Extreme rainfall prediction model, method and equipment based on artificial intelligence and medium
Through an extreme rainfall prediction model based on artificial intelligence, the potential coding sub-model, diffusion sub-model and conditional U-Net sub-model are used to solve the uncertainty and deviation problems of extreme rainfall prediction in the prior art, and more accurate and reliable predictions are achieved.
Patent Information
- Application Number
- CN202411793757.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-12-09
- Publication Date
- 2025-06-24
AI Technical Summary
The prior art has uncertainty and bias in extreme rainfall predictions, and deep learning models are difficult to build a complete probability model covering all predictors and results, resulting in insufficient extreme value estimation and lack of uncertainty quantification.
An extreme rainfall prediction model based on artificial intelligence is proposed, including a potential coding sub-model, a diffusion sub-model and a conditional U-Net sub-model. The rainfall samples are encoded into a potential representation through the potential coding sub-model. The diffusion sub-model performs diffusion and denoising operations on the potential representation, and the conditional U-Net sub-model uses environmental conditional data to perform the output of conditional distribution.
The model is able to capture complex target distributions, provide accurate predictions of extreme rainfall, solves the problems of insufficient extreme estimation and lack of uncertainty quantification in deep learning models, and improves the reliability and integrity of the prediction.
Smart Images
Figure CN120197651A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to an extreme rainfall prediction model, method, device and medium based on artificial intelligence, and belongs to the technical field of rainfall prediction. Background Art
[0002] Rainfall prediction, especially accurate prediction of extreme rainfall events, is of great significance to key fields such as flood control and disaster reduction, optimal allocation of water resources and agricultural planning, and is directly related to the stability and development of social economy.
[0003] Numerical weather prediction models can effectively simulate and predict climate variables such as pressure, temperature and humidity under the atmospheric circulation pattern. However, due to resolution limitations, numerical weather prediction models cannot directly simulate small-scale processes such as convection and cloud microphysics. Therefore, parametric methods are needed to approximate the effects of these processes. And parametric methods are usually based on empirical formulas and simplified physical assumptions, resulting in uncertainties and biases in rainfall simulation at the sub-grid scale in numerical weather prediction models. Therefore, there is limited inherent predictability in extreme rainfall prediction dominated by meteorological processes at the sub-grid scale.
[0004] Deep learning models such as convolutional neural networks and long short-term memory networks can learn data with high accuracy, reliability and integrity, and simulate physical processes at the sub-grid scale through strong complex non-linear dynamics processing capabilities. However, it is difficult to construct a complete probability model covering all predictors and prediction results for such deterministic deep learning models. Therefore, there are problems such as insufficient extreme value estimation, "fuzzy" estimation and lack of uncertainty quantification, which limit the extreme rainfall prediction ability of the models. Summary of the Invention
[0005] In order to solve the problems existing in the above-mentioned prior art, the present invention proposes an extreme rainfall prediction model, method, device and medium based on artificial intelligence.
[0006] The technical solution of the present invention is as follows:
[0007] On the one hand, the present invention proposes an extreme rainfall prediction model based on artificial intelligence, including:
[0008] A latent encoding sub-model, a diffusion sub-model and a conditional U-Net sub-model;
[0009] The latent encoding sub-model includes a first encoder and a first decoder. The first encoder is used to input rainfall samples and encode the rainfall samples into latent representations with a quasi-Gaussian distribution;
[0010] The conditional U-Net sub-model is used to input environmental condition data corresponding to the rainfall samples and output the conditional distribution corresponding to the rainfall samples;
[0011] The diffusion sub-model is used to perform diffusion operations on the latent representation with a quasi-Gaussian distribution, convert it into standard Gaussian distribution noise, and receive the conditional distribution corresponding to the rainfall samples. It uses the conditional distribution to perform denoising operations on the standard Gaussian distribution noise and converts the standard Gaussian distribution noise into output features that approximate the original latent representation;
[0012] The first decoder is used to decode the output features of the diffusion sub-model into rainfall prediction data.
