A short-to-medium-term precipitation forecasting method combining deep learning with traditional numerical models
Patent Information
- Application Number
- CN202510471721.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-04-15
- Publication Date
- 2026-10-23
AI Technical Summary
第二类误差是数值模式云降水物理过程参数化方面的误差
[0058](1)本发明的神经网络模型中使用了特定数量、特定数值的等压面上的要素场，并且其输入变量中不包含大气数值模式的降水预报产品，从而实现了以下效果：通过较少的数据量反映整体的气象状态，从而减小模型参数数量，加快模型收敛，降低训练成本；本发明研究表明选择更多等压面的数据甚至可能造成模型训练效果变差；本发明的神经网络模型的输入变量不包含大气数值模式的降水预报产品，从而避免数值模式降水误差对神经网络模型的误导、并使得神经网络模型可以与任意的传统数值模式结合。
Smart Images
Figure CN122964685A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of weather forecasting, specifically to a short- to medium-term precipitation forecasting method that combines deep learning with traditional numerical models. Background Technology
[0002] Precipitation is the source of almost all freshwater on Earth and a vital resource for human life and production. However, it is also a major cause of natural disasters such as floods and mudslides. Over the past 20 years, floods have affected 1.6 billion people globally, causing over $650 billion in economic losses, accounting for 22% of all economic losses caused by natural disasters. Therefore, precipitation forecasting has always been one of the most important aspects of weather forecasting. Although weather forecasting technology has made significant progress in recent decades, precipitation forecasting still faces considerable challenges.
[0003] Current precipitation forecasts, especially short- and medium-term precipitation forecasts which are crucial for water resource utilization and flood disaster prevention, mainly rely on numerical weather prediction technology.
[0004] Numerical weather prediction (NMR) techniques often suffer from two distinct types of forecast errors, which further contribute to precipitation forecast errors. The first type of forecast error is the prediction error of large-scale meteorological fields (such as the 500 hPa geopotential height field). A suitable large-scale meteorological field is a necessary condition for precipitation formation. This type of error is a key concern in the development and evaluation of NMR techniques. Satellite observations and assimilation techniques have significantly reduced this type of error. The second type of error is the parameterization error of cloud precipitation physical processes in numerical models. Currently, there is insufficient scientific understanding of many physical processes (such as ice facies processes and subgrid processes). Furthermore, there is a lack of satisfactory, accurate, and time-efficient parameterization schemes. Research shows that improving cloud precipitation physical parameterization is extremely difficult. Therefore, even if the large-scale meteorological field predicted by the numerical model is completely correct, inappropriate cloud precipitation physical parameterization can still cause significant errors in precipitation forecasts.
[0005] In recent years, artificial intelligence has made great progress in weather forecasting, especially with large-scale meteorological models such as Pangu, Fuxi, and Fengwu significantly outperforming traditional numerical weather prediction models in large-scale meteorological field prediction. However, large-scale meteorological models all suffer from problems such as insensitivity to small disturbances, difficulty in predicting extreme values, and lack of physical consistency among variables. In addition, current large-scale meteorological models cannot yet provide highly accurate and satisfactory short- to medium-term (e.g., 1-16 day) precipitation forecasts.
[0006] Therefore, there is an urgent need in this field to develop highly accurate short- and medium-term precipitation forecasting methods, systems, and equipment. Summary of the Invention
[0007] The purpose of this invention is to provide a highly accurate short- and medium-term precipitation forecasting method, system, and device.
[0008] This invention provides a short- to medium-term precipitation forecasting method. This method can input the meteorological field forecast results generated by existing weather forecasting systems into a short- to medium-term precipitation forecasting neural network model based on a large-scale meteorological field, thereby obtaining a more accurate surface precipitation forecast.
