A regional precipitation prediction method

Abstract

The present application relates to a kind of regional precipitation prediction methods, comprising: obtaining EC precipitation forecast data;Construct sunny and rainy prediction model, EC precipitation forecast data is input sunny and rainy prediction model, obtain sunny and rainy forecast, wherein sunny and rainy prediction model is obtained by precipitation forecast training set training;Obtain EC numerical prediction data, wherein EC numerical prediction data includes height layer wind field, potential height field and temperature and humidity field data;Construct topographic enhancement strong precipitation model, according to EC numerical prediction data and topographic enhancement strong precipitation model, obtain topographic enhancement forecast;Sunny and rainy forecast and topographic enhancement forecast are fused, obtain precipitation forecast field, obtain final precipitation forecast by precipitation forecast field.The present application solves the limitation of existing forecast model in accuracy, real-time and resolution, significantly improves the accuracy of sunny and rainy and rainstorm forecast.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of precipitation prediction, in particular to a regional precipitation prediction method. BACKGROUND

[0002] As the most common and important part of daily weather service, the accuracy of the rain forecast is directly related to the public life and safety, and is also one of the core standards for evaluating the ability of weather forecast. However, the existing rain forecast method has significant shortcomings, especially in predicting fine precipitation events, which is manifested as frequent false reports of light precipitation. Taking Baise City as an example, the accuracy of urban forecast in 2021 decreased by 1.5% compared with the average level of the past three years. This downward trend is largely due to the uneven distribution of warm area in the flood season, which reduces the accuracy of rain forecast.

[0003] In the field of heavy rain forecast, this task has always been extremely challenging, mainly because of the influence of multiple complex climate systems, such as local convective heavy rain on the edge of the subtropical high, monsoon and typhoon induced precipitation. These factors together cause the difficulty of prediction.

[0004] In recent years, the rapid development of artificial intelligence technology in the field of meteorology, especially in the application of heavy rain research and prediction, has shown great potential. Using deep learning models such as U-net rain forecast model, the ability of feature extraction and data analysis has been improved, which provides a new technical approach to improve the accuracy of prediction. However, although these models have made progress in some aspects, they still lack in integrating terrain effects for precipitation prediction. In particular, the terrain enhancement heavy rain model has not yet achieved the expected prediction accuracy and real-time requirements in dealing with the complexity of terrain and its relationship with precipitation events. Therefore, a regional precipitation prediction method is proposed. SUMMARY

[0005] The purpose of the present application is to provide a regional precipitation prediction method, which aims to solve the limitations of existing prediction models in accuracy, real-time performance and resolution, and significantly improve the accuracy of rain and heavy rain forecast.

[0006] To achieve the above purpose, the present application provides a regional precipitation prediction method, comprising:

[0007] obtaining EC precipitation forecast data;

[0008] constructing a rain forecast model, inputting the EC precipitation forecast data into the rain forecast model, and obtaining rain forecast, wherein the rain forecast model is obtained by training a precipitation forecast training set;

[0009] obtaining EC numerical prediction data, wherein the EC numerical prediction data comprises height layer wind field, geopotential height field and temperature and humidity field data;

[0010] constructing a terrain enhancement heavy rainfall model, and obtaining terrain enhancement prediction according to the EC numerical prediction data and the terrain enhancement heavy rainfall model;

[0011] fusing the sunny and rainy prediction and the terrain enhancement prediction to obtain a precipitation prediction field, and obtaining a final precipitation prediction through the precipitation prediction field.

[0012] Optionally, the sunny and rainy prediction model is constructed by converting each convolution layer in a U-Net network into a Bayesian convolution layer.

[0013] Optionally, the sunny and rainy prediction model is trained through a precipitation prediction training set, including:

[0014] collecting historical EC precipitation prediction data under an ECMWF prediction mode, processing the historical EC precipitation prediction data into grid data of a preset resolution to obtain input data;

[0015] obtaining multi-source real-time precipitation data, processing the multi-source real-time precipitation data into grid data of the preset resolution to obtain label data;

[0016] constructing the precipitation prediction training set through the input data and the label data, and training the sunny and rainy prediction model through the precipitation prediction training set.