[0013] As a preferred embodiment, in the step of using the conditional distribution to perform denoising operations on the standard Gaussian distribution noise and converting the standard Gaussian distribution noise into output features that approximate the original latent representation, a classifier-free guidance is used to approximate the latent representation as the output feature. The classifier-free guidance is expressed as:
[0014]
[0015]
[0016] where, is the classifier-free guidance, z t is the latent representation at time step t; x is the conditional information, corresponding to the environmental condition data of the rainfall samples;
[0017] A guidance weight is introduced to obtain:
[0018]
[0019] where, is the classifier-free guidance after introducing the guidance weight, w is the guidance weight, and 1≥w≥0, ∈ NN (z t ) is the output of the unconditional neural network NN according to the input latent representation z t ; is the conditional network NN c According to the input latent representation z t and the output of the conditional information x;
[0020] By training the unconditional neural network NN and the conditional network NN c to approximate the latent representation z t .
[0021] As a preferred embodiment, the conditional network NN c is implemented based on the denoising U-Net model. The denoising U-Net model includes:
[0022] Three convolutional layers, each convolutional layer contains a residual stack;
[0023] The residual stack is composed of multiple residual convolutional blocks, and each residual convolutional block contains a noise function that embeds temporal noise;
[0024] The noise function generates noise based on the current time step and the latent representation that has undergone max pooling and convolutional processing;
[0025] The output of each residual convolutional block is added to the input by addition.
[0026] As a preferred embodiment, the method for the diffusion sub-model to perform diffusion operations is specifically as follows:
[0027] Noise is gradually added to the latent representation at each time step t in the entire time period T, and the added noise is determined according to the current time step.
[0028] As a preferred embodiment, in the latent coding sub-model:
[0029] The first encoder includes three convolutional layers and a residual stack;
[0030] The first decoder includes one convolutional layer, a residual stack, and two transposed convolutional layers;
[0031] The residual stack includes multiple residual blocks, and each residual block contains two convolutional layers and a ReLU activation function.
[0032] As a preferred embodiment, the conditional U-Net sub-model includes a second encoder and a second decoder. The output of the second encoder is connected to the input of the second decoder, and a residual connection is made between the output of the second encoder and the output of the second decoder;
[0033] The second encoder includes multiple residual convolutional blocks, and each residual convolutional block includes two convolutional layers, followed by two max pooling operations and a GELU activation function;
[0034] The second decoder includes one residual convolutional block and a max pooling layer.
[0035] As a preferred embodiment, the environmental condition data includes static input data and dynamic input data, where:
[0036] The dynamic input data includes: stability, geopotential height, specific humidity, horizontal wind speed, and two-meter temperature at different Kia levels;
[0037] The static input data includes DEM digital elevation information;
[0038] Both the dynamic input data and the static input data are normalized.
[0039] On the other hand, the present invention also proposes an extreme rainfall prediction method based on artificial intelligence, including the following steps:
[0040] Construct an extreme rainfall prediction model based on artificial intelligence as described in any embodiment of the present invention;
[0041] Obtain rainfall samples to construct a training data set, and train the extreme rainfall prediction model based on artificial intelligence through the training data set;
[0042] Perform rainfall prediction through the trained extreme rainfall prediction model based on artificial intelligence.
[0043] In another aspect, the present invention also provides an electronic device, including a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the program, it implements the extreme rainfall prediction method based on artificial intelligence as described in any embodiment of the present invention.
[0044] In another aspect, the present invention also provides a computer-readable storage medium, on which a computer program is stored. When the program is executed by a processor, it implements the extreme rainfall prediction method based on artificial intelligence as described in any embodiment of the present invention.
[0045] The additional aspects and advantages of the present invention will be clarified in the following description, and some of them will be obvious from the description, or can be understood by practicing the present invention. In addition, the various aspects and advantages of the present invention can be realized and obtained through the method steps and combinations specifically pointed out in the appended claims. Description of the Drawings
[0046] Figure 1 It is a schematic diagram of the model structure of Embodiment 1 of the present invention;
[0047] Figure 2 It is a schematic diagram of the method flow of Embodiment 2 of the present invention. Detailed Embodiments
[0048] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.
[0049] It should be understood that the step numbers used in the text are only for convenience of description and do not limit the execution order of the steps.
[0050] It should be understood that the terms used in the specification of the present invention are merely for the purpose of describing specific embodiments and are not intended to limit the present invention. As used in the specification of the present invention and the appended claims, unless the context clearly indicates otherwise, the singular forms "a", "an" and "the" are intended to include the plural forms.
[0051] The terms "comprising" and "including" indicate the presence of the described features, wholes, steps, operations, elements and / or components, but do not preclude the presence or addition of one or more other features, wholes, steps, operations, elements, components and / or their combinations.
