A precipitation forecast correction method combining EOF projection and U-Net network

Abstract

This invention discloses a precipitation forecast correction method combining EOF projection and U-Net network. Based on historical observation data, it extracts the main spatial modes of the precipitation anomaly field and then projects and reconstructs the model's forecast anomaly field to obtain a physically meaningful spatial structure. By introducing observed climatological states to replace model climatological states, it effectively eliminates the model's systematic bias. For residual terms that cannot be explained during the reconstruction process, U-Net is used for modeling and prediction to capture complex nonlinear error components in the model. Finally, the linear reconstruction result is superimposed with the residual terms predicted by deep learning to form the final precipitation forecast correction result. This invention constructs a correction framework that combines physical and data-driven approaches. While retaining the physical interpretability of the traditional EOF method, it introduces a deep learning model to model the residual terms, thereby achieving dual capture and correction of linear and nonlinear errors in model forecasts.

Description

Technical Field

[0001] This invention relates to the fields of meteorological data processing and machine learning technology, specifically to a precipitation forecast correction method that combines EOF projection and U-Net network. Background Technology

[0002] Precipitation forecasting, especially extended-range and sub-seasonal precipitation forecasting, has long been a challenge in operational meteorological forecasting. While numerical models can provide information on the evolution of large-scale circulation backgrounds, they often exhibit significant systematic biases and nonlinear errors in precipitation simulations, particularly in local scales and anomalous weather events. Therefore, effectively post-processing and correcting precipitation outputs from numerical models has become a crucial step in improving forecast accuracy and operational value.

[0003] To address this, the industry has proposed methods for correcting forecasts based on empirical orthogonal functions (EOF) or their variants. These methods perform spatial mode decomposition on historical observation data and model outputs to extract key spatial distribution features, and then establish statistical relationships between observations and models to correct model forecasts. For example, patent document CN115936156A proposes a regional error correction method based on rotated empirical orthogonal function decomposition. This method divides the country into several regions with relatively consistent climate characteristics, performs EOF mode correction on each region separately, and finally splices the results to form the national forecast. This method overcomes the problem of high-order mode instability in traditional EOF methods to a certain extent and has strong physical basis and regional adaptability.

[0004] However, the aforementioned methods still have many limitations in practical applications. First, REOF partitions are often large, leading to lower variance explained by the principal mode in subsequent EOF analyses. This necessitates the introduction of more higher-order modes, increasing computational complexity and limiting their application in refined regional operations (such as provincial scales). Second, existing methods are mostly based on the original field for correction, failing to effectively address the systematic drift of the model to the climate mean state, particularly in the identification and correction of anomalous precipitation events. Furthermore, traditional EOF correction methods typically rely on the linear correlation between observational and model principal mode time series, approximating the observational field through a two-step process of projection and regression. This process not only introduces errors but also places high demands on the stability of historical relationships. However, the correlation between precipitation and circulation factors has been shown to exhibit significant interdecadal variations, leading to uncertainties in the long-term application of this method.

[0005] In recent years, with the development of artificial intelligence technology, some studies have attempted to introduce deep learning models to correct pattern outputs, achieving some results. However, these methods are mostly "end-to-end" black-box models. Although they have strong nonlinear fitting capabilities, they lack physical interpretability and it is difficult to judge the rationality of forecast results from a mechanistic perspective. Especially when facing extreme weather processes not covered by training data, the model's generalization ability and stability are difficult to guarantee, limiting its credibility and promotion value in business applications. Summary of the Invention

[0006] The present invention aims to at least partially solve one of the technical problems existing in the related art.

[0007] One objective of this invention is to provide a precipitation forecast correction method that combines EOF projection and U-Net network. By constructing a correction framework that combines physical and data-driven approaches, while retaining the physical interpretability of the traditional EOF method, a deep learning model is introduced to model the residual terms, thereby achieving dual capture and correction of linear and nonlinear errors in the model forecast.

[0008] To achieve the above objectives, the present invention provides a precipitation forecast correction method combining EOF projection and U-Net network, comprising the following steps:

[0009] S1. Obtain historical precipitation data as the observation field, perform climatological calculations according to calendar days to obtain the daily average observation climate field, and calculate the daily observation anomaly field based on the observed average climate field and historical precipitation data; perform empirical orthogonal function decomposition on the observation anomaly field, extract the first N principal modes, and construct the principal mode space of the observation anomaly field.

