Precipitation forecast correction method combining EOF projection and U-Net network

Abstract

The invention discloses a precipitation forecast correction method combining EOF projection and a U-Net network, and the method comprises the steps: extracting a main spatial mode of a precipitation abnormal field based on historical observation data, and carrying out the projection reconstruction of a mode forecast abnormal field on this basis, and obtaining a spatial structure with physical significance; an observation climate state is introduced to replace a mode climate state, so that systematic deviation of the mode is effectively eliminated; for residual terms which are not explained in the reconstruction process, U-Net is adopted for modeling and prediction so as to capture complex nonlinear error components in the mode; and finally, superposing a linear reconstruction result and a residual term of deep learning prediction to form a final rainfall forecast correction result. According to the method, a correction framework combining physical driving and data driving is constructed, and on the basis of keeping the physical interpretability of a traditional EOF method, a deep learning model is introduced to model a residual term, so that dual capture and correction of linear errors and nonlinear errors in mode forecasting are realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of meteorological data processing and machine learning, and in particular to a precipitation forecast correction method combining EOF projection and U-Net network. BACKGROUND

[0002] Precipitation forecast, especially extended-range and sub-seasonal precipitation forecast, has long been a difficult point in meteorological forecast business. Although numerical models can provide information on the evolution of large-scale circulation background, there are often obvious systematic biases and nonlinear errors in the simulation of precipitation, especially in local scale and abnormal weather processes. Therefore, how to effectively post-process and correct the precipitation output of numerical models has become a key link to improve the accuracy of forecasts and the value of business applications.

[0003] For this reason, the industry has proposed methods for correcting forecasts based on empirical orthogonal functions (EOF) or its variants. Such methods decompose the spatial modes of historical observation data and model output, extract the main spatial distribution characteristics, and establish a statistical relationship between observations and models to correct the model forecasts. For example, the patent document with publication number CN115936156A proposes a partition error correction method based on rotated empirical orthogonal function decomposition. The method divides the country into several regions with relatively consistent climate characteristics, and performs EOF modal correction for each region, and finally splices the national forecast results. This method to some extent overcomes the problem of instability of high-order modes in traditional EOF methods, and has strong physical basis and regional adaptability.

[0004] However, the above method still has many limitations in practical application. First, the REOF partition is often large in scope, resulting in a low explained variance of the main mode in subsequent EOF analysis, which requires the introduction of more high-order modes, increasing the computational complexity and limiting its application ability in fine regional (such as provincial scale) business. Second, existing methods are mostly based on original fields for correction, and cannot effectively handle the systematic drift of the model to the climate average, especially in the identification and correction of abnormal precipitation events. In addition, the traditional EOF correction method usually relies on the linear correlation between the main mode time series of observations and models, and approximates the reconstruction of the observation field through projection and regression. This process not only introduces errors, but also requires a high degree of stability in the historical relationship, while the correlation between precipitation and circulation factors has been proven to have significant interdecadal variations, leading to uncertainty in long-term application of the method.

[0005] In recent years, with the development of artificial intelligence technology, some studies attempt to introduce deep learning models to correct the pattern output, and certain effects have been achieved. However, such methods are mostly "end-to-end" black box models, which have strong non-linear fitting ability, but lack physical interpretability, and it is difficult to judge the rationality of the prediction results from the mechanism, especially when facing extreme weather processes not covered by the training data, the generalization ability and stability of the model are difficult to guarantee, which limits its credibility and promotion value in business. SUMMARY

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

[0007] One object of the present application is to provide a precipitation forecast correction method combining EOF projection and U-Net network, by constructing a correction framework combining physical driving and data driving, on the basis of retaining the physical interpretability of the traditional EOF method, a deep learning model is introduced to model the residual term, thereby realizing the dual capture and correction of linear and nonlinear errors in the model prediction.