[0009] In a first aspect of the present invention, a modeling method for a short- to medium-term precipitation forecast neural network model based on a large-scale meteorological field is disclosed, characterized in that the method includes the following steps:
[0010] (1) Obtain modeling data;
[0011] Acquire meteorological field data for the past N years for a spatial range, where N≥5, and topographic data for the spatial range; the spatial range includes the entire territory of China, and the meteorological field data includes temperature, relative humidity, geopotential height, horizontal wind speed U-component distribution, and horizontal wind speed V-component distribution for five isobaric surfaces; and divide the meteorological field data into training set, validation set, and test set according to the year;
[0012] (2) Establish the model;
[0013] Using the training set data from step (1), a convolutional neural network model with an attention mechanism is established and trained;
[0014] The convolutional neural network model takes 25 channels (including temperature, relative humidity, geopotential height, horizontal wind speed U-component distribution, and horizontal wind speed V-component distribution for each of the five isobaric surfaces) plus the topographic height of the underlying surface as input, and outputs 24-hour cumulative precipitation prediction; wherein, the five isobaric surfaces include isobaric surfaces at 200hPa, 500hPa, 700hPa, 850hPa, and 1000hPa.
[0015] (3) Validate and test the model;
[0016] Using the validation and test set data from step (1), the model established in step (2) is validated and tested to obtain a short- to medium-term precipitation forecast neural network model based on a large-scale meteorological field.
[0017] In a preferred embodiment, the spatial range includes the entire territory of China; and / or
[0018] The terrain data includes the elevation of the underlying surface.
[0019] In a preferred embodiment, in the neural network model, data from 26 channels of each grid point are used as input; and / or
[0020] The size or horizontal resolution of the grid points is 0.25° × 0.25°.
[0021] In a preferred embodiment, the grid points are grid points representing the Chinese region, and the output of the neural network model is configured to be daily precipitation grid data for the Chinese region.
[0022] In a preferred embodiment, during the training of the neural network model, all input meteorological field data are from ERA5 reanalysis data at 12UTC, with a horizontal resolution of 0.25°×0.25°, covering an area of 56°E-152°E and 4°N-68°N, totaling 384×256 grid points. The label data used during training is CHM_PRE data, also with a horizontal resolution of 0.25°×0.25°.
[0023] In a preferred embodiment, the neural network model employs the Attention UNet network architecture.
[0024] In a preferred embodiment, the loss function used during training of the neural network model is as follows:
[0025]
[0026] in,
[0027] MSE stands for Mean Squared Error.
[0028] θ is the threshold;
[0029] TS diff The TS (Threat Score) is a differentiable score; the TS score is used to measure the proportion of grid points in the observation where the precipitation intensity exceeds a certain threshold (θ) and a correct forecast can be obtained.
[0030] μ is the weight of the TS score;
[0031] Hits diff The number of grid points used to observe precipitation intensity exceeding a certain threshold and correctly predicted;
[0032] Misses diff To observe the number of grid points where precipitation intensity exceeds a certain threshold but is not correctly predicted;
[0033] FalseAlarms diff This refers to the number of grid points where the model predicts precipitation exceeding a certain threshold, but the actual observed precipitation does not reach the given threshold.
[0034] P j Predict precipitation intensity at grid point j using the model
[0035] O j To observe the precipitation intensity at grid point j
[0036] In a preferred embodiment, θ and μ are set to 5 mm / day and 0.5, respectively; and / or
[0037] In a preferred embodiment, the short- to medium-term precipitation forecast is configured to forecast for 1 to M days, where M is a positive integer from 10 to 20.
[0038] In a preferred embodiment, M is 14, 15, 16, 17, or 18.
[0039] In a preferred embodiment, the neural network model does not perform parameterization operations on the physical processes of cloud precipitation.
[0040] In a preferred embodiment, step (1) uses 44 years of meteorological field data from 1979 to 2022, with 2001, 2006, 2011, 2016, and 2021 as the training set, 2002, 2007, 2012, 2017, and 2022 as the test set, and the remaining 34 years as the training set.
[0041] A second aspect of the present invention discloses a short- to medium-term precipitation forecasting method based on a large-scale meteorological field, characterized in that the method includes the following steps:
[0042] (1) Acquire and process meteorological field data
[0043] For a given spatial range, variables such as temperature, relative humidity, geopotential height, and the U and V components of horizontal wind speed are extracted from the 1-M day daily meteorological field forecast results of traditional numerical models on five isobaric surfaces within the range of 56°E-152°E and 4°N-68°N, and interpolated at 0.25°×0.25° grid points; where M is a positive integer from 10 to 20; the five isobaric surfaces include isobaric surfaces at 200 hPa, 500 hPa, 700 hPa, 850 hPa, and 1000 hPa.
[0044] (2) Add terrain data
[0045] For each grid point, the terrain height variable of the grid point is added to the interpolated data to form a data structure with 26 channels, so as to meet the parameter requirements when training the model.