[0010] S2. Obtain the historical forecast data from the numerical model, calculate the climate mean field for each forecast lead time, and extract the model forecast anomaly field.

[0011] S3. Project the model-forecasted anomaly field onto the principal mode space of the observed anomaly field to obtain projection coefficients, and reconstruct the model-forecasted anomaly field based on the principal mode space of the observed anomaly field and the projection coefficients; adjust the variance of the reconstructed model-forecasted anomaly field, and superimpose the adjusted model-forecasted anomaly field with the observed climate mean field to obtain the reconstructed forecast field; calculate the residual between the reconstructed forecast field and the corresponding observed field to obtain the residual term;

[0012] S4. Construct a U-Net network, and train the U-Net network with the predicted anomaly field of the model as input and the residual term as the output target to obtain an independent residual correction model under different forecast lead times.

[0013] S5. For any future time period, under each forecast lead time, obtain the model forecast anomaly field and the reconstructed forecast field, and predict the residual terms based on the corresponding residual correction model. Then, superimpose the predicted residual terms with the reconstructed forecast field to obtain the corrected precipitation forecast field.

[0014] In a further preferred embodiment of the present invention, step S1 specifically comprises:

[0015] S11. Collect the observation fields of grid points for all dates during the modeling period. To form a dataset of observation fields ,in, Indicates the time period. Represents latitude grid points, The data is represented by longitude grid points. During the modeling period, the data is grouped by calendar day, forming 365 independent datasets. For each calendar day group, an arithmetic mean is calculated across the spatial grid points, representing the average of all years' observations for that date at each geographic grid point, thus forming a daily observed climate mean field. dataset ;

[0016] S12. For any specific date in the historical sequence, determine the actual observation field for that date. With the corresponding observed climate mean field By performing grid-by-grid subtraction, the daily observed anomaly fields are obtained. The observed anomaly fields for all dates will be obtained. Arranged in chronological order, they form the observed anomaly field matrix. ;

[0017] S13, regarding the observed anomaly field matrix Perform EOF decomposition, retain the first N principal modes and their corresponding explained variances, and ensure that the cumulative explained variance of the first N principal modes exceeds 75%.

[0018] Preferably, step S2 involves obtaining historical forecast data from the numerical model, calculating the climate mean field for each forecast lead time, and extracting the model-forecasted anomalous fields; specifically:

[0019] S21. Obtain historical back-calculation forecast data during the modeling period. ,in, Indicates the forecast lead time;

[0020] S22. For forecast lead time (LD), forecast data calculated from historical data... Extract all forecast data with start times for that forecast lead time to form the forecast field for that forecast lead time (LD). ;

[0021] S23. Align the forecast fields for the forecast lead time (LD) of all reporting times during the modeling period according to the target date to obtain the average climate field for each day of the year under the LD of the forecast lead time, denoted as . ;

[0022] S24. Under the forecast lead time (LD), based on the obtained climate mean field The high-frequency perturbation components of the separable model are used to obtain the model-predicted anomalous field. .

[0023] Preferably, in step S3, the model-predicted anomaly field is projected onto the principal mode space of the observed anomaly field to obtain projection coefficients, and the model-predicted anomaly field is reconstructed based on the principal mode space of the observed anomaly field and the projection coefficients; specifically:

[0024] S31. Under the forecast lead time (LD) at time point T, the model forecast anomaly field The projections are sequentially projected onto the principal mode space of the observed anomaly field to obtain N projection coefficients. Based on the principal mode space of the observed anomaly field and the projection coefficients, a mode prediction of the anomaly field is reconstructed. The calculation formula is:

[0025] ;

[0026] in, The first one represents the observed anomaly field. One principal modal space, This represents the corresponding projection coefficient.