[0008] In order to achieve the above-mentioned purpose, the present application provides a precipitation forecast correction method combining EOF projection and U-Net network, comprising the following steps: S1, obtaining historical observed precipitation data as an observation field, performing climate state calculation according to calendar day to obtain daily observation climate average field, and calculating daily observation anomaly field based on the observation climate average field and historical observed precipitation data; performing empirical orthogonal function decomposition on the observation anomaly field, extracting the first N principal modes to constitute the principal mode space of the observation anomaly field; S2, obtaining the forecast data of the numerical model historical back-calculation, calculating the climate average field for each forecast time, and extracting the model prediction anomaly field; S3, projecting the model prediction anomaly field into the principal mode space of the observation anomaly field to obtain the projection coefficient, and reconstructing the model prediction anomaly field based on the principal mode space of the observation anomaly field and the projection coefficient; performing variance adjustment on the reconstructed model prediction anomaly field, and superimposing the adjusted model prediction anomaly field and the observation climate average field to obtain the reconstructed prediction field; calculating the residual between the reconstructed prediction field and the corresponding observation field to obtain the residual term; S4, constructing a U-Net network, taking the model prediction anomaly field as input and the residual term as output target, training the U-Net network to obtain independent residual correction models under different forecast times; S5, obtaining the mode prediction anomaly field and the reconstruction prediction field under each prediction lead time for any time in the future, and predicting the residual term based on the corresponding residual correction model, and then superimposing the predicted residual term and the reconstruction prediction field to obtain the corrected precipitation prediction field.

[0009] The further preferred technical solution of the present application is that step S1 is specifically: S11, collecting the observation field of the grid points on all dates in the modeling period to form the data set of the observation field , wherein, represents the time, represents the latitude grid point, represents the longitude grid point; the data in the modeling period is grouped according to the calendar day to form 365 independent data sets, and the arithmetic average operation on the spatial grid points is performed on the data in each calendar day grouping, that is, the average value of the observation value corresponding to all years on each geographical grid point on the date is calculated to form the data set containing the observation climate average field of each day ; S12, for any specific date in the historical sequence, performing a grid-to-grid subtraction operation on the real observation field of the date and the corresponding observation climate average field to obtain the observation anomaly field of each day , and arranging the obtained observation anomaly fields of all dates in chronological order to construct the observation anomaly field matrix ; S13, performing EOF decomposition on the observation anomaly field matrix , retaining the first N principal modes and the corresponding explained variances, and ensuring that the cumulative explained variances of the first N principal modes exceed 75%.

[0010] As a preferred, the step S2 obtains the prediction data of the numerical mode historical back-calculation, and calculates the climate average field and extracts the mode prediction anomaly field for each prediction lead time; specifically: S21, obtaining the prediction data of the historical back-calculation in the modeling period , wherein, represents the prediction lead time; S22, for the prediction lead time LD, extracting the prediction data of the prediction lead time from the prediction data of the historical back-calculation to form the prediction field of the prediction lead time LD ; S23, aligning the prediction field of the prediction lead time LD of all starting times in the modeling period according to the target date to obtain the climate average field of each day in a year corresponding to the prediction lead time LD, denoted as ; S24, Under the forecast time LD, according to the climate mean field The high-frequency disturbance component of the separable model can be used to obtain the model forecast anomaly field .

[0011] Preferably, in step S3, the model prediction anomaly field is projected onto the main modal space of the observed anomaly field to obtain a projection coefficient, and the model prediction anomaly field is reconstructed based on the main modal space of the observed anomaly field and the projection coefficient; specifically, S31, under the forecast validity period LD at time point T, the model forecast abnormal field The anomaly field is projected onto the main modal space of the observed anomaly field in sequence to obtain N projection coefficients, and the model prediction anomaly field is reconstructed based on the main modal space of the observed anomaly field and the projection coefficients. , the calculation formula is: ; in, The first The main modal space, Represents the corresponding projection coefficient.

[0012] Preferably, in step S3, the variance of the reconstructed model forecast anomaly field is adjusted, and the adjusted model forecast anomaly field is superimposed on the observed climate mean field to obtain a reconstructed forecast field; the residual between the reconstructed forecast field and the corresponding observation field is calculated to obtain a residual term; specifically, S32. Forecast abnormal fields based on the reconstructed model Perform variance adjustment, the calculation formula is: ; in, is the coefficient of explained variance, , For observing abnormal fields The explained variance of the principal mode space; S33, the model forecast anomaly field after variance adjustment Compared with the observed climate mean field Superposition to obtain the reconstructed forecast field , expressed as: ; S34. Calculate and reconstruct the forecast field Corresponding observation field The residual between ; S35. Calculate the reconstructed forecast field at each time point and forecast time and the residual , we get the reconstructed forecast field set with linear relationship and the residual term set representing the nonlinear change part .