[0052] The term "and / or" refers to any combination and all possible combinations of one or more of the associated listed items, and includes these combinations.
[0053] Example 1:
[0054] See Figure 1 , this embodiment provides an extreme rainfall prediction model based on artificial intelligence, including:
[0055] A potential encoding sub-model, a diffusion sub-model and a conditional U-Net sub-model. Among them:
[0056] The potential encoding sub-model includes a first encoder and a first decoder. The first encoder is used to input rainfall samples and encode the rainfall samples into a latent representation Z with a quasi-Gaussian distribution;
[0057] The conditional U-Net sub-model is used to input the environmental condition data corresponding to the rainfall samples and output the conditional distribution corresponding to the rainfall samples.
[0058] The diffusion sub-model includes a diffusion process and a denoising process; in the diffusion process, a diffusion operation is performed on the latent representation Z with a quasi-Gaussian distribution, and after T time steps, it is converted into noise z close to the standard Gaussian distribution t , and the way of the noise addition process is controllable; in the denoising process, it receives the conditional distribution corresponding to the rainfall samples, and performs a denoising (regression) operation on the standard Gaussian distribution noise using the conditional distribution, and converts the standard Gaussian distribution noise into output features approaching the original latent representation.
[0059] The first decoder is used to decode the output features of the diffusion sub-model into rainfall prediction data.
[0060] In this embodiment, the potential encoding sub-model is composed of a first encoder and a first decoder, where:
[0061] The first encoder includes three convolutional layers and a residual stack;
[0062] The first decoder includes a convolutional layer, a residual stack, and two transposed convolutional layers;
[0063] The residual stack includes multiple residual blocks, and each residual block contains two convolutional layers and a ReLU activation function.
[0064] Specifically, the first encoder takes the rainfall sample X(b×1×M×N), where b represents the number of rainfall samples, and 1×M×N represents the size of the input image of the rainfall sample; after extracting features through two convolutional layers and a residual block, a feature map of the rainfall sample X is formed, and finally, the logarithm of the mean and variance is calculated through a fully connected layer to form the latent representation of X
[0065] The latent representation Z has a quasi-Gaussian distribution and is regularized by D KL (P(μ(z),σ 2 (z))||N(0,I)), where
[0066] In this embodiment, the diffusion process of the diffusion sub-model can be represented by where z t represents the latent representation at time step t, β is the variance schedule that controls the amount of noise added at each step, and I is the identity matrix.
[0067] For the denoising process of the diffusion sub-model, by using the subsequent conditional distribution to invert the Gaussian process to approximate P(z0), specifically: P θ (z t-1 |z t ) = N(z t-1 ; μ θ (z t ), ∑ θ (z t ))), where ∑ θ and μ θ are learned parameters that guide the conversion from noise to the latent representation.
[0068] Furthermore, a classifier-free guidance is used to approximate the latent representation as the output feature, and the classifier-free guidance is represented as:
[0069]
[0070] where, is the classifier-free guidance, z t is the latent representation at time step t; x is the conditional information, corresponding to the environmental condition data of the rainfall sample;
[0071] Furthermore, a guidance weight is introduced to obtain:
[0072]
[0073] Among them, is the classifier-free guidance after introducing the guidance weight, w is the guidance weight, and 1≥w≥0, ∈ NN (z t ) is the output of the unconditional neural network NN according to the input latent representation z t . is the conditional network NN c according to the input latent representation z t and the output of the conditional information x;
[0074] By means of implicit training, the unconditional neural network NN and the conditional network NN are trained c to approximate respectively and to approximate the latent representation z with conditional information as x t .
[0075] In this embodiment, the conditional network NN c is implemented based on a denoising U-Net model, which consists of three convolutional layers. Each convolutional layer contains a residual stack, and the residual stack is composed of multiple residual convolutional blocks. Each residual convolutional block contains a noise function that embeds time, and generates noise according to the current time step and the latent representation of the features further processed by max pooling (MaxPool2d) and 1x1 convolution. The output of each residual block is added to the input by addition to form a residual connection. The latent representation after being processed by the residual stack is used for subsequent denoising steps. After each residual convolutional block, the features in the encoder are combined with the features in the decoder using a residual connection to retain more detailed information. Finally, the latent representation after being processed by the residual stack is transformed into a noise estimate for the current time step using a transposed convolution, with the same size as the latent representation.