[0027] Preferably, in step S3, the variance of the reconstructed model forecast anomaly field is adjusted, and the adjusted model forecast anomaly field is superimposed with the observed climate mean field to obtain the reconstructed forecast field; the residual between the reconstructed forecast field and the corresponding observed field is calculated to obtain the residual term; specifically:

[0028] S32. Anomalies in the reconstructed model forecasts Variance adjustment is performed using the following formula:

[0029] ;

[0030] in, To explain the variance proportion coefficient, , For the observation of the anomaly field The explained variance of each principal modal space;

[0031] S33, The model forecast anomaly field after variance adjustment With the observed climate mean field Superimposed, the reconstructed forecast field is obtained. , represented as:

[0032] ;

[0033] S34, Calculate and reconstruct the forecast field With the corresponding observation field The residuals between them are used to obtain the residual terms. ;

[0034] S35. Calculate the reconstructed forecast field at each time point and forecast lead time. and residuals This yields a set of reconstructed forecast fields with linear relationships. and the set of residual terms representing the nonlinear variation. .

[0035] As a preferred embodiment, the U-Net network constructed in step S4 adopts an encoder-decoder structure, which includes 4 downsampling and 4 upsampling operations. Each convolution uses a 3x3 convolution kernel and a ReLU activation function, and the feature maps of the corresponding layers of the encoder and decoder are fused through skip connections.

[0036] Using the constructed U-Net network, an independent residual correction model is trained for each forecast lead time (LD). During training, a set of pattern-predicted anomaly fields is used. As input features, the set of residual terms As the prediction target, it is represented as:

[0037] ;

[0038] Sixty residual correction models with different forecast lead times were obtained through training.

[0039] As a preferred option, step S5 specifically involves:

[0040] At any future time T, the model forecast anomaly field is calculated under each forecast lead time LD. and reconstructing the forecast field Based on the residual correction model corresponding to the forecast lead time, the corresponding residual terms are predicted. The predicted residual terms With reconstructed forecast field By superimposing the data, the corrected precipitation forecast field is obtained. , represented as:

[0041] .

[0042] In another aspect, the present invention provides a non-transitory computer-readable storage medium having computer instructions stored thereon, the computer instructions causing a computer to execute the above-described precipitation forecast correction method combining EOF projection and U-Net network.

[0043] In another aspect, the present invention provides an electronic device, comprising: a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus, and the processor calls logical instructions in the memory to execute the above-described precipitation forecast correction method combining EOF projection and U-Net network.

[0044] In another aspect, the present invention provides a computer program product comprising a computer program stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer performs the above-described precipitation forecast correction method combining EOF projection and U-Net network.

[0045] Beneficial Effects: The precipitation forecast correction method combining EOF projection and U-Net network of this invention first extracts the main spatial modes of the precipitation anomaly field based on historical observation data, and then projects and reconstructs the model forecast anomaly field to obtain a physically meaningful spatial structure. Subsequently, by introducing observed climatological states to replace model climatological states, the systematic bias of the model is effectively eliminated. Furthermore, for the residual terms that cannot be explained during the reconstruction process, a U-Net deep learning model is used for modeling and prediction to capture the complex nonlinear error components in the model. Finally, the linear reconstruction results are superimposed with the residual terms predicted by deep learning to form the final precipitation forecast correction result. Compared with the prior art, the advantages of this invention are:

[0046] 1. Stability of the Physical Method: This invention projects the high-dimensional, complex original error field onto a low-dimensional feature space with clear physical meaning through empirical orthogonal function (EOF) decomposition. Furthermore, it replaces model-based climate states with observed climate states and simplifies the two-step approximation in traditional projection methods, reducing error propagation and effectively avoiding the interdecadal stability problem of correlations. This step lays the foundation for the physical interpretability of the method and significantly reduces the dimensionality of subsequent processing.

[0047] 2. The Powerful Capabilities of Machine Learning: Machine learning models are used to model and predict the residuals generated by statistical methods. This aims to leverage its powerful nonlinear fitting capabilities and automatic feature mining advantages to accurately capture the complex evolution patterns of the residuals, something that traditional linear methods struggle to achieve.

[0048] 3. Avoiding the shortcomings of machine learning: Finally, the physical methods are combined with the residual terms of machine learning to reconstruct the corrected forecast field. This architecture ensures that the final correction result is strictly constrained by the physical modes, fundamentally avoiding the physical inconsistencies that may arise from pure "black box" models.