[0013] Preferably, the U-Net network constructed in step S4 adopts an encoder-decoder structure, includes 4 downsampling and 4 upsampling operations, each convolution uses a 3x3 convolution kernel and a ReLU activation function, and fuses the feature maps of the corresponding layers of the encoder and decoder through jump connections; Using the constructed U-Net network, an independent residual correction model is trained for each forecast time LD. During training, the set of model forecast anomaly fields is used. As input features, the residual term set As the prediction target, it is expressed as: ; 60 residual correction models with different forecast time limits were obtained through training.

[0014] Preferably, step S5 is specifically as follows: At any time T in the future, under each forecast time LD, the model forecast anomaly field is calculated and reconstructed forecast fields , and based on the residual correction model of the corresponding forecast time, the corresponding residual term is predicted , the predicted residual and reconstructed forecast fields Superposition to obtain the revised precipitation forecast field , expressed as: .

[0015] Another aspect of the present invention provides a non-transitory computer-readable storage medium having computer instructions stored thereon, which enable a computer to execute the above-mentioned precipitation forecast correction method combining EOF projection and U-Net network.

[0016] Another aspect of 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 via the communication bus, and the processor calls logic instructions in the memory to execute the above-mentioned precipitation forecast correction method combining EOF projection and U-Net network.

[0017] On the other hand, the present invention provides a computer program product, which includes a computer program stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer executes the above-mentioned precipitation forecast correction method combining EOF projection and U-Net network.

[0018] Beneficial effects: The precipitation forecast correction method combining EOF projection and U-Net network of the present application firstly extracts the main spatial mode of precipitation anomaly field based on historical observation data, and on this basis, projects and reconstructs the mode forecast anomaly field to obtain a spatial structure with physical significance; then, by introducing the observed climate state to replace the mode climate state, the systematic bias of the mode is effectively eliminated; further, for the residual term that cannot be explained in the reconstruction process, a U-Net deep learning model is used for modeling and prediction to capture the complex nonlinear error components in the mode; finally, the linear reconstruction result and the residual term predicted by the deep learning are superimposed to form the final precipitation forecast correction result. Compared with the prior art, the advantages of the present application are: 1. Stability of physical method: The present application projects the high-dimensional and complex original error field onto a low-dimensional and physically meaningful characteristic space through empirical orthogonal function (EOF) decomposition. In addition, the observed climate state is used to replace the mode climate state, and the two-step approximation in the traditional projection method is simplified, reducing error propagation and effectively avoiding the interdecadal stability problem of correlation. This step lays the foundation for the physical interpretability of the method and significantly reduces the dimensionality of subsequent processing.

[0019] 2. Strong ability of machine learning: For the residual term generated by the statistical method model, a machine learning model is used for modeling and prediction. This is aimed at using the strong nonlinear fitting ability and automatic feature mining advantage to accurately capture the complex evolution law of the residual term, which is difficult to achieve by traditional linear methods.

[0020] 3. Avoiding the short board of machine learning: Finally, the physical method and the residual term of machine learning are combined to reconstruct the corrected forecast field. This architecture ensures that the final correction result is strictly constrained by the physical mode, fundamentally avoiding the physical unreasonable phenomena that may be produced by pure "black box" models.

[0021] In summary, the present application not only significantly improves the accuracy of mode precipitation forecast, but also ensures that the forecast result has good physical interpretability and business stability, and is especially suitable for fine correction of extended period and sub-seasonal scale precipitation forecasts. BRIEF DESCRIPTION OF DRAWINGS

[0022] Fig. 1 is the flow chart for constructing a residual correction model with historical data in the present application; Fig. 2 is the application flow chart of real-time forecast correction of the present application. DETAILED DESCRIPTION

[0023] In order to make the purpose, technical solutions and advantages of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below in conjunction with the drawings in the present invention. Obviously, the embodiments described are part of the embodiments of the present invention, not all of the embodiments, and they should not be understood as limitations on the present invention. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention. In the description of the present invention, it should be understood that the terms used are only for descriptive purposes and cannot be understood as indicating or implying relative importance.

[0024] The following combination Figs. 1-2 The present invention describes a precipitation forecast correction method combining EOF projection and U-Net network.

