusion network based on multi-source meteorological data according to claim 4, characterized in that: Construct and train a representation-diffusion integrated loss function, as follows: in To control the weights of the two losses, This refers to the multi-level reconstruction loss function in a multi-source meteorological data characterization model. This is the loss function in the diffusion model.

6. The method for forecasting heavy precipitation processes using a priori diffusion network based on multi-source meteorological data according to claim 1, characterized in that: The specific implementation process of S4 is as follows: Based on the precipitation data of the past hour a to be predicted, a preliminary future precipitation prediction is made using a pre-trained convolutional long short-term memory network. hourly precipitation data ,Will Input the trained multi-source meteorological data representation network and diffusion network to obtain predictions. Hourly precipitation data .

7. A system for implementing the prior diffusion network forecasting method for heavy precipitation processes based on multi-source meteorological data as described in any one of claims 1-6, characterized in that: include, Data preprocessing module and precipitation prediction module; The data preprocessing module filters and cleans the acquired time-series multi-source meteorological data; The precipitation prediction module includes a data representation module and a diffusion network module. It uses multi-source meteorological data inputs to train a multi-source meteorological data representation network and a diffusion network to predict the heavy precipitation process in the next b hours based on multi-source meteorological data from the past a hours. The data representation module is a three-layer encoding and decoding network structure composed of gated units WEGiU guided by the wave equation, which generates the representation results of multi-source meteorological data using the multi-source meteorological data representation model. The characterization results of multi-source meteorological data are input into the potential diffusion network, and a characterization-diffusion integrated loss is constructed with the characterization network and trained to obtain the diffusion network. The encoding and decoding process of the multi-source meteorological data representation model includes: S21: Perform the first layer of the encoding network, and hide the state. Meteorological data is input into the WEGiU module for state updates based on the wave equation; S22: After the state update calculation guided by the wave equation is completed, the input is processed according to the calculation process of the Convolutional Gated Recurrent Unit (ConvGRU). and hidden states updated based on the wave equation The new hidden state is calculated. After multiple iterations, representation extraction is achieved, resulting in the first layer of encoding. ; S23: Perform the second layer of the encoding network to hide the state. The first layer of encoding The data is passed to the WEGiU module, where the same wave equation-based state update and representation extraction operations as in steps S21 and S22 are performed to obtain the second-layer encoding result. ; S24: Perform the third layer of the encoding network to hide the state. The second layer of encoding is obtained The data is passed to the WEGiU module, where the same wave equation-based state update and representation extraction operations as in steps S21 and S22 are performed to obtain the third-layer encoding result. , recorded as ; S25: Perform decoding operation; the first layer of the decoding network will decode the data from the third encoding network. As the initial hidden state input to the WEGiU module, Deconstructed into The second layer of the decoding network will As input, As the initial hidden state, the reconstruction yields... The third layer of the decoding network will As input, As the initial hidden state, multi-source meteorological data were reconstructed. ; S26: Construct a multi-layer reconstruction loss to ensure that the reconstruction results of 3 different levels are consistent with the input or features in terms of time, space and attributes.

8. An electronic device, characterized in that, include: One or more processors; Storage device for storing one or more programs; When one or more programs are executed by the one or more processors, the one or more processors implement the prior diffusion network forecasting method for heavy precipitation processes based on multi-source meteorological data as described in any one of claims 1-6.

9. A computer-readable medium having a computer program stored thereon, characterized in that: When the program is executed by the processor, it implements the prior diffusion network forecasting method for heavy precipitation processes based on multi-source meteorological data as described in any one of claims 1-6.

Images