[0017] Optionally, obtaining the sunny and rainy prediction includes:

[0018] inputting the precipitation prediction data into the sunny and rainy prediction model, and an encoder of the sunny and rainy prediction model extracts features of the precipitation prediction data to obtain a feature map;

[0019] inputting the feature map into a decoder of the sunny and rainy prediction model, and a Bayesian convolution layer in the decoder assigns weights to the feature map to obtain the sunny and rainy prediction, wherein the sunny and rainy prediction is a sunny and rainy distribution grid field.

[0020] Optionally, the Bayesian convolution layer in the decoder assigns weights to the feature map to obtain the sunny and rainy prediction, including:

[0021] the Bayesian convolution layer in the decoder assigns weights to the feature map to output a plurality of uncertainty information, averages the uncertainty information to obtain a precipitation average probability;

[0022] taking the precipitation average probability as a value of each grid point in the sunny and rainy prediction, which is a probability of precipitation at a corresponding position, to obtain the sunny and rainy prediction.

[0023] Optionally, the terrain enhancement prediction comprises:

[0024] According to the temperature field data, the height layer wind field data and the geopotential height field, the terrain precipitation is determined to be generated;

[0025] After the terrain precipitation is determined to be generated, the water vapor condensation in the atmosphere is calculated based on the humidity field data, and the terrain enhancement prediction is obtained, wherein the terrain enhancement prediction is a heavy rainfall field, and the value of each grid point in the heavy rainfall field is the expected rainfall of the terrain.

[0026] Optionally, the method for determining the terrain precipitation to be generated is that whether the terrain precipitation is generated is determined by calculating a wet Froude number;

[0027] The calculation method of the wet Froude number is:

[0028]

[0029] Wherein, U is the average speed of the airflow; N is the Brunt-Vaisala frequency, g is the acceleration of gravity, and θ is the potential temperature, is the change of the potential temperature with the height, and h is the characteristic height of the terrain.

[0030] Optionally, the heavy rainfall field comprises:

[0031]

[0032] Wherein, P is the condensation amount, that is, the rainfall, R is the condensation rate per unit height, z h is the saturated wet layer height, and z0 is the ground height.

[0033] Optionally, the precipitation prediction field is obtained by fusing the sunny and rainy prediction and the terrain enhancement prediction, which comprises:

[0034] The terrain enhancement prediction is superimposed with the EC precipitation prediction data to obtain a final terrain enhancement prediction;

[0035] The sunny and rainy prediction is multiplied with the final terrain enhancement prediction at the grid points to obtain the precipitation prediction field;

[0036] The final rainfall prediction is obtained through the precipitation prediction field.

[0037] The beneficial effects of the present application are: by introducing the Bayesian U-Net model, the model introduces Bayesian inference on the basis of the classic U-Net architecture, thereby more effectively processing the uncertainty in the input data, and using ECWMF fine grid data for optimization and training to improve the accuracy of rain and shine prediction. The model can more accurately estimate the possibility of precipitation under different weather conditions through the probabilistic weights of the Bayesian neural network, especially in capturing subtle weather changes. Compared with traditional deterministic models, the Bayesian U-Net model can give an uncertainty assessment of the prediction, providing more information for decision makers. The quantification of this uncertainty reduces the sensitivity of the model to noise and outliers, thereby effectively reducing false positives and false negatives in rain and shine prediction, and greatly improving the prediction accuracy. In addition, the Bayesian method also allows the model to make reasonable inferences through prior knowledge in the case of data scarcity or change, further enhancing the adaptability and accuracy of the model under variable weather conditions.

[0038] The present application adopts a physical parameterization method, including the blocking and guiding effect of terrain on atmospheric flow field, the influence on air flow rising and sinking, and temperature and other factors. By accurately incorporating these physical processes into model calculations, the influence of terrain on precipitation distribution and intensity can be more realistically reproduced, improving the prediction accuracy of heavy rain and above precipitation events. The present application combines the Bayesian U-Net model with actual terrain data to improve the accuracy of precipitation prediction. BRIEF DESCRIPTION OF DRAWINGS

[0039] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed in the embodiments. Obviously, the drawings described below are only some embodiments of the present application, and other drawings can also be obtained by those skilled in the art without creative labor.