[0046] (3) Using neural network models for weather forecasting
[0047] The data from the 26 channels processed in step (2) are input daily into a short- to medium-term precipitation forecast neural network model based on a large-scale meteorological field, thereby outputting a 1-M day 24-hour precipitation forecast. The short- to medium-term precipitation forecast neural network model is established using the modeling method described in any one of claims 1-7.
[0048] In a preferred embodiment, the method is not based on cloud precipitation physical process parameterization.
[0049] A third aspect of the present invention discloses a system for short- and medium-term precipitation forecasting, characterized in that the system comprises: a data acquisition module, a data preprocessing module, a short- and medium-term precipitation forecasting module, and an output module, wherein:
[0050] The data acquisition module is configured to acquire the weather field forecast results of traditional numerical models and input them into the data preprocessing module;
[0051] The data preprocessing module is configured to: preprocess the meteorological field forecast results to obtain 25 channels of variables for temperature, relative humidity, geopotential height, and horizontal wind speed U and V components on 5 isobaric surfaces, and combine the variables of the 25 channels with the underlying surface topographic height data of the grid points to form input data for 26 input channels suitable for the neural network model;
[0052] The short-to-medium-term precipitation forecast module is configured to receive data from 26 channels of the data preprocessing module and analyze it using a short-to-medium-term precipitation forecast neural network model based on a large-scale meteorological field, thereby obtaining short-to-medium-term precipitation forecast data.
[0053] The output module is configured to output the short- and medium-term precipitation forecast data.
[0054] In a preferred embodiment, the short- to medium-term precipitation forecast data is daily precipitation gridded data for the Chinese region.
[0055] The fourth aspect of this application discloses an electronic device, which includes a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that the processor executes the computer program to implement the short-to-medium-term precipitation forecasting method based on a large-scale meteorological field as described in claim 9.
[0056] The fifth aspect of this application discloses a non-transitory computer-readable storage medium comprising a computer program, characterized in that the computer program, when executed by a processor, implements the short-to-medium-term precipitation forecasting method based on a large-scale meteorological field as described in claim 9.
[0057] The main advantages of this invention include:
[0058] (1) The neural network model of the present invention uses a specific number and specific values of element fields on isobaric surfaces, and its input variables do not include precipitation forecast products from atmospheric numerical models, thereby achieving the following effects: reflecting the overall meteorological state with a smaller amount of data, thereby reducing the number of model parameters, accelerating model convergence, and reducing training costs; the present invention shows that selecting more isobaric surface data may even lead to a worse model training effect; the input variables of the neural network model of the present invention do not include precipitation forecast products from atmospheric numerical models, thereby avoiding the misleading effect of precipitation errors in numerical models on the neural network model and enabling the neural network model to be combined with any traditional numerical model.
[0059] (2) Compared with traditional precipitation forecasting that relies entirely on numerical forecasting models, this invention uses deep neural networks to replace the parameterization of cloud precipitation physical processes in numerical models, making full use of the advantages of deep neural networks in simulating complex physical correlation mechanisms and avoiding errors caused by parameterization of cloud precipitation physical processes.
[0060] (3) Compared with large meteorological models that use reanalysis data as the training basis, this invention uses observed precipitation data as training labels, which reduces the impact of the label data error on the model accuracy.
[0061] (4) This invention uses meteorological field forecast products from traditional numerical models to drive neural network models, making full use of the advantages of traditional numerical forecast products, such as high physical consistency and accurate extreme value forecasts.
[0062] (5) Compared with traditional precipitation post-processing methods, the neural network model of the present invention is trained with meteorological fields of quasi-observational reanalysis data, which allows it to be combined with any traditional numerical model without the need for complex adjustments.
[0063] The specification of this invention describes numerous technical features distributed across various technical solutions. Listing all possible combinations of these features (i.e., technical solutions) would make the specification excessively lengthy. To avoid this problem, the various technical features disclosed in the above-described invention, the various technical features disclosed in the following embodiments and examples, and the various technical features disclosed in the accompanying drawings can be freely combined to form various new technical solutions (all of which should be considered as described in this specification), unless such a combination of technical features is technically infeasible. For example, one example discloses feature A+B+C, and another example discloses feature A+B+D+E. Features C and D are equivalent technical means that serve the same function, and technically only one needs to be used; they cannot be used simultaneously. Feature E can technically be combined with feature C. Therefore, the solution A+B+C+D should not be considered as described because it is technically infeasible, while the solution A+B+C+E should be considered as described. Attached Figure Description
[0064] Figure 1 This is a schematic diagram of the neural network model of a short-to-medium-term precipitation forecasting method according to an embodiment of the present invention.