[0049] In summary, this invention not only significantly improves the accuracy of model precipitation forecasts, but also ensures that the forecast results have good physical interpretability and operational stability, and is particularly suitable for the need for refined correction of extended-range and sub-seasonal precipitation forecasts. Attached Figure Description

[0050] Figure 1 This is a flowchart of the residual correction model constructed using historical data in this invention;

[0051] Figure 2 This is a flowchart illustrating the application process of real-time forecast correction in this invention. Detailed Implementation

[0052] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, embodiments of this invention, and should not be construed as limiting the invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention. In the description of this invention, it should be understood that the terminology used is for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0053] The following is combined with Figures 1-2 This invention describes a precipitation forecast correction method that combines EOF projection and U-Net network.

[0054] Example 1: This example provides a precipitation forecast correction method that combines EOF projection and U-Net network, such as... Figure 1 As shown, firstly, the main spatial modes of the precipitation anomaly field are extracted based on historical observation data, and then the model-predicted anomaly field is projected and reconstructed to obtain a spatially meaningful structure. Subsequently, the systematic bias of the model is effectively eliminated by introducing observed climatological states to replace the model climatological states. Furthermore, for the residual terms that could not be explained during the reconstruction process, the U-Net deep learning model is used for modeling and prediction to capture the complex nonlinear error components in the model. Finally, the linear reconstruction results are superimposed with the residual terms predicted by deep learning to form the final precipitation forecast correction results.

[0055] The data source for this embodiment is as follows:

[0056] (1) Reanalysis data (historical precipitation data): Precipitation reanalysis data comes from the fifth generation global climate and atmosphere reanalysis daily dataset (ERA5) provided by the European Centre for Medium-Range Weather Forecasts (ECMWF).

[0057] (2) BCC-CPSv3 model data (forecast data from historical back-calculations of the numerical model): The BCC-CPSv3 model is the S2S forecasting system in the National Climate Center's integrated sub-seasonal-seasonal-interannual scale climate model forecasting operational system. It began quasi-operational operation in 2019 and conducts back-calculation experiments on the past in a dynamic manner. The forecast lead time is 1-60 days.

[0058] In the following data representations, lowercase letters represent physical quantities, while uppercase letters indicate that the physical quantity is a fixed value in the specific calculation steps.

[0059] The method implemented here specifically includes the following steps:

[0060] S1. Processing and analysis of historical precipitation data.

[0061] S11. Collect the observation fields of grid points for all dates during the modeling period (in this embodiment, 2008-2019 is selected as the modeling period). To form a dataset of observation fields ,in, Indicates the time period. Represents latitude grid points, Representing longitude grid points ensures the integrity of the time series and the consistency of its spatial range.

[0062] The data during the modeling period are grouped according to calendar days. Specifically, all data for January 1st of each year are grouped into one group, all data for January 2nd into another, and so on, until all data for December 31st are grouped into a single group, forming 365 independent datasets. For each calendar day group, an arithmetic mean is calculated across spatial grid points, meaning the average of all observations for that date across all years is calculated at each geographic grid point, forming a daily observed climatic mean field. dataset The dataset contains 365 "fields", each representing the most stable and predictable weather component of the year, reflecting climate patterns after removing interannual fluctuations.

[0063] S12. For any specific date in the historical sequence, determine the actual observation field for that date. With the corresponding observed climate mean field Performing grid-by-grid subtraction operations yields the daily high-frequency perturbation components, denoted as the observed anomaly field. The observed anomaly fields for all dates will be obtained. Arranged in chronological order, they form the observed anomaly field matrix. Each "field" in the sequence represents the magnitude and spatial distribution of the daily weather deviation from its long-term climate average, containing high-frequency disturbances at the synoptic scale and low-frequency variation signals at the interannual scale.

[0064] S13, regarding the observed anomaly field matrix Perform EOF decomposition (empirical orthogonal decomposition) to retain the first N standardized, typical, and independent two-dimensional spatial distribution patterns of the anomalies, i.e., the main modes, and the corresponding explained variances (Var1-N). It is necessary to ensure that the cumulative explained variance Var_sum of the first N modes exceeds 75% to represent the main low-frequency spatiotemporal evolution characteristics of the observed anomaly field.

[0065] S2. Processing and analysis of forecast data from historical back-calculations of numerical models.