[0040] Figure 1 The overall flowchart of a regional precipitation prediction method according to an embodiment of the present application is shown in Figure 1.

[0041] Figure 2 The Bayesian convolution layer weight distribution diagram according to an embodiment of the present application is shown in Figure 2.

[0042] Figure 3 The comparison diagram of precipitation revision effects from 20:00 on September 9, 2019 to 20:00 on September 10, 2019 according to an embodiment of the present application is shown in Figure 3, wherein (a) is the terrain revision prediction from 20:00 on September 9, 2019 to 20:00 on September 10, 2019, and (b) is the EC original prediction from 20:00 on September 9, 2019 to 20:00 on September 10, 2019.

[0043] Figure 4A comparison chart of the corrected effects of the precipitation of the embodiment of the application from 08:00 on April 28, 2023 to 08:00 on April 29, 2023, wherein (a) is the terrain corrected forecast from 08:00 on April 28, 2023 to 08:00 on April 29, 2023, and (b) is the EC original forecast from 08:00 on April 28, 2023 to 08:00 on April 29, 2023.

[0044] Figure 5 A schematic diagram of the overall network framework of the Bayesian U-Net of the embodiment of the application. DETAILED DESCRIPTION

[0045] The technical solutions in the embodiments of the application will be described clearly and completely below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only part of the embodiments of the application, rather than all the embodiments of the application. Based on the embodiments in the application, all other embodiments obtained by a person of ordinary skill in the art without creative work fall within the protection scope of the application.

[0046] In order to make the above-mentioned objects, features and advantages of the application more obvious and easy to understand, the application will be further described in detail below with reference to the drawings and specific embodiments.

[0047] The embodiment provides a regional precipitation prediction method, as shown in the following formula: Figure 1 The method comprises the following steps:

[0048] EC precipitation forecast data is obtained;

[0049] A sunny and rainy day prediction model is constructed, the EC precipitation forecast data is input into the sunny and rainy day prediction model, and sunny and rainy day prediction is obtained, wherein the sunny and rainy day prediction model is obtained by training a precipitation prediction training set;

[0050] EC numerical prediction data is obtained, wherein the EC numerical prediction data comprises height layer wind field data, potential height field data and temperature and humidity field data;

[0051] A terrain enhancement and heavy precipitation model is constructed, and terrain amplification prediction is obtained according to the EC numerical prediction data and the terrain enhancement and heavy precipitation model;

[0052] The sunny and rainy day prediction and the terrain amplification prediction are fused to obtain a precipitation prediction field, and the final precipitation prediction is obtained through the precipitation prediction field.

[0053] The construction of the sunny and rainy day prediction model is specifically as follows:

[0054] The Bayesian U-Net model still retains the basic architecture of the traditional U-Net in terms of structure, which is composed of two main parts: encoding (downsampling) and decoding (upsampling). Based on the traditional U-Net, each convolutional layer in the present invention is converted into a Bayesian convolutional layer. These Bayesian convolutional layers not only learn the features of the data, but also learn the uncertainty of these features, so as to be able to give not only the predicted value but also the uncertainty evaluation of the prediction in the model's prediction.

[0055] The encoder of the Bayesian U-Net model extracts the features of the input data, obtains the feature maps, and fully utilizes the feature maps from the same resolution in the encoding process in the decoding process, while these feature maps are spliced with the feature maps in the decoding process through cross-layer connection. This not only enables the model to more effectively utilize the shallow features, but also provides important information about the uncertainty of the data during prediction, greatly improving the performance and reliability of the model in dealing with complex or uncertain data.

[0056] Therefore, the Bayesian U-Net model not only inherits the advantages of traditional U-Net in image segmentation, but also enhances the model's ability to handle uncertainty by introducing Bayesian inference, making it more robust and adaptable in various application scenarios, especially in fields that require accurate handling of uncertain information, such as medical image analysis and weather data processing.

[0057] In the Bayesian U-Net model, the introduction of Bayesian convolutional layers (Bayesian neural networks) not only improves the model's ability to handle data uncertainty, but also enables the model to more effectively identify and focus on key features in the input data. The core advantage of Bayesian convolutional layers lies in their ability to evaluate the uncertainty of data features while learning these features, thereby providing important uncertainty information for the model's prediction.