[0065] Figure 2 This is a comparison of the TS and ETS scores of the deep neural network model (aunet) for 24-hour precipitation forecasting in this invention on a 5-year test set with the precipitation product results in ERA5;
[0066] Figure 3 This invention compares the RMSE and spatial correlation coefficient (PCC) of 1-16 day precipitation forecasts generated from NCEP-GFS meteorological fields for 2021-2022 with the precipitation forecast results of NCEP-GFS itself. Solid lines represent the mean, dashed lines represent the median, and shaded areas represent the range between the 25th and 75th percentiles.
[0067] Figure 4 This invention compares the ETS score of the 1-16 day precipitation forecast generated from the 2021-2022 meteorological field using NCEP-GFS data with the precipitation forecast results of NCEP-GFS itself. Detailed Implementation
[0068] Through in-depth and extensive research, the inventors have developed for the first time a prediction method and system for short- and medium-term precipitation forecasting based on a neural network model of large-scale meteorological fields. This invention replaces the parameterization of cloud precipitation physical processes in numerical models with deep neural networks, and uses meteorological field forecast products from traditional numerical models to drive the neural network model, fully leveraging the advantages of high physical consistency and relatively accurate extreme value forecasts of traditional numerical forecast products. Compared to traditional precipitation post-processing methods, the neural network model of this invention is trained using meteorological fields from quasi-observational reanalysis data, allowing it to be combined with any traditional numerical model without complex adjustments.
[0069] Specifically, this invention proposes a short- to medium-term precipitation forecasting method combining deep neural networks and traditional numerical models. This method trains a deep neural network model based on a large-scale meteorological field to forecast cumulative precipitation over the next 24 hours. The large-scale meteorological field data from a traditional numerical model drives the neural network model to generate daily precipitation distribution forecasts for 1-16 days. This method replaces the parameterization of cloud precipitation physical processes in traditional numerical models with a neural network model, avoiding the aforementioned Type II error, namely the precipitation forecast error caused by cloud precipitation physical parameterization. Simultaneously, this method utilizes the large-scale meteorological field forecast products of traditional numerical models, fully leveraging the advantages of high physical consistency and accurate extreme value prediction of traditional models.
[0070] It should be noted that in the invention documents of this patent, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one" does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element. In the invention documents of this patent, if it refers to performing an action according to an element, it means performing the action at least according to that element, including two cases: performing the action only according to that element, and performing the action according to that element and other elements. Expressions such as "multiple," "repeatedly," and "various" include two, two times, two kinds, and more than two, more than two times, and more than two kinds.
[0071] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.
[0072] Example 1
[0073] Example of an embodiment of the present invention Figures 1-2 As shown, it discloses a modeling method for a short-to-medium-term precipitation forecast neural network model based on a large-scale meteorological field. The method includes the following steps:
[0074] An embodiment discloses a modeling method for a short-to-medium-term precipitation forecast neural network model based on a large-scale meteorological field. This neural network model does not perform parameterization operations on cloud precipitation physical processes. The method is characterized by the following steps:
[0075] (1) Obtain modeling data;
[0076] Acquire meteorological field data for the past N years for a spatial range, where N≥5, and topographic data for the spatial range; the spatial range includes the entire territory of China, and the meteorological field data includes temperature, relative humidity, geopotential height, horizontal wind speed U-component distribution, and horizontal wind speed V-component distribution for five isobaric surfaces; and divide the meteorological field data into training set, validation set, and test set according to the year;
[0077] (2) Establish the model;
[0078] Using the training set data in step (1), a convolutional neural network model with an attention mechanism is established and trained;