[0076] Among them, the noise function generates noise according to the current time step and the latent representation.
[0079] Based on the above implementation, the guidance weight w can finely control the influence of the conditional information, and thus balance the model to generate unconditional inputs or conditional outputs. When w = 0, for unconditional information input, the unconditional output of the unconditional neural network NN can be achieved through the denoising U-Net model; w = 1 means that the influence of the conditional information x is calculated by NN c computes, w < 1 means that the influence of the conditional information x is suppressed, and w > 1 means that the influence of the conditional information x is enhanced. This mechanism enables the model to flexibly adjust its sensitivity to conditional information, thereby promoting the generation of data that is more consistent with the required conditions.
[0080] Based on the above solution, the extreme rainfall prediction model in this embodiment maximizes the probability assigned to the observed data (conditional information) through a classifier-free guidance, enabling it to capture complex target distributions and accurately generate extreme events. This method solves the problem of extreme underestimation that plagues deterministic deep learning models. In addition, the denoising process in this embodiment is progressive step by step, capable of capturing high-frequency details and gradually optimizing the generated data, making the output rainfall distribution map clearer and solving the problem of fuzzy estimation that may occur in other deep learning models.
[0081] In addition, the extreme rainfall prediction model introduces a probability element in each step, capable of naturally capturing the uncertainty existing in the data. Through this process, the model can generate diverse samples reflecting the potential uncertainty in the data distribution, providing reliable uncertainty quantification.
[0082] In this embodiment, the conditional U-Net sub-model consists of two parts: a second encoder (downsampling path) and a second decoder (upsampling path). The output of the second encoder is connected to the input of the second decoder, and a residual connection is made between the output of the second encoder and the output of the second decoder.
[0083] The second encoder part consists of multiple residual convolutional blocks, each block containing two convolutional layers. The convolutional layers use two max pooling (MaxPool2d) operations and the GELU activation function to process the output of the encoder, reducing the dimensionality of the feature vector and passing it to the second decoder.
[0084] The second decoder part also consists of a residual convolutional block and a max pooling layer, increasing the spatial dimension of the feature map by using a combination of average pooling and transposed convolution.
[0085] In each upsampling step, the corresponding feature map from the second encoder is combined with the feature map of the second decoder through a residual connection to restore the detailed information of the image. The second decoder part gradually restores the spatial dimension of the feature map through transposed convolution.
[0086] In this embodiment, the environmental condition data includes static input data and dynamic input data, where:
[0087] The dynamic input data includes: temperature (T), geopotential height (GH), specific humidity (Q), horizontal wind speed (U and V components), and 2-meter temperature (T2M) at three different pressure levels (1000, 850, and 500 hPa). The dynamic input data is normalized by subtracting the field mean value and dividing by the field standard deviation to ensure consistent scaling between different variables.
[0088] The static input data includes DEM digital elevation information; the static input data is interpolated to match the requirements of the model, and at the same time, the static input data is also normalized using the same method as the dynamic input data.
[0089] Embodiment 2:
[0090] Refer to Figure 2 , this embodiment proposes an extreme rainfall prediction method based on artificial intelligence, including the following steps:
[0091] S100. Construct an extreme rainfall prediction model based on artificial intelligence as described in any embodiment of the present invention; define a latent encoding sub-model, a conditional U-Net sub-model, and a diffusion sub-model respectively according to the extreme rainfall prediction model; at the same time, set configuration parameters such as learning rate, batch size, model path, etc. for model training and testing.
[0092] S200. Obtain rainfall samples to construct a training data set, set up a model training loop, define a loss function and an optimizer, and perform model training and verification through steps such as loading the training data set, initializing the model, updating model parameters, and saving the best model. During the training process, use the validation set to evaluate the model performance, select the best model according to the validation loss, and then fine-tune the model parameters to improve the prediction accuracy.
[0093] S300. Use the trained model to perform rainfall prediction on new meteorological data, compare and analyze the prediction results in combination with actual rainfall data and other prediction models, and evaluate the model performance. Deploy the trained model to the production environment for actual rainfall prediction tasks.
[0094] In step S200, historical rainfall samples are extracted from the NetCDF file as the data set, and the data set is divided into a training set (90%) and a validation set (10%). Subsequently, random sampling and data augmentation are performed on the training set.