[0058] In the decoding step, the Bayesian convolutional layer assigns weights to the feature maps from the encoding step, which are based not only on the importance of the features but also on the uncertainty of the features. This method makes the model more intelligent and accurate when integrating features from the encoder, giving priority to those features that are most important and reliable for the current task, rather than relying solely on traditional feature importance evaluation.

[0059] The Bayesian U-Net model not only inherits the advantages of traditional U-Net in image segmentation, but also significantly improves the model's ability to identify and utilize key features through the introduction of Bayesian convolutional layers. Especially when facing data with high uncertainty, through the Bayesian convolutional layer, the model not only can more effectively distinguish between small precipitation and non-precipitation areas, but also can quantify the uncertainty of the prediction, thereby greatly reducing the false alarm and missed alarm situations. The structural diagram of the Bayesian U-Net model is as follows:Figure 2 As shown.

[0060] In solving this binary classification problem of sunny and rainy prediction, the binary cross-entropy loss is chosen, and the advantages of the Bayesian method are combined, so that the model can not only measure the difference between the predicted and actual values in the classification problem, but also consider the uncertainty of the prediction. This method helps guide the model to make more accurate and reliable predictions. At the same time, in order to ensure that the model has good generalization ability.

[0061] In terms of constructing the dataset, EC precipitation data from 2019 to 2022 is selected. The inverse distance weighting method is used to interpolate the site precipitation data, i.e. EC precipitation forecast is interpolated to 5km, generating a high-resolution training dataset matching the EC precipitation data. In the training dataset, more than 0.1mm is considered as rain (1), and less than 0.1mm is considered as sunny (0). In the face of the imbalance of sunny and rainy samples in the dataset, appropriate balancing strategies are adopted, i.e. when constructing the dataset, a constraint condition is used, i.e. the number of precipitation grid points greater than 0.1mm is more than 20% of the total number of grid points or there is a strong precipitation event (maximum value greater than 50mm), and sunny and rainy days are labeled. This ensures that both sunny and rainy weather conditions can obtain sufficient learning samples during model training, thereby significantly improving the prediction performance.

[0062] The Bayesian U-Net model in this embodiment takes the 3-hour precipitation data provided by EC (European Center for Medium-Range Weather Forecasts) as input, and uses 5km resolution multi-source real-time precipitation data as training labels to achieve accurate prediction of precipitation events through the following steps. The Bayesian U-Net model framework is as shown in Figure 5 :

[0063] Step 1: Data preprocessing

[0064] Input data: Obtain 3-hour historical precipitation prediction data provided by EC and perform necessary preprocessing such as data normalization, missing value processing, etc.

[0065] Label data: Integrate multi-source real-time precipitation data and convert it to 5km resolution grid data matching the model input as the true label in the training and validation process.

[0066] Step 2: Model construction

[0067] Construct the Bayesian U-Net model, which integrates the Bayesian inference mechanism on the basis of the traditional U-Net architecture.

[0068] Step 3: Model training and validation

[0069] The Bayesian U-Net model is trained using pre-processed historical EC precipitation data and precipitation data (label data);

[0070] During the training process, techniques such as cross-validation are used to avoid overfitting and ensure that the model has good generalization ability;

[0071] The model is verified using an independent test set to ensure the accuracy and reliability of the model's predictions.

[0072] Step four: uncertainty assessment

[0073] During the prediction process, the model not only provides prediction values, but also gives the corresponding uncertainty range, such as the posterior probability distribution to represent the uncertainty of the prediction result;

[0074] The assessment of uncertainty is of great significance for guiding actual meteorological decision-making, and helps to assess the risk and credibility of the prediction.

[0075] Step five: output and application

[0076] The EC precipitation prediction data is input into the sunny and rainy prediction model to obtain the sunny and rainy prediction, and the model outputs the 5km resolution precipitation prediction result, including the predicted precipitation and its uncertainty assessment; the output result can be directly used for meteorological prediction service, and more accurate and reliable precipitation information is provided for the public and relevant departments.