[0079] The convolutional neural network model takes 25 channels (including temperature, relative humidity, geopotential height, and horizontal wind speed U-component and V-component distributions for each of the five isobaric surfaces) plus the terrain height of the underlying surface as input, and outputs a 24-hour cumulative precipitation prediction. The five isobaric surfaces include those at 200 hPa, 500 hPa, 700 hPa, 850 hPa, and 1000 hPa. The loss function used during training is as follows:
[0080]
[0081] Where MSE is the traditional Mean Squared Error, and TS diff The ThreatScore (TS) is a differentiable TS score. The TS score is a commonly used evaluation metric in precipitation forecasting, used to measure the proportion of grid points where precipitation intensity exceeds a certain threshold θ and a correct forecast is obtained. A higher TS score indicates better forecast performance. μ is the weight of the TS score and is an adjustable hyperparameter. Adding the TS score to the loss function can significantly improve the forecast performance of heavy precipitation. MSE stands for Mean Squared Error. diffMisses is the number of grid points where precipitation intensity exceeds a certain threshold and is correctly predicted. diff To observe the number of grid points where precipitation intensity exceeds a certain threshold but is not correctly predicted;
[0082] FalseAlarms diff P represents the number of grid points where the model predicts precipitation exceeding a certain threshold, but the actual observed precipitation does not reach the given threshold. j To predict the precipitation intensity O at grid point j in the model j To observe the precipitation intensity at grid point j
[0083] The neural network architecture is as follows Figure 1 As shown, the neural network architecture includes an encoder and a decoder. Specifically, the input to this neural network architecture is data with a horizontal resolution of 0.25°×0.25° and a grid of 384×256 points (i.e., 381×256×26) across 26 channels. This data is fed into the encoder, where local meteorological features are extracted through 3×3 convolution + ReLU, progressively compressing the spatial dimension (384×256→24×16) while increasing the number of channels to 1024 to capture multi-scale atmospheric models. Max pooling (step size 2) simulates the spatial downscaling of the meteorological field while preserving key circulation structures.
[0084] In the decoder portion of the neural network architecture, upsampling (×2) is progressively restored to high resolution (384×256), incorporating fine-grained features (such as terrain effects) injected into the encoder via skip connections. Furthermore, different regions are dynamically weighted using attention gates to suppress noise.
[0085] Finally, CHM Precip, i.e. precipitation forecast, is generated in the output layer. The resolution is consistent with the input (384×256) and directly corresponds to the geographic grid prediction.
[0086] (3) Validate and test the model;
[0087] Using the validation and test set data from step (1), the model established in step (2) is validated and tested to obtain a short- to medium-term precipitation forecast neural network model based on a large-scale meteorological field.
[0088] Optionally, in one embodiment, the present invention uses 44 years of data from 1979 to 2022 to train and evaluate the neural network model. Five years are used as the validation set, comprising data from 2001, 2006, 2011, 2016, and 2021; five years are used as the test set, comprising data from 2002, 2007, 2012, 2017, and 2022; and the remaining 34 years are used as the training set. This data partitioning strategy ensures that all three datasets cover different phases of multi-year atmospheric oscillations (such as ENSO), thereby avoiding model performance bias towards a particular phase.
[0089] Optionally, in one embodiment, the spatial range includes parts of China (excluding Hong Kong, Macao, and Taiwan) or the entire territory of China; and / or
[0090] The terrain data includes the elevation of the underlying surface.
[0091] Optionally, in one embodiment, in the neural network model, the data from 26 channels of each grid point are used as input; and / or
[0092] The size or horizontal resolution of the grid points is 0.25° × 0.25°.
[0093] Optionally, in one embodiment, the grid points are grid points in the Chinese region, and the output of the neural network model is configured as daily precipitation grid data for the Chinese region.
[0094] Optionally, in one embodiment, during the training of the neural network model, all input meteorological field data are from ERA5 reanalysis data at 12UTC, with a horizontal resolution of 0.25°×0.25°, covering an area of 56°E-152°E, 4°N-68°N, and a total of 384×256 grid points. The label data used during training is CHM_PRE data, also with a horizontal resolution of 0.25°×0.25°.
[0095] Optionally, in one embodiment, the neural network model employs the Attention UNet network architecture.
[0096] Optionally, in one embodiment, θ and μ are set to 5 mm / day and 0.5, respectively; and / or
[0097] Optionally, in one embodiment, the short- to medium-term precipitation forecast is configured to forecast for 1-M days, where M is a positive integer from 10 to 20. Optionally, in another embodiment, M is 14, 15, 16, 17, or 18.
