Meteorological large model precipitation prediction method, system, equipment and medium based on satellite inversion precipitation data

Abstract

The present application belongs to the technical field of precipitation prediction, and discloses a meteorological large model precipitation prediction method, system, device and medium based on satellite inversion precipitation data, to solve the problem of insufficient prediction accuracy of meteorological large model. The method comprises: obtaining ERA5 reanalysis data and satellite inversion precipitation data; constructing a bias quantization model for different climate regions and precipitation intensities based on satellite precipitation data; using a dynamic weighted fusion strategy to correct the bias of ERA5 precipitation data, and splicing with other meteorological variables to form an optimized data set; freezing the core parameters of the pre-trained meteorological large model to fine-tune the precipitation prediction layer with the optimized data set and satellite precipitation label; inputting the ERA5 data of the region to be predicted after bias quantization and dynamic weighted fusion correction into the fine-tuned model to output multi-time precipitation prediction results in a recursive prediction manner. The present application can effectively eliminate the weak bias of ERA5 precipitation and significantly improve the precipitation prediction accuracy.

Description

Technical Field

[0001] This invention belongs to the field of precipitation prediction technology, specifically relating to a method, system, equipment, and medium for precipitation prediction based on a large meteorological model using satellite-retrieved precipitation data. Background Technology

[0002] Precipitation forecasting has significant guiding importance for agricultural production, disaster prevention and mitigation, water resource management, and transportation. In recent years, large-scale meteorological models based on architectures such as Transformer and CNN have been widely used in precipitation forecasting, and their forecast accuracy has gradually approached or even surpassed that of traditional numerical weather prediction models.

[0003] Currently, the training and forecasting of mainstream meteorological models largely rely on ERA5 reanalysis data released by the European Centre for Medium-Range Weather Forecasts (ECMWF). This data boasts advantages such as high spatiotemporal resolution, comprehensive meteorological variables, and wide coverage, making it the core foundational data for training meteorological models. However, during the precipitation variable inversion process, ERA5 reanalysis data suffers from a systematic precipitation underestimation bias due to physical factors such as uneven distribution of observation stations, limitations of assimilation algorithms, and cloud top observation errors. Specifically, it underestimates low-to-medium intensity precipitation such as light and moderate rain, and fails to capture the peak values ​​of high-intensity precipitation such as heavy rain and storms. This bias is not due to human manipulation or data selection, but rather stems from the physical modeling limitations of the existing atmospheric reanalysis technology system, representing an inherent defect in the technology itself. This defect makes it difficult for meteorological models trained on raw ERA5 data to meet practical application requirements for precipitation prediction accuracy, especially in scenarios with high precision requirements such as agricultural irrigation scheduling and storm disaster warnings, where the bias problem is even more pronounced.

[0004] To address the aforementioned issues, existing technologies primarily focus on improving the architecture of large meteorological models or increasing the amount of training data. While satellite-retrieved precipitation data (such as GPM and TRMM satellite products) offers global coverage, high observation frequency, and high precipitation retrieval accuracy, serving as a valid reference for actual precipitation values, existing methods still suffer from the following technical shortcomings when utilizing this type of data: First, bias correction often employs a fixed-weight fusion strategy, failing to adaptively adjust weights based on the distribution characteristics of precipitation bias across different regions and intensities, resulting in limited correction effectiveness. Second, the core parameters of the pre-trained model are not properly frozen during model adjustment, easily compromising the model's existing meteorological feature extraction capabilities and leading to a decline in model generalization ability. Third, training loss functions often use a single numerical error index, making it difficult to comprehensively measure the bias in precipitation prediction, resulting in insufficient precision in model adjustment direction.

[0005] In summary, existing technologies for precipitation forecasting using ERA5 reanalysis data in large-scale meteorological models have failed to effectively eliminate the systematic precipitation bias inherent in the data source itself, and also lack targeted bias correction and model adjustment strategies, resulting in limited precipitation forecast accuracy. Summary of the Invention

[0006] Based on the aforementioned shortcomings and deficiencies in the existing technology, one of the objectives of this invention is to at least solve one or more of the aforementioned problems in the existing technology. In other words, one of the objectives of this invention is to provide a meteorological large model precipitation prediction method, system, equipment, and medium that meets one or more of the aforementioned requirements, so as to accurately solve the systematic precipitation weakness bias in ERA5 reanalysis data and significantly improve the precipitation prediction accuracy of the meteorological large model.

[0007] To achieve the above-mentioned objectives, the present invention adopts the following technical solution: In a first aspect, the present invention provides a meteorological large-scale model precipitation prediction method based on satellite inversion precipitation data, comprising the following steps: S1. Obtain ERA5 reanalysis data and satellite-retrieved precipitation data; S2. Using the satellite-retrieved precipitation data as a reference, quantitatively analyze the bias characteristics of the precipitation data in the ERA5 reanalysis data for different climate regions and different precipitation intensity ranges, so as to construct a bias quantification model that characterizes the degree of deviation between the precipitation data and the true precipitation value. S3. Using a dynamic weighted fusion strategy, based on the output of the deviation quantification model, determine the fusion weight of the satellite-retrieved precipitation data and the precipitation data to correct the deviation of the precipitation data. Then, stitch the corrected precipitation data with other meteorological variable data in the ERA5 reanalysis data to form an optimized model input dataset. S4. Obtain and fix the core network parameters of the pre-trained meteorological model, use the optimized model input dataset as training data and the satellite inverted precipitation data as labels, adjust the precipitation prediction output layer and related feature fusion layer of the pre-trained meteorological model to obtain the adjusted meteorological model. S5. After sequentially performing deviation quantification and dynamic weighted fusion correction on the ERA5 reanalysis data of the area to be predicted, input the adjusted meteorological big model and output multi-time precipitation prediction results in a recursive forecasting manner.

