Method and device for constructing extreme rainfall forecasting model

Abstract

The invention discloses a construction method and device of an extreme rainfall forecasting model, and belongs to the technical field of extreme rainfall forecasting, and the method comprises the steps: obtaining historical meteorological data and historical rainfall data; constructing an initial extreme rainfall forecasting model; distributing precipitation weights for the historical precipitation data based on the precipitation level information entropy; establishing a continuous rainfall regression loss function based on the rainfall weight to realize continuous rainfall prediction; during training, taking historical meteorological data as input, taking historical rainfall data as a prediction target, and training the initial extreme rainfall forecasting model based on a continuous rainfall regression loss function to obtain an extreme rainfall forecasting model. According to the method, information balance is carried out by distributing the rainfall weight for the historical rainfall data based on the rainfall level, so that the problem of non-uniform sample data distribution caused by small proportion of extreme rainfall data quantity is solved, and the prediction effect of heavy rainfall is remarkably improved.

Description

Technical Field

[0001] This application belongs to the technical field of extreme precipitation forecasting, and specifically relates to a method and device for constructing an extreme precipitation forecasting model. Background Art

[0002] In recent years, the rapidly developing artificial intelligence has shown potential in the field of weather forecasting. Especially since 2022, a series of AI medium-term weather forecasting models have emerged, such as FourCastNet of NVIDIA, "Pangu-Weather" of Huawei Cloud, GraphCast of Google DeepMind, and "FuXi" of Fudan University

[0003] (FuXi), etc., providing new possible methods for weather forecasting and extreme disaster early warning. These AI models are no longer limited by the traditional numerical model framework, but directly use data-driven. The AI weather large model can autonomously learn the internal relationship between input parameters and output results based on data, bypass the complex physical process modeling and small-scale processes, and directly generate more accurate forecasting results. Starting from "Pangu-Weather" of Huawei Cloud, the accuracy of the medium-term forecasting of the AI model has exceeded that of the best-performing European Centre for Medium-Range Weather Forecasts High-Resolution Forecast (ECMWFHRES) in the traditional numerical model for the first time, which marks a major turning point of the AI model in the field of meteorological forecasting, and the numerical forecasting model is no longer the only option. Taking "Pangu-Weather" of Huawei Cloud as an example, the trained model only takes a few seconds to generate the forecasting results for the next 24 hours, which is tens of thousands of times faster than the Integrated Forecast System (ECMWFIFS) of the European Centre. Although "Pangu-Weather" has exceeded the traditional numerical model in the medium-term forecasting of multiple upper-air and surface meteorological elements for the first time, it does not provide precipitation forecasting. The subsequent AI weather large model "FuXi" provides precipitation forecasting, but the forecasting effect for extreme processes such as heavy precipitation is still weak.

[0004] To cope with the increasingly frequent extreme disastrous weather under the global climate change situation, meet the service requirements of extreme precipitation weather, and ensure the safe operation of cities, it is necessary to provide extreme precipitation forecasting products that are faster and more accurate than traditional numerical models. Summary of the Invention

[0005] The purpose of this application is to provide a method and device for constructing an extreme precipitation forecasting model to solve the problem of great difficulty in extreme precipitation prediction.

[0006] According to the first aspect of the embodiments of this application, a method for constructing an extreme precipitation forecasting model is provided, including:

[0007] Obtain historical meteorological data and historical precipitation data;

[0008] Build an initial extreme precipitation prediction model;

[0009] Assign precipitation weights to the historical precipitation data based on the information entropy of precipitation levels;

[0010] Establish a continuous precipitation regression loss function based on the precipitation weights;

[0011] Use the historical meteorological data as the input and the historical precipitation data as the prediction target, and train the initial extreme precipitation prediction model based on the continuous precipitation regression loss function to obtain an extreme precipitation prediction model.

[0012] In some alternative embodiments of the present application, the historical meteorological data includes upper-air meteorological data and surface meteorological data.