[0098] The results show that the 24-hour precipitation forecast of the neural network model developed using the method in this embodiment is significantly better than the precipitation forecast of ERA5 (European Centre for Medium-Range Weather Forecasts) itself. Specifically, the RMSE of the precipitation forecast using the neural network model (aunet) is 2.52 mm per day. -1 The result is 3.36 mm day, compared to the ERA5 model. -1 The error rate was 24.9% lower. The output of the neural network model was configured with daily precipitation gridded data for China. In this embodiment, the label data used during the training of the neural network model was CHM_PRE data with a horizontal resolution of 0.25°×0.25°. The input variables of the neural network model did not include precipitation forecast products from atmospheric numerical models, thereby avoiding the misleading influence of numerical model precipitation errors on the neural network model, and enabling the neural network model to be combined with any traditional numerical model. Meanwhile, as... Figure 3 As shown, the neural network model predicts precipitation at 1, 5, 10, 25, 50, and 100 mm days. -1 The TS and ETS scores at the equal threshold were significantly higher than the ERA5 results by 0.054 and 0.059, respectively.
[0099] Example 2
[0100] This embodiment discloses a short- to medium-term precipitation forecasting method based on a large-scale meteorological field. The method includes the following steps:
[0101] (1) Acquire and process meteorological field data
[0102] For a given spatial range, variables such as temperature, relative humidity, geopotential height, and the U and V components of horizontal wind speed are extracted from the 1-M day daily meteorological field forecast results of traditional numerical models on five isobaric surfaces within the range of 56°E-152°E and 4°N-68°N, and interpolated at 0.25°×0.25° grid points; where M is a positive integer from 10 to 20; the five isobaric surfaces include isobaric surfaces at 200 hPa, 500 hPa, 700 hPa, 850 hPa, and 1000 hPa.
[0103] (2) Add terrain data
[0104] For each grid point, the terrain height variable of the grid point is added to the interpolated data to form a data structure with 26 channels, so as to meet the parameter requirements when training the model.
[0105] (3) Using neural network models for weather forecasting
[0106] The data from the 26 channels processed in step (2) are input daily into a short- to medium-term precipitation forecast neural network model based on a large-scale meteorological field, thereby outputting a 1-M day 24-hour precipitation forecast. The short- to medium-term precipitation forecast neural network model is established using the modeling method described in any one of claims 1-7.
[0107] This embodiment also discloses a short-to-medium-term precipitation forecasting system, characterized in that the system includes: a data acquisition module, a data preprocessing module, a short-to-medium-term precipitation forecasting module, and an output module, wherein:
[0108] The data acquisition module is configured to acquire the weather field forecast results of traditional numerical models and input them into the data preprocessing module;
[0109] The data preprocessing module is configured to: preprocess the meteorological field forecast results to obtain 25 channels of variables for temperature, relative humidity, geopotential height, and horizontal wind speed U and V components on 5 isobaric surfaces, and combine the variables of the 25 channels with the underlying surface topographic height data of the grid points to form input data for 26 input channels suitable for the neural network model;
[0110] The short-to-medium-term precipitation forecast module is configured to receive data from 26 channels of the data preprocessing module and analyze it using a short-to-medium-term precipitation forecast neural network model based on a large-scale meteorological field, thereby obtaining short-to-medium-term precipitation forecast data.
[0111] The output module is configured to output the short- and medium-term precipitation forecast data.
[0112] In a preferred embodiment, the short- to medium-term precipitation forecast data is daily precipitation gridded data for the Chinese region.
[0113] A comparison of precipitation forecast data and NCEP-GFS precipitation forecast results in this embodiment. Figures 3-4 As shown, this NCEP-GFS precipitation forecast uses the 12 UTC-based product from the National Centers for Environmental Prediction (NCEP) Global Forecast System (GFS). The forecast has a horizontal resolution of 0.25° × 0.25° (same as ERA5 data) and a forecast duration of 384 hours (16 days). Weather forecasts using the European Centre for Medium-Range Weather Forecasts (ECMWF) model, the China Meteorological Administration (CMA-GFS) model, or large-scale meteorological models may produce precipitation forecasts with comparable accuracy.