[0008] Secondly, the present invention provides a meteorological large-scale model precipitation prediction system based on satellite inversion precipitation data, which is used to implement the meteorological large-scale model precipitation prediction method as described in the first aspect.

[0009] Thirdly, the present invention provides an electronic device, the electronic device including a memory, a processor and a computer program, wherein when the computer program is executed by the processor, it implements the meteorological large model precipitation prediction method as described in the first aspect.

[0010] Fourthly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the large-scale meteorological model precipitation prediction method as described in the first aspect.

[0011] Compared with the prior art, the present invention has the following beneficial effects: 1. This invention introduces high-precision satellite-retrieved precipitation data as a reference benchmark, and constructs deviation quantification models for different climate regions and precipitation intensity ranges to accurately characterize the systematic precipitation underestimation bias of ERA5 reanalysis data. Based on this, a dynamic weighted fusion strategy is adopted to adaptively adjust the fusion weights of satellite data and ERA5 data according to the degree of deviation, effectively correcting the underestimation bias of ERA5 precipitation data. The corrected data is then input into a large meteorological model for fine-tuning training, enabling the model to learn more accurate precipitation distribution patterns. Practical verification shows that compared to models trained based on original ERA5 data, the precipitation prediction accuracy of this invention is significantly improved, and the root mean square error (RMSE) and mean absolute error (MAE) are greatly reduced, effectively solving the technical problem of underestimation in existing precipitation forecasts.

[0012] 2. The precipitation bias characteristics of ERA5 reanalysis data differ significantly across different climatic regions and precipitation intensities. To address this, this invention establishes separate bias quantification models and dynamically adjusts the fusion weights based on the bias rate. This combination of scenario-specific modeling and dynamic weighting allows bias correction to adaptively adapt to changes in bias characteristics across different scenarios, overcoming the limitation of existing fixed-weight fusion strategies that cannot adjust according to changes in bias characteristics. Whether in humid or arid regions, and regardless of light rain or heavy rain, the correction strategy of this invention maintains high correction accuracy and demonstrates excellent scenario adaptability.

[0013] 3. Pre-trained meteorological models have learned general meteorological feature extraction capabilities (such as the spatial distribution patterns of temperature, wind, and pressure fields) through extensive data. Globally adjusting these core parameters can easily disrupt the model's original physical consistency modeling ability. To address this issue, this invention employs a fine-tuning strategy that freezes core network parameters. This involves keeping the feature extraction layer parameters unchanged and optimizing only the precipitation prediction output layer and related feature fusion layers. This design effectively protects the model's ability to represent other meteorological variables while correcting precipitation bias, thus ensuring the model's overall generalization performance and physical rationality.

[0014] 4. Traditional single numerical error loss functions are insufficient to comprehensively measure the bias in precipitation forecasting, easily leading to inaccurate model fine-tuning. Therefore, this invention employs a weighted loss function that integrates mean squared error (MSE) and mean absolute error (MAE). MSE has a larger penalty weight for large errors, effectively suppressing extreme and outrageous prediction biases; MAE focuses on the average bias of overall predictions, ensuring the model maintains good prediction accuracy on most samples. This weighted fusion allows the model to remain sensitive to peak predictions of high-intensity precipitation while effectively controlling systematic biases in medium- and low-intensity precipitation during training, resulting in more comprehensive and balanced precipitation forecasting performance.

[0015] 5. This invention starts from the data source and forms a complete technical chain including deviation quantification, dynamic weighted fusion correction, model fine-tuning, prediction verification, and feedback optimization. Each step supports and synergizes with the others. Specifically, the deviation quantification model provides accurate deviation feature input for dynamic weighted fusion; dynamic weighted fusion correction provides high-quality input data for model fine-tuning; model fine-tuning provides optimized model parameters for precipitation prediction; and verification feedback provides a clear direction for parameter adjustment. The improvements in each stage are cumulative rather than mutually canceling, ultimately achieving a systematic improvement in the accuracy of precipitation prediction.

[0016] 6. The core improvement of this invention lies in the data fusion and correction strategy rather than a specific model structure, thus it does not rely on a specific meteorological large-scale model architecture or a specific reanalysis data source. On the one hand, it can be extended to meteorological large-scale models trained on other reanalysis data such as MERRA-2, JRA-55, and CFSR; on the other hand, it can be adapted to pre-trained meteorological large-scale models with different architectures such as Transformer, CNN, LSTM, and GNN. In application, only the data preprocessing parameters and the bias quantification model fitting parameters need to be adjusted, without modifying the core process of the method, demonstrating good industrial practical value and promising prospects for widespread application.

[0017] Further or more detailed beneficial effects will be described in conjunction with specific embodiments in the detailed implementation. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a flowchart illustrating the large-scale meteorological model precipitation prediction method described in Embodiment 1 of the present invention.

[0020] Figure 2 This is a schematic diagram of the meteorological large-scale precipitation prediction system described in Embodiment 2 of the present invention.

[0021] Figure 3 This is a structural diagram of the electronic device described in Embodiment 3 of the present invention.