[0066] S21. Obtain the four-dimensional field from historical back-calculations during the modeling period, and denote it as forecast data. ,in, It indicates the forecast lead time, meaning it can predict the daily precipitation for the next 1 day;

[0067] S22. Because numerical weather prediction models have different systematic errors at different forecast lead times, data needs to be processed separately for each forecast lead time. Therefore, for the forecast lead time (LD), the forecast data calculated from historical data... Extract all forecast data with start times for that forecast lead time to form a three-dimensional dataset of the forecast lead time (LD), denoted as the forecast field. ;

[0068] S23. Align the forecast fields for the forecast lead time (LD) of all reporting times during the modeling period according to the target date to obtain the average climate field for each day of the year under the LD of the forecast lead time, denoted as . ;

[0069] S24. Under the forecast lead time (LD), based on the obtained climate mean field Computable model prediction field Low-frequency stable components (daily climatological field: ) and daily high-frequency disturbance components (daily anomaly field: The high-frequency perturbation components are denoted as the model forecast anomaly field. .

[0070] S3. Model forecast reconstruction based on projection and system drift.

[0071] The purpose of this step is to reconstruct and correct the model's forecast results at each forecast start time T and each forecast lead time LD. The core idea is to decompose the model-predicted anomalous field into the main modes of the observed anomalous field, reconstruct a forecast field that more closely approximates the observed spatial structure, and replace the model's climate mean field with the observed climate mean field to eliminate systematic bias. Specifically:

[0072] S31. Under the forecast lead time (LD) at time point T, the model forecast anomaly field The observed anomaly fields are sequentially projected onto their principal mode spaces to obtain N projection coefficients. These projection coefficients quantify the "weight" or "similarity" of the model-predicted anomaly field on each observed principal spatial distribution. Based on the principal mode space of the observed anomaly field and the projection coefficients, the model-predicted anomaly field is reconstructed. The calculation formula is:

[0073] ;

[0074] in, The first one represents the observed anomaly field. One principal modal space, Indicates the corresponding projection coefficient;

[0075] S32. Since the first N modes in the EOF analysis do not represent 100% of the original variance (e.g., in this embodiment, the cumulative explained variance Var_sum of the first N modes is >75%), the field reconstructed from them is directly used. There will be some attenuation in amplitude. The variance adjustment step is to compensate for the amplification in amplitude, making it match the magnitude of the original observed anomaly field. Therefore, the reconstructed model-predicted anomaly field... Variance adjustment is performed using the following formula:

[0076] ;

[0077] in, To explain the variance proportion coefficient, , For the observation of the anomaly field The explained variance of each principal modal space;

[0078] S33. This step aims to correct the system's mean state bias using the observed climate mean field. The alternative model's own climate mean field fundamentally eliminates the model's bias in understanding the "normal state".

[0079] The variance-adjusted model forecast anomaly field With the observed climate mean field Superimposed, the reconstructed forecast field is obtained. , represented as:

[0080] ;

[0081] S34, Calculate and reconstruct the forecast field With the corresponding observation field The residuals between them are used to obtain the residual terms. ;

[0082] This residual term represents the remaining error that was not explained by the current reconstruction process. It may originate from information contained in the discarded (smaller variance) EOF modes, nonlinear errors, and random errors.

[0083] S35. Calculate the reconstructed forecast field at each time point and forecast lead time. and residuals This yields a set of reconstructed forecast fields with linear relationships. and the set of residual terms representing the nonlinear variation. .

[0084] S4. Residual correction model construction and training.

[0085] This step aims to utilize a deep learning model (U-Net) to learn the residuals generated during the reconstruction of the forecast field, thereby capturing the nonlinear components in the model forecast error. The core process is as follows: train a model independently for each forecast lead time, using the original model forecast anomaly field as input and the reconstructed residual field as the output target.

[0086] The constructed U-Net network adopts an encoder-decoder structure, which includes 4 downsampling and 4 upsampling operations. Each convolution uses a 3x3 convolution kernel and a ReLU activation function, and the feature maps of the corresponding layers of the encoder and decoder are fused through skip connections.

[0087] Using the constructed U-Net network, an independent residual correction model is trained for each forecast lead time (LD). During training, a set of pattern-predicted anomaly fields is used. As input features, the set of residual terms As the prediction target, it is represented as:

[0088] ;

[0089] Sixty residual correction models with different forecast lead times were obtained through training.