[0114] Figure 3and Figure 4 A comparison of the accuracy of the precipitation forecast results of this invention based on 2021-2022 data with the NCEP-GFS precipitation forecast results is presented. Among them, Figure 3 a shows that the average RMSE (root mean square error) of the precipitation forecast of this invention for 1-16 days is significantly lower than that of the NCEP-GFS product results, by an average of 17.7%; Figure 3 b shows that the average PCC (Pearson correlation coefficient) was significantly higher than that of the NCEP-GFS product results in the 1-16 day range, by an average of 0.078; if PCC > 0.5 is used to define effective forecast duration, then it can be achieved through... Figure 3 b. It was found that the effective forecast duration of the precipitation forecast of the present invention is 2 days longer than that of NCEP-GFS; that is, the effective forecast duration of the present invention is about 6.5 days, while the effective forecast duration of NCEP-GFS is about 4.5 days.
[0115] Figure 4 a-4e refers to the Equitable Threat Score (ETS) of the 1-16 day precipitation forecasts generated using NCEP-GFS data from the 2021-2022 meteorological field (in terms of 1, 5, 10, 25, and 50 mm day values). -1 The comparison is between the ETS score (with a threshold) and the precipitation forecast results of NCEP-GFS itself. The two-year average of the ETS score is represented by the solid colored line, corresponding to the left axis; the proportion of days in the two years where the ETS score of the forecast results of this invention is higher than that of the NCEP-GFS forecast results is represented by the solid black line, corresponding to the right axis.
[0116] The results show that the prediction results of this invention are significantly better than those of NCEP-GFS in two aspects.
[0117] first, Figure 4 a- Figure 4 The green line in figure e represents the two-year average ETS score of the forecast results of this invention, while the orange line represents the two-year average ETS score of NCEP-GFS. It is clear from the figure that the two-year average ETS score of the forecast results of this invention is significantly higher than that of NCEP-GFS. Figure 4 In ae, the green line is always above the orange line.
[0118] Secondly, the probability that the prediction results of this invention are better than NCEP-GFS exceeds 50%, that is... Figure 4 In AE, the solid black line always lies above the dashed black line. (Using a 100 mm day...) -1 On the ETS score calculated for the threshold, the accuracy of the prediction results of this invention is comparable to that of the NCEP-GFS results (see...). Figure 4 f).
[0119] Example 3
[0120] This embodiment discloses an electronic device and a non-transitory computer-readable storage medium. The electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. The feature is that when the processor executes the computer program, it implements the short- and medium-term precipitation forecasting method based on a large-scale meteorological field as described in Embodiment 2.
[0121] The non-transitory computer-readable storage medium includes a computer program, characterized in that, when the computer program is executed by a processor, it implements the short-to-medium-term precipitation forecasting method based on a large-scale meteorological field as described in Embodiment 2.
[0122] All documents mentioned in this invention are considered to be incorporated integrally into the disclosure of this invention so that they can serve as the basis for modifications if necessary. Furthermore, it should be understood that the above embodiments are merely preferred embodiments of this specification and are not intended to limit the scope of protection of this specification. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of one or more embodiments of this specification should be included within the scope of protection of one or more embodiments of this specification.
Claims
1. A modeling method for a short- to medium-term precipitation forecast neural network model based on a large-scale meteorological field, characterized in that, The method includes the following steps: (1) Obtain modeling data; Acquire meteorological field data (N≥5) for the past N years for a spatial range, as well as topographic data for the spatial range; the spatial range includes the entire territory of China, and the meteorological field data includes temperature, relative humidity, geopotential height, U-component distribution of horizontal wind speed, and V-component distribution of horizontal wind speed on five isobaric surfaces; and divide the meteorological field data into training set, validation set, and test set according to the year; (2) Establish the model; Using the training set data from step (1), a convolutional neural network model with an attention mechanism is established and trained; The convolutional neural network model takes 25 channels (including temperature, relative humidity, geopotential height, horizontal wind speed U-component distribution, and horizontal wind speed V-component distribution for each of the five isobaric surfaces) plus the topographic height of the underlying surface as input, and outputs 24-hour cumulative precipitation prediction; wherein, the five isobaric surfaces include isobaric surfaces at 200hPa, 500hPa, 700hPa, 850hPa, and 1000hPa. (3) Validate and test the model; Using the validation and test set data from step (1), the model established in step (2) is validated and tested to obtain a short- to medium-term precipitation forecast neural network model based on a large-scale meteorological field.
2. The method according to claim 1, characterized in that, The spatial scope includes the entire territory of China; and / or The terrain data includes the elevation of the underlying surface.