[0022] Icon labels: 300. Electronic devices; 301. Processor; 302. Communication bus; 303. User interface; 304. Network interface; 305. Memory. Detailed Implementation

[0023] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0024] In the following description, several embodiments of the present invention are provided. Different embodiments can be substituted or combined. Therefore, the present invention can also be considered to include all possible combinations of the same and / or different embodiments described. Thus, if one embodiment includes features A, B, and C, and another embodiment includes features B and D, then the present invention should also be considered to include embodiments containing one or more other possible combinations of A, B, C, and D, even if such embodiments are not explicitly described in the following text.

[0025] The following description provides examples and does not limit the scope, applicability, or examples set forth in the claims. Changes may be made to the function and arrangement of the described elements without departing from the scope of the invention. Various processes or components may be appropriately omitted, substituted, or added to the various examples. For example, the described methods may be performed in a different order than described, and various steps may be added, omitted, or combined. Furthermore, features described with respect to some examples may be combined into other examples.

[0026] To facilitate a better understanding of the embodiments of the present invention, its application scenarios will be explained before providing a detailed explanation of the specific implementation methods.

[0027] The meteorological large-scale model precipitation prediction method described in the embodiments of this specification is applied to scenarios with high requirements for precipitation prediction accuracy, such as agricultural irrigation scheduling, rainstorm disaster early warning, water resource regulation, and transportation safety. In these scenarios, the application of the meteorological large-scale model precipitation prediction method aims to accurately eliminate the systematic precipitation bias inherent in the ERA5 data by fusing satellite inversion precipitation data with ERA5 reanalysis data. This improves the prediction accuracy of the meteorological large-scale model for low-to-medium intensity precipitation (light rain, moderate rain) and high-intensity precipitation (heavy rain, rainstorm), providing reliable meteorological support for agricultural production decisions, disaster prevention and mitigation deployment, optimal water resource allocation, and transportation scheduling.

[0028] The following is a brief explanation of the ERA5 reanalysis data, satellite-retrieved precipitation data, climate regions, precipitation intensity ranges, bias characteristics, fusion weights, pre-trained large meteorological models, and recursive forecasting methods involved in the various embodiments of this specification: ERA5 reanalysis data is the fifth generation of global atmospheric reanalysis data released by the European Centre for Medium-Range Weather Forecasts (ECMWF). It covers meteorological variables across the entire atmospheric pressure layer from 50 hPa to 1000 hPa globally. Its core components include more than 70 types of meteorological variables such as precipitation, temperature, air pressure, relative humidity, zonal wind, and meridional wind. It has a spatial resolution of 0.25° × 0.25° and a temporal resolution of 6 hours, and is the core foundational data for training large-scale meteorological models.

[0029] Satellite-retrieved precipitation data refers to precipitation observation products retrieved from GPM (Global Precipitation Measurement Satellite) or TRMM (Tropical Rainfall Measurement Satellite). These data are acquired using satellite remote sensing technology and are characterized by global coverage, high observation frequency, and high precipitation retrieval accuracy.

[0030] Climate Regions: Based on the spatial distribution characteristics of precipitation bias, this invention divides climate regions into four categories: arid, humid, semi-arid, and semi-humid. Due to differences in atmospheric circulation, water vapor transport, and topography, the precipitation bias characteristics of ERA5 reanalysis data vary significantly across different climate regions. Therefore, it is necessary to construct bias quantification models for each region.

[0031] Precipitation Intensity Ranges: This invention classifies precipitation into four levels based on intensity: light rain, moderate rain, heavy rain, and torrential rain or above. Light rain is defined as daily precipitation of 0.1 mm to 9.9 mm, moderate rain as 10.0 mm to 24.9 mm, heavy rain as 25.0 mm to 49.9 mm, and torrential rain or above as ≥50.0 mm. Within different precipitation intensity ranges, the degree of bias in ERA5 reanalysis data varies; the bias for low-to-medium intensity precipitation (light and moderate rain) is generally more significant than that for high-intensity precipitation (heavy and torrential rain).

[0032] The underestimation bias refers to the systematic underestimation of precipitation data in ERA5 reanalysis data relative to the true precipitation values ​​(based on satellite-retrieved precipitation data). Specifically, it manifests as lower precipitation values, with a greater underestimation of low-to-medium intensity precipitation such as light and moderate rain, and insufficient capture of peak values ​​for high-intensity precipitation such as heavy rain and torrential rain. This bias stems from physical factors such as uneven distribution of observation stations, limitations of the assimilation algorithm, and cloud top observation errors, and is an inherent defect of the reanalysis technology system itself.

[0033] In a dynamic weighted fusion strategy, the fusion weight is a coefficient used to determine the proportion of contribution of satellite-retrieved precipitation data and ERA5 precipitation data to the fusion result. The weight value is adaptively adjusted according to the severity of the ERA5 precipitation bias: when the bias is significant or the precipitation intensity is low, the satellite-retrieved precipitation data has a higher weight; when the bias is small or the precipitation intensity is high, the ERA5 precipitation data has a higher weight.

[0034] Pre-trained large-scale meteorological models: These are deep learning models based on Transformer or CNN architectures, which have undergone initial training using raw ERA5 data and possess good basic capabilities for meteorological feature extraction and precipitation prediction. The Transformer architecture model uses a 4×4 patch embedding method to process the meteorological field, containing 12 to 16 feature extraction layers and 4 to 6 decoder layers, with a total of 500 million to 1 billion parameters. This type of model serves as the foundation for fine-tuning and optimization in this invention.