[0090] During training, mean squared error (MSE) was used as the loss function, the Adam optimizer (learning rate set to 0.001) was used for parameter optimization, and Dropout (ratio 0.2) was introduced to prevent overfitting.

[0091] Specifically, the period from 2008 to 2019 is used as the training set, and 2020 to 2023 is used as the validation set. (1) Repeat one epoch: Input all the data from the training set into the model in batches, calculate the loss, and backpropagate to update the model parameters. (2) After one epoch: Run the model on the validation set and calculate the evaluation metric MSE. (3) Judge the loss curves of the training set and the validation set: If both the training loss and the validation loss continue to decrease, then the training is complete; otherwise, if there is an overfitting situation where the training loss continues to decrease, but the validation loss starts to rise or stops decreasing, then further adjustments are needed. Optionally, when the validation set loss stops decreasing for several consecutive epochs (e.g., 10), immediately stop training and roll back to the model parameters of the epoch with the best performance on the validation set, or increase the Dropout ratio.

[0092] S5, Final revision of the forecast field.

[0093] like Figure 2 As shown, at any future time T, under each forecast lead time LD, the model forecast anomaly field is calculated. and reconstructing the forecast field Based on the residual correction model corresponding to the forecast lead time, the corresponding residual terms are predicted. The predicted residual terms With reconstructed forecast field By superimposing the data, the corrected precipitation forecast field is obtained. , represented as:

[0094] .

[0095] Example 2: This example provides a non-transitory computer-readable storage medium storing computer instructions that cause a computer to execute a precipitation forecast correction method combining EOF projection and U-Net network. The method includes the following steps:

[0096] S1. Obtain historical precipitation data as the observation field, perform climatological calculations according to calendar days to obtain the daily average observation climate field, and calculate the daily observation anomaly field based on the observed average climate field and historical precipitation data; perform empirical orthogonal function decomposition on the observation anomaly field, extract the first N principal modes, and construct the principal mode space of the observation anomaly field.

[0097] S2. Obtain the historical forecast data from the numerical model, calculate the climate mean field for each forecast lead time, and extract the model forecast anomaly field.

[0098] S3. Project the model-forecasted anomaly field onto the principal mode space of the observed anomaly field to obtain projection coefficients, and reconstruct the model-forecasted anomaly field based on the principal mode space of the observed anomaly field and the projection coefficients; adjust the variance of the reconstructed model-forecasted anomaly field, and superimpose the adjusted model-forecasted anomaly field with the observed climate mean field to obtain the reconstructed forecast field; calculate the residual between the reconstructed forecast field and the corresponding observed field to obtain the residual term;

[0099] S4. Construct a U-Net network, and train the U-Net network with the predicted anomaly field of the model as input and the residual term as the output target to obtain an independent residual correction model under different forecast lead times.

[0100] S5. For any future time period, under each forecast lead time, obtain the model forecast anomaly field and the reconstructed forecast field, and predict the residual terms based on the corresponding residual correction model. Then, superimpose the predicted residual terms with the reconstructed forecast field to obtain the corrected precipitation forecast field.

[0101] Example 3: This example provides an electronic device that may include a processor, a communication interface, a memory, and a communication bus. The processor, communication interface, and memory communicate with each other via the communication bus. The processor can call logical instructions from the memory to execute a precipitation forecast correction method combining EOF projection and the U-Net network. This method includes the following steps:

[0102] S1. Obtain historical precipitation data as the observation field, perform climatological calculations according to calendar days to obtain the daily average observation climate field, and calculate the daily observation anomaly field based on the observed average climate field and historical precipitation data; perform empirical orthogonal function decomposition on the observation anomaly field, extract the first N principal modes, and construct the principal mode space of the observation anomaly field.

[0103] S2. Obtain the historical forecast data from the numerical model, calculate the climate mean field for each forecast lead time, and extract the model forecast anomaly field.

[0104] S3. Project the model-forecasted anomaly field onto the principal mode space of the observed anomaly field to obtain projection coefficients, and reconstruct the model-forecasted anomaly field based on the principal mode space of the observed anomaly field and the projection coefficients; adjust the variance of the reconstructed model-forecasted anomaly field, and superimpose the adjusted model-forecasted anomaly field with the observed climate mean field to obtain the reconstructed forecast field; calculate the residual between the reconstructed forecast field and the corresponding observed field to obtain the residual term;

[0105] S4. Construct a U-Net network, and train the U-Net network with the predicted anomaly field of the model as input and the residual term as the output target to obtain an independent residual correction model under different forecast lead times.