[0095] When training, the loss function of the latent encoding sub-model is set as:
[0096]
[0097] where x is the rainfall sample, is the rainfall sample reconstructed from the latent representation z, μ(z) and σ(z) are the mean and standard deviation of the Gaussian distribution of the latent representation z sampled from, and β is a hyperparameter used to balance the reconstruction loss and the KL divergence. To improve the reconstruction quality, β = 0 can be set in the initial stage of training and β = 0.01 in the later stage.
[0098] The loss function of the conditional U-Net model is set as:
[0099] L(θ) = MSE(z, f(c; θ));
[0100] Where z represents the latent representation of rainfall, and f(c; θ) represents the latent representation of rainfall restored based on conditional information c and model parameters θ.
[0101] The loss function of the diffusion model is set as:
[0102]
[0103] Where t represents the time step, ∈ is the added noise following a normal distribution, ∈ θ (z t , t|z′) is the output of the denoising model.
[0104] In step S300, deterministic metrics such as root mean square error (RMSE), critical success index (CSI), and Pearson correlation coefficient (PCC), as well as probabilistic metrics such as continuous ranked probability score (CRPS) and coverage rate (CR) are used to evaluate the model performance. Among them, the root mean square error (RMSE) is used to measure the accuracy of the prediction model and is defined as n is the number of observed values, y i is the actual value, is the predicted value; the critical success index CSI is also used to measure the accuracy of the prediction and is defined as (TP represents the number of true positive results, indicating an accurate prediction of an event or occurrence; FP represents the number of false positive results, referring to the situation where an event is predicted but does not occur; FN represents the number of false negative results, meaning that an event occurs but is not correctly predicted.). The Pearson correlation coefficient PCC measures the linear correlation between two variables and is defined as: Where x i and y i are the actual values; and are the means. The continuous ranked probability score CRPS is a measure of the accuracy of probability prediction and is defined as: Where F(y) is the cumulative prediction distribution, H is the Heaviside step function, and y obs is the observed value. The coverage rate (CR) is used to evaluate the performance of probability forecasts and is defined as CR = number of observations within the prediction interval / total number of observations.
[0105] Example 3:
[0106] This example proposes an electronic device, including a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the program, it implements the extreme rainfall prediction method based on artificial intelligence as described in any embodiment of the present invention.
[0107] Example 4:
[0108] This embodiment provides a computer-readable storage medium storing a computer program, which when executed by a processor, implements the extreme rainfall prediction method based on artificial intelligence as described in any embodiment of the present invention.
[0109] In the embodiments of the present application, "at least one" means one or more, and "a plurality" means two or more. "And / or" describes the association relationship of associated objects, indicating that three relationships may exist. For example, A and / or B may represent the cases of A existing alone, A and B existing simultaneously, and B existing alone. Where A and B may be singular or plural. The character " / " generally represents an "or" relationship between the associated objects before and after. "At least one of the following" and its similar expressions refer to any combination of these items, including any combination of single or plural items. For example, at least one of a, b, and c may represent: a, b, c, a and b, a and c, b and c, or a and b and c, where a, b, and c may be single or multiple.
[0110] Those of ordinary skill in the art can realize that the units and algorithm steps described in the embodiments disclosed herein can be implemented by electronic hardware, computer software, or a combination of both. Whether these functions are executed in hardware or software depends on the specific application and design constraints of the technical solution. Skilled professionals can use different methods to implement the described functions for each specific application, but such implementation should not be considered to exceed the scope of the present application.
[0111] Those skilled in the art can clearly understand that for the convenience and brevity of description, the specific working processes of the systems, devices, and units described above can refer to the corresponding processes in the foregoing method embodiments, and will not be repeated here.
[0112] In several embodiments provided by the present application, if any function is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application, in essence, or the part that contributes to the prior art, or a part of this technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present application. The foregoing storage medium includes: various media such as USB flash drives, mobile hard disks, read-only memories (hereinafter referred to as ROMs), random access memories (hereinafter referred to as RAMs), magnetic disks, or optical discs that can store program codes.
[0113] The above are only embodiments of the present invention, and do not thereby limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made by using the content of the specification and drawings of the present invention, or directly or indirectly applied in other related technical fields, shall equally be included within the patent protection scope of the present invention.