[0035] Recursive forecasting is a method for predicting precipitation over multiple time periods. It uses the current time period's forecast as input for the next time period, sequentially outputting precipitation forecasts for each time period. This invention employs recursive forecasting to sequentially output precipitation forecasts for the next 6, 12, 24, 48, and 72 hours. Each forecast includes information on precipitation intensity, precipitation range, and precipitation duration.

[0036] Example 1: like Figure 1 As shown, this embodiment provides a meteorological large-scale model precipitation prediction method based on satellite inversion precipitation data, including the following steps: Step S1, Data Acquisition and Preprocessing: Obtain ERA5 reanalysis data and satellite inversion precipitation data.

[0037] Specifically, in this embodiment, the ERA5 reanalysis data comes from publicly available products released by the European Centre for Medium-Range Weather Forecasts (ECMWF), covering meteorological variables across the entire pressure layer from 50 hPa to 1000 hPa globally. The core data includes more than 70 types of meteorological variables such as precipitation, temperature, air pressure, relative humidity, zonal wind, and meridional wind. As shown in Table 1, this embodiment selects 23 key meteorological variables as model input features, with a spatial resolution of 0.25° × 0.25°, a temporal resolution of 6 hours, and a time span covering at least nearly 30 years.

[0038] Table 1:

[0039] The satellite-retrieved precipitation data are precipitation observation products retrieved using GPM (Global Precipitation Measurement Satellite) or TRMM (Tropical Rainfall Measurement Satellite), with a spatial resolution of 0.1°×0.1° and a temporal resolution adjusted to 6 hours, consistent with the temporal resolution of ERA5 data. This satellite-retrieved precipitation data must be verified against ground-based meteorological station observation data, with a verification accuracy of no less than 90%, to ensure its reliability as a reference for true precipitation values.

[0040] After acquiring the ERA5 reanalysis data and the satellite-retrieved precipitation data, the two types of data are preprocessed, including spatial alignment, time synchronization, outlier removal, and standardization.

[0041] Spatial alignment uses bilinear interpolation with an interpolation kernel size of 4×4 and an interpolation step size of 1. The satellite-retrieved precipitation data is interpolated to a spatial resolution of 0.25°×0.25°, which perfectly matches the grid specifications of the ERA5 data.

[0042] Time synchronization uses linear interpolation with an interpolation window of two time steps before and after, filling in the time step gaps caused by missing observations and transmission delays in the two types of data.

[0043] Outlier removal employs the 3σ criterion. By calculating the global mean μ and global standard deviation σ for each type of data, outlier data points that deviate from the mean by more than three times the standard deviation are removed. For data gaps after outlier removal, linear interpolation is used to fill in the gaps.

[0044] The standardization process uses the Z-score standardization method, and the standardization formula is: (1), In formula (1): For standardized data, The original data is represented by μ, which is the global mean of this data type, and σ, which is the global standard deviation of this data type. The global mean μ and global standard deviation σ are obtained by statistically analyzing ERA5 historical reanalysis data (nearly 30 years) and satellite-retrieved historical precipitation data (nearly 30 years), respectively. The mean and standard deviation are updated annually to ensure the timeliness of the standardization process. The global mean μ and global standard deviation σ are stored in the mean.npy and std.npy files for easy access during subsequent data preprocessing.

[0045] After the above preprocessing operations, this step yields preprocessed data with a uniform format that can be merged.

[0046] Step S2, Deviation Quantification Model Construction: Using the satellite-retrieved precipitation data as a reference, the bias characteristics of precipitation data in the ERA5 reanalysis data are quantitatively analyzed for different climate regions and different precipitation intensity ranges, so as to construct a bias quantification model that characterizes the degree of deviation between the precipitation data and the true precipitation value.

[0047] Specifically, this embodiment uses preprocessed satellite-retrieved precipitation data as a reference benchmark. It quantitatively analyzes the weak bias characteristics of ERA5 precipitation data for four climate regions (arid, humid, semi-arid, and semi-humid) and four precipitation intensity ranges (light rain (0.1 mm / 24h to 9.9 mm / 24h), moderate rain (10.0 mm / 24h to 24.9 mm / 24h), heavy rain (25.0 mm / 24h to 49.9 mm / 24h), and torrential rain and above (≥50.0 mm / 24h). The relative, absolute, and bias rates of ERA5 precipitation data compared to satellite-retrieved precipitation data are calculated for each region and precipitation intensity.

[0048] The deviation quantification model is a linear regression model, and its formula is: (2), In equation (2): x is the preprocessed ERA5 precipitation data (input), y is the preprocessed satellite-retrieved precipitation data (output, used as a reference for the true value), k is the bias correction coefficient, and b is the intercept. The specific values ​​of k and b are obtained by least squares fitting, and the fitting process satisfies: (3), In equation (3): n is the total number of data samples, and the total number of samples is not less than 10,000 to ensure the reliability of the fitting results; For the i-th ERA5 precipitation data sample, Let i be the precipitation data sample retrieved from the i-th satellite, and and Corresponding to the same grid point and the same time step.

[0049] After fitting, the model's fit is verified using the R² test. The R² value must be ≥0.85. If the R² value is lower than 0.85, the samples need to be re-screened (outlier samples removed, sample numbers increased), and fitting is performed again until the fitting accuracy requirement is met. Simultaneously, for different climate regions and different precipitation intensity ranges, corresponding k and b values ​​are fitted separately to ensure that the deviation quantification model can adapt to the deviation characteristics under different scenarios, thereby improving the accuracy of deviation quantification.