[0077] Obtain EC numerical prediction data, wherein the EC numerical prediction data includes high-level wind field, potential height field and temperature and humidity field data; construct a terrain enhancement strong precipitation model, and obtain terrain enhancement prediction according to the EC numerical prediction data and the terrain enhancement strong precipitation model.

[0078] The terrain enhancement strong precipitation model is constructed, which is specifically:

[0079] Traditional terrain precipitation estimation methods, such as the formula proposed by Smith (1979), although effective in many cases, show certain limitations in dealing with precipitation prediction under complex terrain conditions. Especially in the integral upper limit, precipitation efficiency and judgment criteria, there are obvious shortcomings, which are particularly obvious in dealing with strong precipitation prediction in specific areas.

[0080] The present application proposes to use the saturated wet layer height as the integral upper limit to more reasonably reflect the actual precipitation process. This improvement is based on a deeper understanding of water vapor condensation and vertical motion in the precipitation process.

[0081] Precipitation efficiency: the region is divided at a resolution of 5km, and the precipitation efficiency of different height terrains is re-evaluated and quantified. This method takes into account the influence of terrain height on the precipitation process, and more accurately simulates the influence of terrain on precipitation.

[0082] a. Wet Froude number calculation:

[0083] The wet Froude number (Fr) is a dimensionless number used to evaluate the flow pattern of fluid flow over an obstacle. In the context of orographic precipitation, it can be used to determine whether the airflow is strong enough to overcome the orographic obstacle and result in precipitation. The formula for calculating the wet Froude number is:

[0084]

[0085] where U is the average speed of the airflow (which can be obtained from the ECMWF model's height layer wind field data), N is the Brunt-Vaisala frequency, which represents atmospheric stability, and h is the characteristic height of the terrain.

[0086] b. Brunt-Vaisala frequency (N) calculation:

[0087] The Brunt-Vaisala frequency is an indicator of atmospheric stability, and its formula is:

[0088]

[0089] where g is the acceleration due to gravity, θ is the potential temperature (a meteorological quantity that represents temperature variation), is the rate of change of potential temperature with height.

[0090] c. Orographic precipitation determination:

[0091] When the calculated wet Froude number is greater than 1, it indicates that the airflow is strong enough to overcome the orographic obstacle and may result in orographic precipitation. Conversely, it is unlikely to produce orographic precipitation.

[0092] d. Condensation amount calculation:

[0093] If orographic precipitation is determined, the calculation of condensation amount can be based on the condensation of water vapor in the atmosphere. The formula for calculating the condensation amount can be:

[0094]

[0095] where P is the condensation amount, i.e., the precipitation amount, R is the condensation rate per unit height, z h and z0 are the lower and upper limits of the integral, respectively, and the upper limit is the saturation wet layer height mentioned earlier, and z0 is the ground height.

[0096]

[0097] where ρ w = r s ρ air , r s is the saturated mixing ratio, and ρ airFor air density, saturated water vapor density (p ws ) is replaced by actual air water vapor density (p w ), and a is the terrain slope, U(z) is the horizontal wind.

[0098] Judgment criteria: the dimensionless wet Froude number is introduced as a new judgment criterion for terrain precipitation. When it is greater than 1, it indicates that there is terrain precipitation. This criterion provides a new and more scientific method for terrain precipitation prediction.

[0099] This embodiment uses the upper-level wind field, potential height field and temperature field data predicted by the ECMWF model to calculate the wet Froude number, so as to determine whether terrain precipitation occurs. If it is determined that there is terrain precipitation, the condensation amount (the condensation of water vapor in the atmosphere is processed through the humidity field to obtain) and the upper limit of integration are calculated to obtain the terrain heavy precipitation amount, that is, the terrain enhancement prediction, wherein the obtained terrain heavy precipitation amount is a terrain heavy precipitation amount field with a resolution of 5Km.

[0100] Precipitation efficiency consideration: the precipitation efficiency of different terrains may be different, and the calculated precipitation amount can be adjusted according to the terrain characteristics.