3. The method according to claim 1, characterized in that, In the neural network model, the data from 26 channels of each grid point are used as input; and / or The size or horizontal resolution of the grid points is 0.25° × 0.25°.
4. The method according to claim 3, characterized in that, The grid points are grid points for the Chinese region, and the output of the neural network model is configured to be daily precipitation grid data for the Chinese region.
5. The method according to claim 1, characterized in that, During the training of the neural network model, all input meteorological field data came from the ERA5 reanalysis data at 12UTC, with a horizontal resolution of 0.25°×0.25°, covering an area of 56°E-152°E and 4°N-68°N, totaling 384×256 grid points. The label data used during training was CHM_PRE data, also with a horizontal resolution of 0.25°×0.25°.
6. The method according to claim 1, characterized in that, The neural network model adopts the Attention UNet network architecture.
7. The method according to claim 1, characterized in that, The loss function used during the training of the neural network model is as follows: in, MSE stands for Mean Squared Error. θ is the threshold; TS diff The TS (Threat Score) is a differentiable score; the TS score is used to measure the proportion of grid points in the observation where the precipitation intensity exceeds a certain threshold (θ) and can be correctly predicted. μ is the weight of the TS score; Hits diff The number of grid points used to observe precipitation intensity exceeding a certain threshold and correctly predicted; Misses diff To observe the number of grid points where precipitation intensity exceeds a certain threshold but is not correctly predicted; FalseAlarms diff This refers to the number of grid points where the model predicts precipitation exceeding a certain threshold, but the actual observed precipitation does not reach the given threshold. P j The model predicts the precipitation intensity at grid point j; O j The precipitation intensity at grid point j in the observation is given.
8. The method according to claim 7, characterized in that, The short- to medium-term precipitation forecast is configured to forecast for 1 to M days, where M is a positive integer from 10 to 20.
9. A short- to medium-term precipitation forecasting method based on a large-scale meteorological field, characterized in that, The method includes the following steps: (1) Acquire and process meteorological field data For a given spatial range, variables such as temperature, relative humidity, geopotential height, and the U and V components of horizontal wind speed are extracted from the 1-M day daily meteorological field forecast results of traditional numerical models on five isobaric surfaces within the range of 56°E-152°E and 4°N-68°N, and interpolated at 0.25°×0.25° grid points; where M is a positive integer from 10 to 20; the five isobaric surfaces include isobaric surfaces at 200 hPa, 500 hPa, 700 hPa, 850 hPa, and 1000 hPa. (2) Add terrain data For each grid point, the terrain height variable of the grid point is added to the interpolated data to form a data structure with 26 channels, so as to meet the parameter requirements when training the model. (3) Using neural network models for weather forecasting The data from the 26 channels processed in step (2) are input daily into a short- to medium-term precipitation forecast neural network model based on a large-scale meteorological field, thereby outputting a 1-M day 24-hour precipitation forecast. The short- to medium-term precipitation forecast neural network model is established using the modeling method described in any one of claims 1-7.
10. A short- to medium-term precipitation forecasting system, characterized in that, The system includes: a data acquisition module, a data preprocessing module, a short- and medium-term precipitation forecasting module, and an output module, wherein: The data acquisition module is configured to acquire the weather field forecast results of traditional numerical models and input them into the data preprocessing module; The data preprocessing module is configured to: preprocess the meteorological field forecast results to obtain 25 channels of variables for temperature, relative humidity, geopotential height, and horizontal wind speed U and V components on 5 isobaric surfaces, and combine the variables of the 25 channels with the underlying surface topographic height data of the grid points to form input data for 26 input channels suitable for the neural network model; The short-to-medium-term precipitation forecast module is configured to receive data from 26 channels of the data preprocessing module and analyze it using a short-to-medium-term precipitation forecast neural network model based on a large-scale meteorological field, thereby obtaining short-to-medium-term precipitation forecast data. The output module is configured to output the short- and medium-term precipitation forecast data.
11. 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 computer program, it implements the short-to-medium-term precipitation forecasting method based on a large-scale meteorological field as described in claim 9.
12. A non-transitory computer-readable storage medium, the non-transitory computer-readable storage medium comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the short-to-medium-term precipitation forecasting method based on a large-scale meteorological field as described in claim 9.