[0050] Step S3, Data Fusion and Bias Correction: A dynamic weighted fusion strategy is adopted. Based on the output of the deviation quantification model, the fusion weight of the satellite-retrieved precipitation data and the precipitation data is determined to correct the deviation of the precipitation data. The corrected precipitation data is then spliced ​​with other meteorological variable data in the ERA5 reanalysis data to form an optimized model input dataset.

[0051] Specifically, the ERA5 precipitation deviation coefficients for each grid point and time step are first obtained based on the deviation quantification model, and then the dynamic weights are determined based on these deviation coefficients. The fusion formula for the dynamic weighted fusion strategy is as follows: (4), In equation (4): For the merged precipitation data, For satellite inversion of precipitation data, This is ERA5 precipitation data. Weights for satellite-retrieved precipitation data. The weights for ERA5 precipitation data are given, and they satisfy the following conditions: .

[0052] The weighting rules are as follows: When the ERA5 precipitation is significantly weak (deviation rate ≥30%) and the precipitation intensity falls within the range of heavy rain or torrential rain and above, The value range is [0.7, 0.8] and The value range is [0.2, 0.3]. The higher the value within this range, the higher the deviation rate. The closer the value is to 0.8, the better. The closer the value is to 0.2; When the ERA5 precipitation is significantly weak (deviation rate ≥30%) and the precipitation intensity falls within the range of light or moderate rain, The value range is [0.6, 0.7] and The value range is [0.3, 0.4]; When the ERA5 precipitation is weak with a small deviation (deviation rate <30%) and the precipitation intensity falls within the range of heavy rain or torrential rain and above, The value range is [0.4, 0.6] and The value range is [0.4, 0.6]; When the ERA5 precipitation is weak with a small deviation (deviation rate <30%) and the precipitation intensity falls within the range of light or moderate rain, The value range is [0.2, 0.4] and The value range is [0.6, 0.8]. When selecting a value within this range, the lower the deviation rate, The closer the value is to 0.8, the better. The closer the value is to 0.2, the better.

[0053] The weight update frequency is consistent with the satellite inversion data update frequency, which is once a day or once every 6 hours, to ensure that the weights can be adapted to the latest bias characteristics in real time.

[0054] During the fusion process, the rationality of the fused precipitation data is checked, and data that exceeds the reasonable precipitation range (0-500mm / 24h) is removed.

[0055] The corrected precipitation data was concatenated with other meteorological variable data (temperature, air pressure, humidity, wind field, etc.) from the ERA5 reanalysis data according to grid points and time steps to form the optimized model input dataset. The optimized model input dataset was then divided into training, validation, and test sets in a ratio of 7:2:1.

[0056] Step S4, Model Adjustment: The core network parameters of the pre-trained meteorological model are obtained and fixed. The optimized model input dataset is used as training data and the satellite inverted precipitation data is used as labels. The precipitation prediction output layer and related feature fusion layer of the pre-trained meteorological model are adjusted to obtain the adjusted meteorological model.

[0057] Specifically, this embodiment loads a pre-trained meteorological model based on the Transformer architecture. This model uses a 4×4 patch embedding method to process the meteorological field, and includes 12 to 16 feature extraction layers and 4 to 6 decoder layers. The total number of parameters is 500 million to 1 billion, and it has completed preliminary training using the original ERA5 data.

[0058] During the adjustment process, the core network parameters of the model are frozen, meaning all parameters below the feature extraction layer remain unchanged, and parameter optimization is performed only on the precipitation prediction output layer and related feature fusion layers. This avoids damaging the model's existing meteorological feature extraction capabilities, optimizes only parameters directly related to precipitation prediction, and improves adjustment efficiency.

[0059] Using the optimized model input dataset as training data and the verified true values ​​of satellite-retrieved precipitation data as labels, a weighted loss function fusing mean squared error (MSE) and mean absolute error (MAE) is set. The loss function formula is: (5), In the formula: MSE is the mean squared error; MAE is the mean absolute error; α is the MSE weighting coefficient (preferably in the range of 0.5-0.7 in this embodiment), β is the MAE weighting coefficient (preferably in the range of 0.3-0.5 in this embodiment), and satisfies the following conditions: The specific calculation formulas for MSE and MAE are as follows: (6), (7), In the formula: n is the total number of training samples, and n is not less than 7000; This represents the true value of precipitation data retrieved from satellite. These are the model's predicted values.

[0060] The Adam optimizer was used for training, with an initial learning rate of 0.001-0.01, which was adaptively adjusted based on model convergence. The learning rate gradually decreased with each iteration, with a decay coefficient of 0.95, decreasing once every 10 iterations; the number of iterations ranged from 100 to 500. An early stopping strategy was employed: training was stopped when the validation set loss value did not decrease for 10 consecutive iterations to prevent overfitting. During the tuning process, the training and validation set loss values ​​were monitored in real time to ensure both convergence and that the difference between the validation and training set loss values ​​did not exceed 0.1, ensuring the model's generalization ability. After tuning, the model tuning effect was validated using test set data; the test set prediction accuracy needed to be ≥85% to ensure the effectiveness and convergence of the model tuning.

[0061] Step S5, Precipitation Forecast: The ERA5 reanalysis data of the region to be predicted are sequentially corrected by bias quantification and dynamic weighted fusion, and then input into the adjusted meteorological big model to output multi-time precipitation prediction results in a recursive forecasting manner.