[0013] In some alternative embodiments of the present application, the initial extreme precipitation prediction model includes a feature extraction module, a precipitation prediction module, and a precipitation upscaling module connected in sequence;

[0014] The feature extraction module is used to process the upper-air meteorological data by a three-dimensional convolution method to obtain upper-air meteorological features;

[0015] The feature extraction module is also used to process the surface meteorological data by a two-dimensional convolution method to obtain surface meteorological features;

[0016] The feature extraction module is also used to extract the terrain features and land surface features of the meteorological data through a static layer structure;

[0017] The feature extraction module is also used to splice the upper-air meteorological features, the surface meteorological features, the terrain features, and the land surface features and input them into the precipitation prediction module.

[0018] In some alternative embodiments of the present application, the precipitation prediction module includes an encoder, a hidden layer, and a decoder connected in sequence;

[0019] The encoder includes two layers of residual convolutional networks;

[0020] The hidden layer includes 16 MogaNet blocks.

[0021] In some alternative embodiments of the present application, the precipitation upscaling module is used to improve the spatial resolution through PixelShuffle;

[0022] The neural network boundary conditions in the initial extreme precipitation prediction model are established through a cyclic padding technique.

[0023] In some alternative embodiments of the present application, the upper-air meteorological data includes geopotential height, absolute humidity, temperature, zonal wind speed, and meridional wind speed;

[0024] The ground meteorological data includes the temperature at a height of 2 meters, the zonal wind speed at a height of 10 meters, the meridional wind speed at a height of 10 meters, and the mean sea level pressure.

[0025] In some alternative embodiments of the present application, assigning precipitation weights to the historical precipitation data based on the information entropy of precipitation levels includes:

[0026] Obtaining the occurrence probability of the historical precipitation data;

[0027] Assigning the precipitation weights based on the occurrence probability;

[0028] Wherein, the precipitation weight is positively correlated with the negative logarithm of the occurrence probability.

[0029] In some alternative embodiments of the present application, the continuous precipitation regression loss function is established based on the domain space test method.

[0030] According to the second aspect of the embodiments of the present application, an extreme precipitation forecasting method is provided.

[0031] Including:

[0032] Inputting the real-time meteorological data into an extreme precipitation forecasting model constructed by using the method for constructing an extreme precipitation forecasting model according to any one of claims 1-8 to obtain real-time precipitation forecasting data.

[0033] According to the third aspect of the embodiments of the present application, a device for constructing an extreme precipitation forecasting model is provided, including:

[0034] An acquisition module, configured to acquire historical meteorological data and historical precipitation data;

[0035] A first data processing module, configured to construct an initial extreme precipitation forecasting model;

[0036] An assignment module, configured to assign precipitation weights to the historical precipitation data based on the information entropy of precipitation levels;

[0037] A second data processing module, configured to establish a continuous precipitation regression loss function based on the precipitation weights;

[0038] A training module, configured to use the historical meteorological data as input, use the historical precipitation data as a prediction target, and train the initial extreme precipitation forecasting model based on the continuous precipitation regression loss function to obtain an extreme precipitation forecasting model.

[0039] According to the fourth aspect of the embodiments of the present application, an electronic device is provided, and the electronic device may include:

[0040] A processor;

[0041] A memory for storing processor-executable instructions;

[0042] Wherein, the processor is configured to execute instructions to implement the method for constructing an extreme precipitation prediction model as described in any of the embodiments of the first aspect.

[0043] The above technical solution of the present application has the following beneficial technical effects:

[0044] The embodiment of the present application provides a method for constructing an extreme precipitation prediction model. By allocating precipitation weights to historical precipitation data based on the information entropy of precipitation levels for information balance, the problem of uneven distribution of sample data caused by the small proportion of extreme precipitation data is solved, thereby significantly improving the prediction effect of heavy precipitation. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 is a flowchart of a method for constructing an extreme precipitation prediction model in an exemplary embodiment of the present application;

[0046] Figure 2 is a schematic structural diagram of an extreme precipitation prediction model in an exemplary embodiment of the present application;

[0047] Figure 3 is a distribution diagram of the occurrence times of different precipitation levels in an exemplary embodiment of the present application.