[0101] The fusion of the sunny and rainy forecast and the heavy precipitation field is specifically:

[0102] In the complete terrain, AI fusion precipitation model strategy, combining the advantages of the Bayesian U-Net sunny and rainy forecast model and the terrain enhancement heavy precipitation model is a crucial step. This integration strategy aims to ensure that the model can provide comprehensive and accurate predictions under various precipitation conditions. To achieve the fusion of the two models, the terrain enhancement prediction and the sunny and rainy forecast are multiplied at the grid points to present a continuous and physically meaningful precipitation prediction field. Through this strategy, not only the accuracy of the prediction is improved, but also the spatial continuity of the prediction result is ensured. Before multiplying the terrain enhancement prediction and the sunny and rainy forecast at the grid points, the terrain enhancement prediction and the EC precipitation prediction data are superimposed.

[0103] (1) Bayesian U-Net sunny and rainy forecast model output:

[0104] The Bayesian U-Net model outputs a sunny and rainy forecast, which is a sunny and rainy distribution grid field presented in grid form. The value of each grid point represents the probability of precipitation at that location. The decoder of the Bayesian U-Net model calculates 30 uncertain values, which are averaged as output to achieve deterministic prediction. The output is the average probability of precipitation, and the output probability value is between 0 and 1. Finally, greater than 0.5 is 1 and less than 0.5 is 0, where 0 represents no precipitation and 1 represents precipitation.

[0105] (2) Terrain enhancement heavy precipitation model output:

[0106] The terrain-enhanced heavy precipitation model outputs a heavy precipitation field, also represented in grid form.

[0107] The value at each grid point indicates the expected precipitation amount under that terrain condition. The final terrain-enhanced prediction is obtained by superimposing the original EC 3-hour precipitation forecast.

[0108] (3) Grid multiplication technique:

[0109] The Bayesian U-Net rain and shine prediction model output is multiplied with the terrain-enhanced heavy precipitation model and EC precipitation forecast data superposition results to obtain the fusion output. That is, (terrain-enhanced prediction + EC precipitation forecast data) x rain and shine prediction = final rain prediction field.

[0110] (4) Final fusion output:

[0111] The fused output provides a comprehensive and accurate precipitation prediction field. This prediction field not only improves the prediction accuracy, but also maintains continuity in spatial distribution, enabling the model to effectively handle various precipitation and terrain conditions.

[0112] The method of the present application can significantly improve the prediction performance of the model for heavy precipitation events, especially in the prediction accuracy when the precipitation system and the terrain interact strongly. Table 1 shows the Bayesian U-Net model 24-hour precipitation prediction TS score from April to September 2023.

[0113] Table 1

[0114] Precipitation threshold (mm) TS score of the present model Original TS score of the control group Note Greater than 0.1 mm 0.7343 0.6021 Accuracy of light precipitation events significantly improved Greater than 10 mm 0.3645 0.3630 Performance of moderate precipitation events comparable Greater than 25 mm 0.2630 0.2533 Performance of moderate to heavy precipitation events slightly improved Greater than 50 mm 0.1757 0.1372 Accuracy of heavy precipitation events significantly improved Greater than 100 mm 0.0744 0.0543 Accuracy of extremely heavy precipitation events improved

[0115] The data in Table 1 shows the comparison of the Bayesian U-Net model of the present application and the terrain-enhanced heavy precipitation model in terms of TS score at different precipitation thresholds and the original EC numerical prediction 24-hour rainfall. The data shows that the model exhibits equal or better prediction ability than the existing method at all tested precipitation thresholds, especially in the prediction of light precipitation and heavy precipitation events.

[0116] Table 2 shows the comparison of precipitation correction effect from 20:00 on September 9, 2019 to 20:00 on September 10, 2019, and Table 2 corresponds to Figure 3 (a)-(b), which improves the TS score of heavy rain.

[0117] Table 2

[0118] Precipitation threshold (mm) TS score of the present model Original TS score of the control group Greater than 10 mm 0.58 0.57 Greater than 25 mm 0.38 0.31 Greater than 50 mm 0.19 0.16 Greater than 100 mm 0.04 0.03

[0119] Table 3 shows the comparison of precipitation from 08:00 on April 28, 2023 to 08:00 on April 29, 2023, and Table 3 corresponds to Figure 4(a)-(b) match, data show that the present application can improve the TS score of rainstorm.