[0062] Specifically, the ERA5 reanalysis data of the area to be predicted is corrected for bias using the bias quantification model described in step S2 and the dynamic weighted fusion strategy described in step S3, and then input into the adjusted meteorological big data model described in step S4. The adjusted meteorological big data model uses a recursive forecasting method to sequentially output precipitation forecasts for the next 6 hours, 12 hours, 24 hours, 48 ​​hours, and 72 hours. Each forecast result includes information on precipitation intensity, precipitation range, and precipitation duration.

[0063] Furthermore, this embodiment uses preset accuracy indicators to verify the multi-time precipitation prediction results. The preset accuracy indicators include at least precipitation prediction accuracy, root mean square error (RMSE), and mean absolute error (MAE). During the verification process, measured precipitation data from at least 50 ground meteorological stations in the area to be predicted are selected as real labels.

[0064] The specific calculation methods and judgment criteria for each accuracy index are as follows: Precipitation forecast accuracy: (8), Root Mean Square Error (RMSE): (9), Mean Absolute Error (MAE): (10) In equations (8) to (10): n is the total number of verification samples, n is not less than 2000, and the verification samples must cover different climate regions and different precipitation intensity ranges to ensure the comprehensiveness and objectivity of the verification results; The actual precipitation values ​​are derived from measured data from ground meteorological stations in the area to be predicted. The measured data must undergo quality control to remove outliers. These are the model's predicted precipitation values.

[0065] When the multi-time precipitation prediction results fail to meet the preset accuracy requirements, the system returns to step S3 to adjust the weight range and weight update frequency of the dynamic weighted fusion, or returns to step S4 to optimize the model and adjust parameters (learning rate, number of iterations, loss function weight coefficients, etc.), until the preset accuracy requirements are met before outputting the final precipitation prediction result. The preset accuracy requirements are: accuracy ≥ 85%, RMSE ≤ 5mm / 24h, MAE ≤ 3mm / 24h.

[0066] Example 2: like Figure 2 As shown, this embodiment provides a meteorological large-scale model precipitation prediction system based on satellite inversion precipitation data, including: The data acquisition module is used to acquire ERA5 reanalysis data and satellite-retrieved precipitation data; The deviation quantification module is used to quantify and analyze the weak deviation characteristics of the precipitation data in the ERA5 reanalysis data for different climate regions and different precipitation intensity ranges, using the satellite-retrieved precipitation data as a reference benchmark, so as to construct a deviation quantification model that characterizes the degree of deviation between the precipitation data and the true precipitation value. The data fusion correction module is used to determine the fusion weight of the satellite-retrieved precipitation data and the precipitation data based on the output of the deviation quantification model using a dynamic weighted fusion strategy, so as to correct the deviation of the precipitation data, and then splice the corrected precipitation data with other meteorological variable data in the ERA5 reanalysis data to form an optimized model input dataset. The model adjustment module is used to acquire and fix the core network parameters of the pre-trained meteorological large model. Using the optimized model input dataset as training data and the satellite inverted precipitation data as labels, the precipitation prediction output layer and related feature fusion layer of the pre-trained meteorological large model are adjusted to obtain the adjusted meteorological large model. The precipitation prediction module is used to input the ERA5 reanalysis data of the area to be predicted into the adjusted meteorological model after sequentially performing deviation quantification and dynamic weighted fusion correction, and output multi-time precipitation prediction results in a recursive forecasting manner.

[0067] Example 3: like Figure 3 As shown, this embodiment provides an electronic device, which may include: at least one processor, at least one network interface, a user interface, a memory, and at least one communication bus.

[0068] The communication bus can be used to enable communication between the various components mentioned above.

[0069] The user interface may include buttons, and optional user interfaces may also include standard wired interfaces and wireless interfaces.

[0070] The network interface may include, but is not limited to, Bluetooth modules, NFC modules, Wi-Fi modules, etc.

[0071] The processor may include one or more processing cores. It connects various parts of the electronic device via various interfaces and lines, executing instructions, programs, code sets, or instruction sets stored in memory, and accessing data stored in memory to perform various functions and process data. Optionally, the processor can be implemented using at least one hardware form of DSP, FPGA, or PLA. The processor may integrate one or more of the following: CPU, GPU, and modem. The CPU primarily handles the operating system, user interface, and applications; the GPU is responsible for rendering and drawing the content required for display; and the modem handles wireless communication. It is understood that the modem may also be implemented as a separate chip without being integrated into the processor.

[0072] The memory may include RAM or ROM. Optionally, the memory may include a non-transitory computer-readable medium. The memory can be used to store instructions, programs, code, code sets, or instruction sets. The memory may include a program storage area and a data storage area, wherein the program storage area may store instructions for implementing an operating system, instructions for at least one function (such as touch function, sound playback function, image playback function, etc.), instructions for implementing the above-described method embodiments, etc.; the data storage area may store data involved in the above-described method embodiments, etc. Optionally, the memory may also be at least one storage device located remotely from the aforementioned processor. The memory, as a computer storage medium, may include an operating system, a network communication module, a user interface module, and a prediction application. The processor can be used to call the prediction application stored in the memory and execute the steps of the meteorological large-scale model precipitation prediction method mentioned in the foregoing embodiments.

[0073] Example 4: This embodiment provides a computer-readable storage medium storing instructions that, when executed on a computer or processor, cause the computer or processor to perform the above-described instructions. Figure 1 One or more steps in the illustrated embodiment. If the constituent modules of the above-described electronic device are implemented as software functional units and sold or used as independent products, they can be stored in the computer-readable storage medium.