[0048] Figure 4 is a comparison diagram of predicted data and actual data in an exemplary embodiment of the present application.

[0049] Figure 5 is a flowchart of a method for extreme precipitation prediction in an exemplary embodiment of the present application.

[0050] Figure 6 is a schematic diagram of an apparatus for constructing an extreme precipitation prediction model in an exemplary embodiment of the present application;

[0051] Figure 7 is a schematic structural diagram of an electronic device in an exemplary embodiment of the present application. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0052] To make the objectives, technical solutions, and advantages of the present application clearer, the present application will be further described in detail below in conjunction with the specific embodiments and with reference to the accompanying drawings. It should be understood that these descriptions are exemplary only and are not intended to limit the scope of the present application. In addition, in the following description, descriptions of well-known structures and technologies are omitted to avoid unnecessarily confusing the concepts of the present application.

[0053] The schematic diagram of the layer structure according to the embodiments of the present application is shown in the accompanying drawings. These figures are not drawn to scale, where for the purpose of clarity, some details are enlarged and some details may be omitted. The shapes of various regions and layers shown in the figures, as well as their relative sizes and positional relationships, are merely exemplary, and in practice, there may be deviations due to manufacturing tolerances or technical limitations. And those skilled in the art can design regions / layers with different shapes, sizes, and relative positions according to actual needs.

[0054] Obviously, the described embodiments are part of the embodiments of the present application, rather than all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present application.

[0055] In the description of the present application, it should be noted that the terms "first", "second", and "third" are only used for descriptive purposes and cannot be understood as indicating or implying relative importance.

[0056] In addition, the technical features involved in different embodiments of the present application described below can be combined with each other as long as they do not conflict with each other.

[0057] Extreme rainstorms and corresponding floods and geological disasters have caused serious losses to people's lives and property. Against the backdrop of global climate change, extreme precipitation events will generally increase and intensify. At present, it is urgent to improve the forecasting accuracy of precipitation weather, especially extreme precipitation, to ensure safe and stable development.

[0058] Current AI medium-term forecasting models all have deficiencies or are lacking in heavy precipitation forecasting. Among these models, "Pangu" Meteorology and "Fengwu" do not provide precipitation forecasts; GraphCast can provide medium-term precipitation forecasts, but their research points out that there are biases in the precipitation part of the European Centre for Medium-Range Weather Forecasts Fifth Generation Reanalysis Data (ERA5) they use, so the precipitation forecast performance is poor, and even the accuracy test of precipitation is not given; "Fuxi" also uses ERA5 as training data and can provide precipitation forecasts, but the accuracy test of precipitation forecasts is not given in the article either. When using the model training files open-sourced by "Fuxi" for precipitation forecasting, it is found that the forecasting performance for heavy precipitation is poor.

[0059] (1) The extreme imbalance of precipitation data itself makes it difficult for AI models to train and learn. On datasets with uneven precipitation distributions like this, AI models tend to predict the more numerous categories and ignore the less numerous ones, ultimately leading to weak or even missed precipitation forecasts. For extreme rainstorm events that people are more concerned about, although the precipitation intensity is high, due to their low occurrence frequency, it is difficult for the model to accurately learn these heavy precipitation features through direct training, and accurate forecasts for extreme rainstorms cannot be achieved. (2) At the same time, there are also biases in the precipitation data in ERA5. It depends on the cloud microphysics parameterization scheme and convective parameterization scheme specified in the model and does not directly reference actual precipitation data. Especially in ocean areas, there is a significant underestimation of precipitation rates.