[0106] S5. For any future time period, under each forecast lead time, obtain the model forecast anomaly field and the reconstructed forecast field, and predict the residual terms based on the corresponding residual correction model. Then, superimpose the predicted residual terms with the reconstructed forecast field to obtain the corrected precipitation forecast field.

[0107] Furthermore, the logical instructions in the aforementioned memory can be implemented as software functional units and sold or used as independent products, and can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0108] Example 4: This example provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute a precipitation forecast correction method combining EOF projection and U-Net network. The method includes the following steps:

[0109] S1. Obtain historical precipitation data as the observation field, perform climatological calculations according to calendar days to obtain the daily average observation climate field, and calculate the daily observation anomaly field based on the observed average climate field and historical precipitation data; perform empirical orthogonal function decomposition on the observation anomaly field, extract the first N principal modes, and construct the principal mode space of the observation anomaly field.

[0110] S2. Obtain the historical forecast data from the numerical model, calculate the climate mean field for each forecast lead time, and extract the model forecast anomaly field.

[0111] S3. Project the model-forecasted anomaly field onto the principal mode space of the observed anomaly field to obtain projection coefficients, and reconstruct the model-forecasted anomaly field based on the principal mode space of the observed anomaly field and the projection coefficients; adjust the variance of the reconstructed model-forecasted anomaly field, and superimpose the adjusted model-forecasted anomaly field with the observed climate mean field to obtain the reconstructed forecast field; calculate the residual between the reconstructed forecast field and the corresponding observed field to obtain the residual term;

[0112] S4. Construct a U-Net network, and train the U-Net network with the predicted anomaly field of the model as input and the residual term as the output target to obtain an independent residual correction model under different forecast lead times.

[0113] S5. For any future time period, under each forecast lead time, obtain the model forecast anomaly field and the reconstructed forecast field, and predict the residual terms based on the corresponding residual correction model. Then, superimpose the predicted residual terms with the reconstructed forecast field to obtain the corrected precipitation forecast field.

[0114] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0115] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0116] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A precipitation forecast correction method combining EOF projection and U-Net network, characterized in that, Includes the following steps: S1. Obtain historical precipitation data as the observation field, perform climatological calculations according to calendar days to obtain the daily average observation climate field, and calculate the daily observation anomaly field based on the observed average climate field and historical precipitation data; perform empirical orthogonal function decomposition on the observation anomaly field, extract the first N principal modes, and construct the principal mode space of the observation anomaly field. S2. Obtain the historical forecast data from the numerical model, calculate the climate mean field for each forecast lead time, and extract the model forecast anomaly field. S3. Project the model-forecasted anomaly field onto the principal mode space of the observed anomaly field to obtain projection coefficients, and reconstruct the model-forecasted anomaly field based on the principal mode space of the observed anomaly field and the projection coefficients; adjust the variance of the reconstructed model-forecasted anomaly field, and superimpose the adjusted model-forecasted anomaly field with the observed climate mean field to obtain the reconstructed forecast field; calculate the residual between the reconstructed forecast field and the corresponding observed field to obtain the residual term; specifically: S31. Under the forecast lead time (LD) at time point T, the model forecast anomaly field The projections are sequentially projected onto the principal mode space of the observed anomaly field to obtain N projection coefficients. Based on the principal mode space of the observed anomaly field and the projection coefficients, a mode prediction of the anomaly field is reconstructed. The calculation formula is: ； in, The first one represents the observed anomaly field. One principal modal space, Indicates the corresponding projection coefficient; S32. Anomalies in the reconstructed model forecasts Variance adjustment is performed using the following formula: ； in, To explain the variance proportion coefficient, , For the observation of the anomaly field The explained variance of each principal modal space; S33, The model forecast anomaly field after variance adjustment With the observed climate mean field Superimposed, the reconstructed forecast field is obtained. , represented as: ； S34, Calculate and reconstruct the forecast field With the corresponding observation field The residuals between them are used to obtain the residual terms. ; S35. Calculate the reconstructed forecast field at each time point and forecast lead time. and residuals This yields a set of reconstructed forecast fields with linear relationships. and the set of residual terms representing the nonlinear variation. ; S4. Construct a U-Net network, and train the U-Net network with the predicted anomaly field of the model as input and the residual term as the output target to obtain an independent residual correction model under different forecast lead times. S5. For any future time period, under each forecast lead time, obtain the model forecast anomaly field and the reconstructed forecast field, and predict the residual terms based on the corresponding residual correction model. Then, superimpose the predicted residual terms with the reconstructed forecast field to obtain the corrected precipitation forecast field.