[0074] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this specification are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in or transmitted through a computer-readable storage medium. The computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, Digital Subscriber Line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium accessible to a computer or a data storage device such as a server or data center that integrates one or more available media. The available media may be magnetic media (e.g., floppy disks, hard disks, magnetic tapes), optical media (e.g., Digital Versatile Discs (DVDs)), or semiconductor media (e.g., Solid State Disks (SSDs)).

[0075] Those skilled in the art will understand that all or part of the processes in the method of Embodiment 1 described above can be implemented by a computer program instructing related hardware. This program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments of the methods described above. The aforementioned storage medium includes various media capable of storing program code, such as ROM, RAM, magnetic disks, or optical disks. Unless otherwise specified, the technical features of this embodiment and the implementation scheme can be combined arbitrarily.

[0076] Example 5: To verify the effectiveness of the meteorological large-scale model precipitation prediction method based on satellite inversion precipitation data described in this specification, this embodiment uses the 2024 flood season (June to August) precipitation prediction in the Jiangnan region (humid area) of my country as an application scenario to verify the method of the present invention in practice. The verification process is as follows: Application scenarios and data preparation: In this embodiment, the Jiangnan region of my country (22°N-32°N, 110°E-122°E) is selected as the area to be predicted, the prediction period is from August 10 to August 12, 2024, and the prediction lead time is 72 hours.

[0077] During the data preparation phase, ERA5 reanalysis data and GPM satellite-retrieved precipitation data were acquired according to step S1 of Example 1. Verification with measured precipitation data from 50 ground meteorological stations in the Jiangnan region showed that the accuracy of the GPM satellite-retrieved precipitation data was 92.3%, meeting the requirements for use as a reference benchmark for true precipitation values.

[0078] Deviation quantification model construction: Following the method in step S2 of Example 1, the Jiangnan region was divided into three sub-regions: eastern, central, and western. The deviation rate between ERA5 precipitation data and GPM satellite data was calculated for each of the four intensity ranges: light rain, moderate rain, heavy rain, and torrential rain. Quantitative analysis showed that the deviation rate was 25%-35% in the eastern region, 30%-40% in the central region, and 20%-30% in the western region.

[0079] A linear regression deviation quantification model was constructed, and the fitting correction coefficients k and intercept b were determined by region and precipitation intensity. In this embodiment, k=1.32 and b=0.25 for light rain in the central region, k=1.28 and b=0.31 for moderate rain in the eastern region, and k=1.15 and b=0.18 for heavy rain in the western region.

[0080] Dynamic weighted fusion correction: Following the method in step S3 of Example 1, the fusion weights are determined based on the deviation rate: for the central region (deviation rate 30%-40%) and the light to moderate rain range, the satellite-retrieved precipitation data weight w1=0.7-0.8, and the ERA5 precipitation data weight w2=0.2-0.3; for the eastern region (deviation rate 25%-35%) and the light to moderate rain range, w1=0.65-0.75, w2=0.25-0.35; for the western region (deviation rate 20%-30%) and the heavy rain to torrential rain range, w1=0.25-0.35, w2=0.65-0.75. The fusion weights are updated every 6 hours.

[0081] The corrected precipitation data is calculated by fusion formula (formula (4) described in Example 1), and the corrected precipitation data is spliced ​​with other meteorological variable data in ERA5 reanalysis data to form an optimized model input dataset, which is divided into training set, validation set and test set in a ratio of 7:2:1.

[0082] Model Adjustment: Following the method in step S4 of Example 1, load the pre-trained meteorological model based on the Transformer architecture. Freeze all parameters from the feature extraction layer down, and adjust only the precipitation prediction output layer and the feature fusion layer.

[0083] A weighted loss function was set, using the Adam optimizer with an initial learning rate of 0.005, decaying by 5% every 10 iterations for a total of 300 iterations, and an early stopping strategy was employed. After adjustment, the training set loss was 0.82, the validation set loss was 0.89, and the test set prediction accuracy was 88.7%.

[0084] Precipitation forecasting and effect verification: Following the method in step S5 of Example 1, the ERA5 reanalysis data of the Jiangnan region from August 10 to August 12, 2024, were sequentially corrected by bias quantification and dynamic weighted fusion, and then input into the adjusted meteorological big model to output multi-time precipitation forecast results in a recursive forecasting manner.

[0085] Measured precipitation data from 30 surface meteorological stations in the Jiangnan region during this period were selected as real labels to verify the prediction results. The verification results are as follows: the precipitation prediction accuracy is 89.2%, the root mean square error (RMSE) is 4.3 mm / 24h, and the mean absolute error (MAE) is 2.7 mm / 24h, all of which meet the preset accuracy requirements.

[0086] Based on the above, this embodiment verifies the effectiveness of the meteorological large-scale model precipitation prediction method based on satellite inversion precipitation data described in this specification. This invention, through an integrated technical solution of bias quantification, dynamic weighted fusion correction, and model adjustment, accurately solves the problem of systematically weak precipitation bias in ERA5 reanalysis data, significantly improving the precipitation prediction accuracy of the meteorological large-scale model.

[0087] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that the present invention is not limited to the described order of actions, as some steps can be performed in other orders or simultaneously according to the present invention. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions and modules involved are not necessarily essential to the present invention.