[0060] To solve the problem of extreme imbalance in precipitation data, the present invention provides a method for constructing an extreme precipitation forecasting model:

[0061] (1) The present invention will use a balanced information scheme and specifically design a continuous precipitation regression loss function to solve the problem of extreme imbalance in precipitation data. After statistically analyzing the long-term precipitation probability distribution results, precipitation weights are set for different precipitation level distributions to improve the prediction effect of heavy precipitation. (2) Considering the bias between the precipitation data in ERA5 and actual precipitation, the present invention selects the precipitation inversion product NOAA CMORPH based on satellite observation data as the model training dataset, trains a basic precipitation forecasting model with a 0.25-degree resolution globally, and generates precipitation forecasts through the circulation field.

[0062] The following will, with reference to the accompanying drawings, through specific embodiments and their application scenarios, elaborate in detail on a method and device for constructing an extreme precipitation forecasting model provided by an embodiment of the present application.

[0063] As Figure 1 shown, in Embodiment 1 of the present application, a method for constructing an extreme precipitation forecasting model is provided, including the following steps:

[0064] Step S101: Obtain historical meteorological data and historical precipitation data;

[0065] Step S102: Construct an initial extreme precipitation forecasting model;

[0066] Step S103: Assign precipitation weights to historical precipitation data based on the information entropy of precipitation levels;

[0067] Step S104: Establish a continuous precipitation regression loss function based on precipitation weights;

[0068] Step S105: Using the historical meteorological data as the input and the historical precipitation data as the prediction target, train the initial extreme precipitation prediction model based on the continuous precipitation regression loss function to obtain the extreme precipitation prediction model.

[0069] This embodiment provides a method for constructing an extreme precipitation prediction model. By calculating the information entropy based on the precipitation level to allocate precipitation weights for historical precipitation data for information balance, it solves the problem of uneven distribution of sample data caused by the small proportion of extreme precipitation data, thereby significantly improving the prediction effect of heavy precipitation.

[0070] As Figure 2 shown, in Embodiment 2 of the present application, a method for constructing an extreme precipitation prediction model is provided. The extreme precipitation prediction model consists of three parts: a feature extraction module, a precipitation prediction module, and a precipitation upscaling module.

[0071] The feature extraction module includes a circulation field feature extraction part and a static layer. The circulation field feature extraction part will separately perform feature extraction and dimensionality reduction on the upper-air and surface circulation fields, and then merge them and splice them with the static layer parameters that can autonomously learn to represent elements such as terrain; the precipitation prediction module uses an encoder-decoder structure and selects MogaNet (Multi-order Gated Aggregation Network) as the core part of the prediction model; the precipitation upscaling module then raises the spatial resolution of the output result of the precipitation prediction module back to 0.25 degrees and outputs the predicted precipitation distribution. The extreme precipitation prediction model modifies the continuous precipitation regression loss function according to the information balance method, uses a weighted continuous precipitation regression loss function, and uses the field space verification method (FSS) to replace the traditional root mean square error RMSE as the screening index, which can more reasonably evaluate the prediction accuracy of extreme precipitation of different magnitudes.

[0072] Specifically, the training samples used in the present invention include two types of data: reanalysis circulation field and satellite-inverted precipitation. Among them, the reanalysis circulation field uses the European Centre for Medium-Range Weather Forecasts fifth-generation reanalysis data (ERA5), and the precipitation uses the global satellite-inverted precipitation data (NOAA CMORPH) produced by the National Oceanic and Atmospheric Administration of the United States.