2. The precipitation forecast correction method combining EOF projection and U-Net network according to claim 1, characterized in that, Step S1 is as follows: S11. Collect the observation fields of grid points for all dates during the modeling period. To form a dataset of observation fields ,in, Indicates the time period. Represents latitude grid points, The data is represented by longitude grid points. During the modeling period, the data is grouped by calendar day, forming 365 independent datasets. For each calendar day group, an arithmetic mean is calculated across the spatial grid points, representing the average of all years' observations for that date at each geographic grid point, thus forming a daily observed climate mean field. dataset ; S12. For any specific date in the historical sequence, determine the actual observation field for that date. With the corresponding observed climate mean field By performing grid-by-grid subtraction, the daily observed anomaly fields are obtained. The observed anomaly fields for all dates will be obtained. Arranged in chronological order, they form the observed anomaly field matrix. ; S13, regarding the observed anomaly field matrix Perform EOF decomposition, retain the first N principal modes and their corresponding explained variances, and ensure that the cumulative explained variance of the first N principal modes exceeds 75%.

3. The precipitation forecast correction method combining EOF projection and U-Net network according to claim 2, characterized in that, Step S2 involves acquiring historical forecast data from the numerical model, calculating the climate mean field for each forecast lead time, and extracting the model-forecasted anomalous fields; specifically: S21. Obtain historical back-calculation forecast data during the modeling period. ,in, Indicates the forecast lead time; S22. For forecast lead time (LD), forecast data calculated from historical data... Extract all forecast data with start times for that forecast lead time to form the forecast field for that forecast lead time (LD). ; S23. Align the forecast fields for the forecast lead time (LD) of all reporting times during the modeling period according to the target date to obtain the average climate field for each day of the year under the LD of the forecast lead time, denoted as . ; S24. Under the forecast lead time (LD), based on the obtained climate mean field The high-frequency perturbation components of the separable model are used to obtain the model-predicted anomalous field. .

4. The precipitation forecast correction method combining EOF projection and U-Net network according to claim 3, characterized in that, The U-Net network constructed in step S4 adopts an encoder-decoder structure, which includes 4 downsampling and 4 upsampling operations. Each convolution uses a 3x3 convolution kernel and a ReLU activation function, and the feature maps of the corresponding layers of the encoder and decoder are fused through skip connections. Using the constructed U-Net network, an independent residual correction model is trained for each forecast lead time (LD). During training, a set of pattern-predicted anomaly fields is used. As input features, the set of residual terms As the prediction target, it is represented as: ； Sixty residual correction models with different forecast lead times were obtained through training.

5. The precipitation forecast correction method combining EOF projection and U-Net network according to claim 3, characterized in that, Step S5 is as follows: At any future time T, the model forecast anomaly field is calculated under each forecast lead time LD. and reconstructing the forecast field Based on the residual correction model corresponding to the forecast lead time, the corresponding residual terms are predicted. The predicted residual terms With reconstructed forecast field By superimposing the data, the corrected precipitation forecast field is obtained. , represented as: 。 6. A non-transitory computer-readable storage medium, characterized in that, It stores computer instructions that cause the computer to execute the precipitation forecast correction method combining EOF projection and U-Net network as described in any one of claims 1-5.

7. An electronic device, characterized in that, include: The system includes a processor, a communication interface, a memory, and a communication bus. The processor, communication interface, and memory communicate with each other via the communication bus. The processor calls logical instructions from the memory to execute the precipitation forecast correction method combining EOF projection and U-Net network as described in any one of claims 1-5.

8. A computer program product, characterized in that, The computer program product includes a computer program stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer performs the precipitation forecast correction method combining EOF projection and U-Net network as described in any one of claims 1-5.

Images