[0088] In the above embodiments, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0089] The foregoing description is merely an exemplary embodiment of the present invention and should not be construed as limiting the scope of the invention. Any equivalent changes and modifications made in accordance with the teachings of this invention are still within the scope of this invention. Those skilled in the art will readily conceive of embodiments of the invention upon considering the specification and practicing the disclosure herein. This invention is intended to cover any variations, uses, or adaptations of the invention that follow the general principles of the invention and include knowledge or means in the art not described herein. The specification and embodiments are to be considered exemplary only, and the scope and spirit of the invention are defined by the claims.

Claims

1. A meteorological large-scale model precipitation prediction method based on satellite inversion precipitation data, characterized in that, Including the following steps: S1. Obtain ERA5 reanalysis data and satellite-retrieved precipitation data; S2. Using the satellite-retrieved precipitation data as a reference, quantitatively analyze the bias characteristics of the precipitation data in the ERA5 reanalysis data for different climate regions and different precipitation intensity ranges, so as to construct a bias quantification model that characterizes the degree of deviation between the precipitation data and the true precipitation value. S3. Using a dynamic weighted fusion strategy, based on the output of the deviation quantification model, determine the fusion weight of the satellite-retrieved precipitation data and the precipitation data to correct the deviation of the precipitation data. Then, stitch the corrected precipitation data with other meteorological variable data in the ERA5 reanalysis data to form an optimized model input dataset. The dynamic weighted fusion strategy is specifically as follows: When the deviation rate of precipitation data in the ERA5 reanalysis data reaches a preset threshold and its precipitation intensity belongs to the light rain or moderate rain range, the fusion weight of the satellite-retrieved precipitation data is higher than the fusion weight of the precipitation data in the ERA5 reanalysis data. When the deviation rate of precipitation data in the ERA5 reanalysis data is lower than a preset threshold and its precipitation intensity belongs to the heavy rain or rainstorm range, the fusion weight of precipitation data in the ERA5 reanalysis data is higher than the fusion weight of precipitation data retrieved from satellites. The fusion weights are dynamically updated according to the update frequency of the satellite-retrieved precipitation data. S4. Obtain and fix the core network parameters of the pre-trained meteorological model, use the optimized model input dataset as training data and the satellite inverted precipitation data as labels, adjust the precipitation prediction output layer and feature fusion layer of the pre-trained meteorological model to obtain the adjusted meteorological model. S5. After sequentially performing deviation quantification and dynamic weighted fusion correction on the ERA5 reanalysis data of the area to be predicted, input the adjusted meteorological big model and output multi-time precipitation prediction results in a recursive forecasting manner.

2. The meteorological large-scale model precipitation prediction method according to claim 1, characterized in that: The deviation quantification model is established for four climate regions: arid, humid, semi-arid, and semi-humid, as well as four precipitation intensity ranges: light rain, moderate rain, heavy rain, and rainstorm. The deviation quantification model obtains the correction coefficients and intercepts for correcting precipitation deviations by fitting the linear relationship between precipitation data in the ERA5 reanalysis data and the satellite-retrieved precipitation data.

3. The precipitation prediction method based on a large meteorological model according to claim 1, characterized in that: The core network parameters of the fixed pre-trained meteorological model refer to keeping the parameters of the feature extraction layer used to extract meteorological features in the pre-trained meteorological model unchanged, and only optimizing the parameters of the precipitation prediction output layer and the feature fusion layer. The parameter optimization process employs a weighted loss function that integrates mean squared error and mean absolute error, along with an optimizer with a progressively decreasing learning rate and an early stopping strategy.

4. The meteorological large-scale model precipitation prediction method according to claim 1, characterized in that: The recursive forecasting method refers to using the forecast result of the current time as the input of the next time, and outputting the precipitation forecast for each time in sequence. The multi-time precipitation forecast results include precipitation forecasts for the next 6 hours, 12 hours, 24 hours, 48 ​​hours, and 72 hours. Each forecast result includes information on precipitation intensity, precipitation range, and precipitation duration.

5. The meteorological large-scale model precipitation prediction method according to claim 1, characterized in that, After outputting the multi-time precipitation prediction results in step S5, the following steps are also included: The multi-time precipitation prediction results were verified using preset accuracy indicators. The preset accuracy indicators are categorized into precipitation prediction accuracy, root mean square error, and mean absolute error based on the indicator type. When the multi-time precipitation prediction results do not meet the preset accuracy requirements, the weight parameters of the dynamic weighted fusion strategy are adjusted according to the type of index that did not meet the requirements, or the fine-tuning parameters of the pre-trained meteorological model are adjusted, until the preset accuracy requirements are met and the final precipitation prediction results are output.

6. The meteorological large-scale model precipitation prediction method according to claim 1, characterized in that, The steps following step S1 and before step S2 include: The acquired ERA5 reanalysis data and the satellite-retrieved precipitation data are spatially aligned, time-synchronized, outlier removed, and standardized to obtain preprocessed data in a unified format. The spatial alignment uses bilinear interpolation to interpolate the satellite-retrieved precipitation data to the same spatial resolution as the ERA5 reanalysis data. The time synchronization uses linear interpolation to fill in time step gaps.

7. A meteorological large-scale model precipitation prediction system based on satellite inversion precipitation data, characterized in that, Used to implement the large-scale meteorological model precipitation prediction method as described in any one of claims 1 to 6.

8. An electronic device, the electronic device comprising a memory, a processor, and a computer program, characterized in that, When the computer program is executed by the processor, it implements the meteorological large-scale precipitation prediction method as described in any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the meteorological large-scale precipitation prediction method as described in any one of claims 1 to 6.

Images