[0073] For the re-analysis of the circulation field, 5 commonly used upper-air variables (geopotential height Z, absolute humidity Q, temperature T, zonal wind speed U, and meridional wind speed V) from ERA5 are selected, with a total of 13 levels (50 hPa, 100 hPa, 150 hPa, 200 hPa, 250 hPa, 300 hPa, 400 hPa, 500 hPa, 600 hPa, 700 hPa, 850 hPa, 925 hPa, and 1000 hPa), and 4 surface variables (temperature at 2 m height T2m, zonal wind speed at 10 m height U10, meridional wind speed at 10 m height V10, and mean sea-level pressure MSL), resulting in a total of 69 meteorological element variables. The satellite-inverted precipitation data is a global geostationary satellite-inverted precipitation product, covering the region between 60 degrees north and south latitudes. The time range is from 1998 to 2022, with a time resolution of 30 minutes. To highlight heavy precipitation processes, the precipitation data with an original resolution of 8 km is upsampled to a resolution of 0.25 degrees, and the maximum method interpolation is used to generate hourly rainfall intensity and 6-hour accumulated precipitation data. When processing accumulated precipitation, the original data is first summed, and then the maximum method interpolation is performed to avoid over-exaggerating the precipitation intensity problem and ensure the performance of the dataset during heavy precipitation processes. Since the precipitation data of NOAA CMORPH has included precipitation data within the range of 60 degrees north and south latitudes since 1998, the final time range of the circulation field meteorological elements and precipitation data is 25 years from 1998 to 2022. Among them, the first 24 years are used for training, and the data of the last year is used for testing. When training the basic model, the circulation field and precipitation data within the range of 60 degrees north and south latitudes are used, and the resolution is unified to 0.25 degrees.

[0074] The extreme imbalance of the precipitation data itself makes it difficult for the AI model to train and learn. The AI model tends to predict the categories with a larger number on the unevenly distributed precipitation dataset, ignoring the categories with a smaller number, ultimately resulting in a weak prediction or even a missed prediction in precipitation forecasting. For extreme rainstorm events that people are more concerned about, although the precipitation intensity is large, due to the low occurrence frequency, it is difficult for the model to directly learn these heavy precipitation characteristics accurately during training and unable to achieve accurate forecasting of extreme rainstorms.

[0075] To solve the problem of extreme imbalance in precipitation data, the present invention uses an information balance scheme for continuous precipitation prediction, specifically designs a continuous precipitation regression loss function to solve the problem of extreme imbalance in precipitation data, and sets weights for the information entropy distribution of different precipitation levels after statistically analyzing the long-term precipitation probability distribution results, thereby improving the prediction effect of heavy precipitation. In the present invention, first, all historical precipitation data in the training set is counted according to the distribution of different precipitation levels to obtain a precipitation level frequency distribution map, and precipitation weights are assigned to the historical precipitation data of different levels in turn for calculating the loss value during training. For example Figure 3As shown, all precipitation data is divided into 91 ranges (Bins) according to the magnitude, with the minimum being 0 - 0.1 mm and the maximum being 400 mm. The frequency of precipitation occurrences in each magnitude is counted, and the weight of precipitation in the corresponding magnitude is calculated based on the frequency. Referring to the theory of information entropy, the different precipitation weights (W i ) are positively correlated with the negative logarithm of the occurrence probability (P(y i )) as shown in formula (1), where τ is the temperature coefficient hyperparameter. This means that the smaller the precipitation weight with more occurrences, and the highest weight for extreme precipitation. As Figure 3 shown, the first magnitude has the highest occurrence frequency because the situation of no precipitation is the most in the precipitation data, so the number of 0 values is the largest.

[0076]

[0077] When establishing the continuous precipitation regression loss function, the historical precipitation data includes the target precipitation in a two-dimensional distribution. The target precipitation is used to represent the actual precipitation of each grid point after the historical precipitation area is gridded. The weight calculated based on the target precipitation is added to the original root mean square error, as shown in formula (2). X i is the predicted precipitation of the i-th grid point, Y i is the target precipitation of the i-th grid point, and W i (Y i ) is the precipitation weight of the i-th grid point. The precipitation weight is the weight obtained by looking up according to the true precipitation at each grid point in 91 ranges. After matrix multiplication, the continuous precipitation regression loss function considering the precipitation weight is obtained, thus effectively enhancing the influence of heavy precipitation on the continuous precipitation regression loss function.