[0120] Table 3

[0121]

[0122]

[0123] The above described embodiments are only to describe the preferred modes of the present application, and are not intended to limit the scope of the present application. Without departing from the design spirit of the present application, various modifications and improvements to the technical solutions of the present application made by those skilled in the art shall fall within the protection scope determined by the claims of the present application.

Claims

1. A method of regional precipitation prediction, characterized by, The method comprises: obtaining EC precipitation forecast data; constructing a sunny and rainy prediction model, inputting the EC precipitation forecast data into the sunny and rainy prediction model, and obtaining a sunny and rainy forecast, wherein the sunny and rainy prediction model is obtained by training a precipitation forecast training set, and the sunny and rainy prediction model is constructed by converting each convolutional layer in a U-Net network into a Bayesian convolutional layer; obtaining the sunny and rainy forecast comprises: inputting the precipitation forecast data into the sunny and rainy prediction model, and the encoder of the sunny and rainy prediction model extracts features of the precipitation forecast data to obtain a feature map; inputting the feature map into the decoder of the sunny and rainy prediction model, and the Bayesian convolutional layer in the decoder assigns weights to the feature map to obtain the sunny and rainy forecast, wherein the sunny and rainy forecast is a sunny and rainy distribution grid field; the Bayesian convolutional layer in the decoder assigns weights to the feature map to obtain the sunny and rainy forecast, which comprises: the Bayesian convolutional layer in the decoder assigns weights to the feature map to output a plurality of uncertainty information, averages the uncertainty information, and obtains a precipitation average probability; taking the precipitation average probability as the value of each grid point in the sunny and rainy forecast as the probability of precipitation at the corresponding position to obtain the sunny and rainy forecast; obtaining EC numerical prediction data, wherein the EC numerical prediction data comprises height layer wind field, geopotential height field and temperature and humidity field data; constructing a terrain enhancement heavy precipitation model, and obtaining a terrain amplification forecast according to the EC numerical prediction data and the terrain enhancement heavy precipitation model; obtaining the terrain amplification forecast comprises: determining whether terrain precipitation occurs according to the temperature field data, the height layer wind field data and the geopotential height field; after determining whether terrain precipitation occurs, calculating the condensation of water vapor in the atmosphere based on the humidity field data to obtain the terrain amplification forecast, wherein the terrain amplification forecast is a heavy precipitation field, and the value of each grid point in the heavy precipitation field is the expected precipitation of the terrain; fusing the sunny and rainy forecast and the terrain amplification forecast to obtain a precipitation forecast field, and obtaining a final precipitation forecast through the precipitation forecast field, comprising: superimposing the terrain amplification forecast and the EC precipitation forecast data to obtain a final terrain amplification forecast; multiplying the sunny and rainy forecast and the final terrain amplification forecast by grid points to obtain the precipitation forecast field; obtaining the final precipitation forecast through the precipitation forecast field.

2. The regional precipitation prediction method according to claim 1, characterized by, training the sunny and rainy prediction model through a precipitation forecast training set comprises: collecting historical EC precipitation forecast data under the ECMWF prediction model, processing the historical EC precipitation forecast data into grid data of a preset resolution to obtain input data; obtaining multi-source real-time precipitation data, processing the multi-source real-time precipitation data into grid data of the preset resolution to obtain label data; constructing the precipitation forecast training set through the input data and the label data, and training the sunny and rainy prediction model through the precipitation forecast training set.

3. The regional precipitation prediction method according to claim 1, characterized by, The method for determining whether terrain precipitation occurs is to calculate the wet Froude number to determine whether terrain precipitation occurs; the calculation method of the wet Froude number is: where U is the average velocity of the air flow; N is the Brunt-Vaisala frequency, g is the acceleration of gravity, θ is the potential temperature, is the change of potential temperature with height, h is the characteristic height of the terrain.

4. The regional precipitation prediction method according to claim 1, characterized by, obtaining the heavy precipitation field comprises: where P is the condensation amount, i.e. the precipitation, R is the condensation rate per unit height, z h is the saturated wet-layer height, and z0is the ground height.

Images