Claims
1. An artificial intelligence-based extreme rainfall prediction model, characterized in that: include: Latent encoding sub-model, diffusion sub-model and conditional U-Net sub-model; The latent coding sub-model includes a first encoder and a first decoder, wherein the first encoder is used to input a rainfall sample and encode the rainfall sample into a latent representation having a quasi-Gaussian distribution; The conditional U-Net sub-model is used to input the environmental condition data corresponding to the rainfall samples and output the conditional distribution of the corresponding rainfall samples; The diffusion sub-model is used to perform diffusion operation on the potential representation with quasi-Gaussian distribution, convert it into standard Gaussian distribution noise, and receive the conditional distribution of the corresponding rainfall sample. The conditional distribution is used to perform denoising operation on the standard Gaussian distribution noise, and the standard Gaussian distribution noise is converted into output features close to the original potential representation. The first decoder is used to decode the output features of the diffusion sub-model into rainfall prediction data.
2. The artificial intelligence-based extreme rainfall prediction model according to claim 1, characterized in that: In the step of using conditional distribution to perform denoising operation on standard Gaussian distribution noise and converting the standard Gaussian distribution noise into output features approximating the original potential representation, a classifier-free guide is used to approximate the potential representation as the output feature, and the classifier-free guide is expressed as: in, is a classifier-free guide, z t is the potential representation of time step t; x is the conditional information, corresponding to the environmental condition data of the rainfall sample; Introducing the guide weight, we get: in, is the classifier-free guide after the introduction of the guide weight, w is the guide weight, and 1≥w≥0,∈ NN (z t ) is the unconditional neural network NN according to the potential representation z of the input t The output, For the conditional network NN c According to the potential representation z of the input t And the output of conditional information x; By training the unconditional neural network NN and the conditional network NN c The approximation conditional information is the potential representation z of x t .
3. The artificial intelligence-based extreme rainfall prediction model according to claim 2, characterized in that: The conditional network NN c It is implemented based on a denoising U-Net model, which includes: Three convolutional layers, each containing a residual stack; The residual stack consists of multiple residual convolution blocks, and each residual convolution block contains a noise function embedded with temporal noise; The noise function generates noise based on the current time step and the potential representation after maximum pooling and convolution; The output of each residual convolution block is added to the input via addition.
4. The artificial intelligence-based extreme rainfall prediction model according to claim 1, characterized in that: The method for the diffusion sub-model to perform diffusion operation is specifically as follows: Noise is added to the latent representation step by step at each time step t in the entire time period T, and the added noise is determined according to the current time step.
5. The artificial intelligence-based extreme rainfall prediction model according to claim 1, characterized in that: In the latent encoding submodel: The first encoder includes three convolutional layers and a residual stack; The first decoder includes a convolutional layer, a residual stack, and two transposed convolutional layers; The residual stack consists of multiple residual blocks, each of which contains two convolutional layers and a ReLU activation function.
6. The artificial intelligence-based extreme rainfall prediction model according to claim 1, characterized in that: The conditional U-Net sub-model includes a second encoder and a second decoder, the output of the second encoder is connected to the input of the second decoder, and the output of the second encoder and the output of the second decoder are residually connected; The second encoder includes a plurality of residual convolution blocks, each residual convolution block includes two convolution layers, and the convolution layers are subsequently connected to two maximum pooling operations and a GELU activation function; The second decoder includes a residual convolution block and a maximum pooling layer.
7. The artificial intelligence-based extreme rainfall prediction model according to claim 1, characterized in that: The environmental condition data includes static input data and dynamic input data, wherein: Dynamic input data include: stability, geopotential height, specific humidity, horizontal wind speed, and two-meter temperature at different KIA levels; Static input data includes DEM digital elevation information; Both dynamic input data and static input data are normalized.
8. An extreme rainfall prediction method based on artificial intelligence, characterized in that: The following steps are involved: Constructing an artificial intelligence-based extreme rainfall prediction model as described in any one of claims 1 to 7; Obtain rainfall samples to construct a training data set, and use the training data set to train the AI-based extreme rainfall prediction model; Rainfall prediction is performed through the trained AI-based extreme rainfall prediction model.
9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, the extreme rainfall prediction method based on artificial intelligence as described in claim 8 is implemented.
10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the extreme rainfall prediction method based on artificial intelligence as described in claim 8 is implemented.