[0078]

[0079] The precipitation weight allocation method provided in this embodiment is calculated through precipitation statistical probability and is more scientific and effective compared to manually setting weights. The information balance scheme used in this solution predicts continuous precipitation and solves the problem of uneven distribution of precipitation samples in regression problems, which is different from the previous precipitation classification prediction schemes.

[0080] The specific structure of the extreme precipitation prediction model is as follows:

[0081] (1) Feature extraction module, which extracts spatial feature information from upper-air and surface meteorological fields respectively. Upper-air meteorological elements include geopotential height, specific humidity, meridional wind speed, zonal wind speed and temperature at 13 height levels. These elements are processed using 3D convolution method, which mixes and compresses information of different height levels. Although the spatial resolution is reduced, the number of information channels increases, ensuring the integrity of details. Surface meteorological elements include sea surface pressure, meridional wind speed at 10 m height, zonal wind speed at 10 m height and surface temperature at 2 m height. These elements are processed by 2D convolution method to achieve extraction and compression of spatial features. Subsequently, a learnable static layer structure is introduced to enable the model to autonomously learn and generate static information such as terrain and land surface features, thereby enhancing the adaptability of the model under different terrain and land surface conditions. Finally, all the extracted spatial feature information is merged in the model to achieve comprehensive feature learning and training.

[0082] (2) Precipitation prediction module, which is the core module for the model to predict precipitation, consisting of an encoder, a hidden layer and a decoder. The encoder contains two layers of residual convolutional networks. The hidden layer is the core temporal processing module of the model, containing 16 MogaNet blocks. The decoder is responsible for restoring the features output by the hidden layer to high-resolution precipitation prediction.

[0083] (3) Precipitation upscaling module, which is achieved by PixelShuffle (a deep learning upsampling method), and the residual term of the original resolution is considered to enhance the detailed features of the precipitation forecast of the basic model. Finally, the optimal solution in the model training process is selected by calculating the FSS (Fractionskill score) of heavy precipitation between the forecast result and the original precipitation.

[0084] In addition, in order to ensure the continuity of the longitude boundary, the neural network in the extreme precipitation forecast model adopts the Circular padding technology. This technology can effectively handle the periodic boundary problem in the latitude direction and ensure that the prediction of the model at the boundary does not show discontinuity.

[0085] As Figure 4 shown, the upper figure is the predicted data and the lower figure is the actual data, which is a comparison chart of the six-hour cumulative precipitation prediction results of the extreme precipitation forecast model in the test set. The consistency between the predicted data and the actual data of the extreme precipitation forecast model for heavy precipitation above 25 mm is good.

[0086] The FSS (Fractionskill score) of the extreme precipitation forecast model provided by the present invention for the prediction of heavy precipitation above 25 mm can reach above 0.5:

[0087] Table 1 Scores of precipitation models at different magnitudes in the test set

[0088]

[0089] The extreme precipitation forecasting model provided by the present invention has a higher forecasting accuracy for heavy precipitation than existing mainstream AI large models.

[0090] The third embodiment of the present invention provides an extreme precipitation forecasting method, including inputting real-time meteorological data into an extreme precipitation forecasting model constructed by using the construction method of an extreme precipitation forecasting model provided in any embodiment of the present invention to obtain real-time precipitation forecasting data.

[0091] As Figure 5 shown, in some embodiments, an extreme precipitation forecasting method includes the following steps:

[0092] Step S201: Obtain historical meteorological data and historical precipitation data;

[0093] Step S202: Construct an initial extreme precipitation forecasting model;

[0094] Step S203: Assign precipitation weights to historical precipitation data based on the information entropy of precipitation levels;

[0095] Step S204: Establish a continuous precipitation regression loss function based on the precipitation weights;

[0096] Step S205: Use the historical meteorological data as the input and the historical precipitation data as the prediction target, and train the initial extreme precipitation forecasting model based on the continuous precipitation regression loss function to obtain an extreme precipitation forecasting model;

[0097] Step S206: Input the real-time meteorological data into the extreme precipitation forecasting model to obtain real-time precipitation forecasting data.

[0098] The extreme precipitation forecasting method provided by the present invention significantly improves the detection rate of extreme precipitation and helps to forecast and prevent disasters.

[0099] As Figure 6 shown, based on the same inventive concept, the fourth embodiment of the present application provides a device for constructing an extreme precipitation forecasting model, including:

[0100] An acquisition module 11, configured to acquire historical meteorological data and historical precipitation data;

[0101] A first data processing module 12, configured to construct an initial extreme precipitation forecasting model;

[0102] An assignment module 13, configured to assign precipitation weights to historical precipitation data based on the information entropy of precipitation levels;

[0103] A second data processing module 14, configured to establish a continuous precipitation regression loss function based on the precipitation weights;

[0104] A training module 15, which is configured to use historical meteorological data as input, historical precipitation data as a prediction target, and train an initial extreme precipitation prediction model based on a continuous precipitation regression loss function to obtain an extreme precipitation prediction model.

[0105] Optionally, as Figure 7 shown, an embodiment of the present application further provides an electronic device 1100, including a processor 1101, a memory 1102, and a program or instruction stored on the memory 1102 and executable on the processor 1101. When the program or instruction is executed by the processor 1101, it implements each process of the embodiment of the method for constructing the above extreme precipitation prediction model and can achieve the same technical effect. To avoid repetition, it will not be elaborated here.

[0106] It should be noted that the electronic device in the embodiment of the present application includes the above-mentioned mobile electronic device and non-mobile electronic device.

[0107] An embodiment of the present application further provides a readable storage medium, on which a program or instruction is stored. When the program or instruction is executed by a processor, it implements each process of the embodiment of the method for constructing the above extreme precipitation prediction model and can achieve the same technical effect. To avoid repetition, it will not be elaborated here.

[0108] Wherein, the processor is the processor in the electronic device described in the above embodiment. The readable storage medium includes a computer-readable storage medium, such as a computer read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disc, etc.

[0109] Another embodiment of the present application provides a chip, which includes a processor and a communication interface. The communication interface is coupled to the processor, and the processor is configured to run a program or instruction to implement each process of the embodiment of the method for constructing the above extreme precipitation prediction model and can achieve the same technical effect. To avoid repetition, it will not be elaborated here.

[0110] It should be understood that the chip mentioned in the embodiment of the present application may also be referred to as a system-on-chip, a system chip, a chip system, or a system-on-chip.

[0111] It should be noted that in this article, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements not only includes those elements, but also includes other elements not explicitly listed, or elements inherent to such process, method, article or device. Without further limitation, an element defined by the statement "including one..." does not exclude the existence of additional identical elements in the process, method, article or device including that element. In addition, it should be pointed out that the scope of the methods and devices in the embodiments of the present application is not limited to performing functions in the order shown or discussed, but may also include performing functions in a substantially simultaneous manner or in the reverse order according to the functions involved. For example, the described methods may be performed in an order different from that described, and various steps may be added, omitted, or combined. Additionally, the features described with reference to certain examples may be combined in other examples.

[0112] Through the description of the above embodiments, those skilled in the art can clearly understand that the above-described example methods can be implemented by means of software plus a necessary general hardware platform. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation. Based on this understanding, the technical solution of the present application, in essence, or the part that contributes to the prior art, can be embodied in the form of a computer software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) and includes several instructions for causing a terminal (which can be a mobile phone, computer, server, or network device, etc.) to execute the methods described in the various embodiments of the present application.

[0113] The embodiments of the present application have been described above in conjunction with the accompanying drawings. However, the present application is not limited to the above specific embodiments. The above specific embodiments are merely illustrative and not restrictive. Under the inspiration of the present application, those of ordinary skill in the art can also make many forms without departing from the purpose of the present application and the scope protected by the claims, and all of them belong to the protection scope of the present application.

Claims

1. A method for constructing an extreme precipitation forecast model, characterized in that: include: Obtain historical meteorological data and historical precipitation data; Constructing an initial extreme precipitation forecast model; assigning precipitation weights to the historical precipitation data based on information entropy of precipitation magnitude; Establishing a continuous precipitation regression loss function based on the precipitation weight; The historical meteorological data is used as input, the historical precipitation data is used as a prediction target, and the initial extreme precipitation forecast model is trained based on the continuous precipitation regression loss function to obtain an extreme precipitation forecast model.

2. The method for constructing an extreme precipitation forecast model according to claim 1, characterized in that: The historical meteorological data includes high-altitude meteorological data and ground-based meteorological data.

3. The method for constructing an extreme precipitation forecast model according to claim 2, characterized in that: The initial extreme precipitation forecast model includes a feature extraction module, a precipitation forecast module and a precipitation dimension enhancement module which are connected in sequence; The feature extraction module is used to process the high-altitude meteorological data by a three-dimensional convolution method to obtain high-altitude meteorological features; The feature extraction module is also used to process the ground meteorological data by a two-dimensional convolution method to obtain ground meteorological features; The feature extraction module is also used to extract the terrain features and land surface features of the meteorological data through a static layer structure; The feature extraction module is also used to splice the high-altitude meteorological features, the ground meteorological features, the terrain features and the land surface features and input them into the precipitation forecast module.

4. The method for constructing an extreme precipitation forecast model according to claim 3, characterized in that: The precipitation forecast module includes an encoder, a hidden layer and a decoder connected in sequence; The encoder comprises a two-layer residual convolutional network; The hidden layer includes 16 MogaNet blocks.

5. The method for constructing an extreme precipitation forecast model according to claim 3, characterized in that: The descaling module is used to improve the spatial resolution through PixelShuffle; The neural network boundary conditions in the initial extreme precipitation forecast model are established through a loop filling technique.

6. The method for constructing an extreme precipitation forecast model according to claim 2, characterized in that: The high-altitude meteorological data include geopotential height, absolute humidity, temperature, latitudinal wind speed and meridional wind speed; The ground meteorological data include temperature at 2 meters, latitudinal wind speed at 10 meters, longitudinal wind speed at 10 meters and average sea level pressure.

7. The method for constructing an extreme precipitation forecast model according to claim 1, characterized in that: Allocating precipitation weights to the historical precipitation data based on the information entropy of precipitation magnitude includes: Obtaining the occurrence probability of the historical precipitation data; assigning the precipitation weight based on the occurrence probability; The precipitation weight is positively correlated with the negative logarithm of the occurrence probability.

8. The method for constructing an extreme precipitation forecast model according to claim 1, characterized in that: The continuous precipitation regression loss function is established based on the domain space verification method.

9. An extreme precipitation forecasting method, characterized in that: include: The real-time meteorological data is input into the extreme precipitation forecast model constructed by the method for constructing an extreme precipitation forecast model according to any one of claims 1 to 8 to obtain real-time precipitation forecast data.

10. A device for constructing an extreme precipitation forecast model, characterized in that: include: An acquisition module is used to acquire historical meteorological data and historical precipitation data; The first data processing module is used to construct an initial extreme precipitation forecast model; An allocation module, configured to allocate precipitation weights to the historical precipitation data based on information entropy of precipitation magnitude; A second data processing module, used to establish a continuous precipitation regression loss function based on the precipitation weight; A training module is used to take the historical meteorological data as input, take the historical precipitation data as a prediction target, and train the initial extreme precipitation forecast model based on the continuous precipitation regression loss function to obtain an extreme precipitation forecast model.

11. An electronic device, characterized in that: include: A processor, a memory, and a program or instruction stored in the memory and executable on the processor, wherein the program or instruction, when executed by the processor, implements a method for constructing an extreme precipitation forecast model as described in any one of claims 1 to 8.

Images