[0025] Example 1: This example provides a precipitation forecast correction method combining EOF projection and U-Net network, such as Fig. 1 As shown in the figure, first, the main spatial modes of the precipitation anomaly field are extracted based on historical observation data, and on this basis, the model forecast anomaly field is projected and reconstructed to obtain a spatial structure with physical significance; then, by introducing the observed climate state to replace the model climate state, the systematic bias of the model is effectively eliminated; further, for the residual terms that cannot be explained in 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 result is superimposed with the residual term predicted by deep learning to form the final precipitation forecast correction result.

[0026] The data sources for this embodiment are as follows: (1) Reanalysis data (historical observed precipitation data): The precipitation reanalysis data come from the fifth generation of global climate atmospheric reanalysis daily dataset (ERA5) provided by the European Centre for Medium-Range Weather Forecasts (ECMWF).

[0027] (2) BCC-CPSv3 model data (forecast data from historical back-calculation of numerical models): The BCC-CPSv3 model is the S2S prediction system within the National Climate Center's subseasonal-seasonal-interannual scale integrated climate model prediction operational system. It began quasi-operational operation in 2019 and conducts dynamic back-calculation experiments on the past. The forecast timeframe is 1-60 days.

[0028] In the following data representation, lowercase English represents physical quantities, and uppercase English represents that the physical quantity is a fixed value in a specific calculation step.

[0029] The method of this implementation specifically includes the following steps: S1. Processing and analysis of historical precipitation data.

[0030] S11. Collect the observation fields of the grid points for all dates during the modeling period (in this embodiment, 2008-2019 is selected as the modeling period) , forming a data set of the observation field ,in, Indicates the time, represents the latitude grid, Represents the longitude grid points to ensure the time series is complete and the spatial range is consistent.

[0031] The data of the modeling period are grouped by calendar day. Specifically, all the data of January 1st of each year in the modeling period are grouped into one group, all the data of January 2nd are grouped into another group, and so on, until all the data of December 31st are grouped into one group, forming 365 independent data sets. The arithmetic mean operation on the spatial grid is performed on the data in each calendar day group, that is, the average value of the observations corresponding to all years on that date is calculated at each geographical grid point, forming a daily observation climate mean field. Dataset The dataset contains 365 "fields", each of which represents the most stable and predictable weather component of the year, reflecting the climate laws after eliminating interannual fluctuations.

[0032] S12. For any specific date in the historical sequence, the actual observation field of that date The corresponding observed climate mean field Perform grid-to-grid subtraction to obtain the daily high-frequency disturbance component, which is recorded as the observed anomaly field. , the observed anomaly fields of all dates will be obtained Arrange in chronological order and construct an observation anomaly field matrix Each "field" in the sequence represents the amplitude and spatial distribution of the deviation of the weather on that day from its long-term climate average state, including high-frequency disturbances on the weather scale and low-frequency change signals on the interannual scale.

[0033] S13, the observed abnormal field matrix Perform EOF decomposition (empirical orthogonal decomposition) and retain the first N standardized typical and independent two-dimensional spatial distribution patterns of anomalies, namely 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.

[0034] S2. Processing and analysis of forecast data from historical back-calculations of numerical models.

[0035] S21. Obtain the four-dimensional field calculated historically during the modeling period and record it as forecast data ,in, represents the forecast lead time, i.e. the number of days in the future that can be predicted; S22, since the numerical prediction model has different systematic errors at different forecast lead times, it is necessary to process the data separately for each forecast lead time. Therefore, for the forecast lead time LD, the forecast data from the historical back-calculation is extracted for all the start times of the forecast to form a three-dimensional data set for the forecast lead time LD, denoted as the forecast field ; S23, the forecast field for all the start times of the forecast in the modeling period is aligned according to the target date to obtain the climatological average field corresponding to each day of the year under the forecast lead time LD, denoted as ; S24, under the forecast lead time LD, the obtained climatological average field can be used to calculate the low-frequency stable component (daily climatological field: ) and the daily high-frequency perturbation component (daily anomaly field: ) of the model prediction field , and the high-frequency perturbation component is denoted as the model prediction anomaly field .

[0036] S3, reconstruction of the model prediction based on projection and systematic drift.

[0037] The purpose of this step is to reconstruct and correct the prediction results of the model under each start time T and each forecast lead time LD. The core idea is to decompose the anomaly field of the model prediction into the principal modes of the observed anomaly field, to reconstruct a prediction field that is closer to the structure of the observation space, and to replace the climatological average field of the model with the observed climatological average field to eliminate the systematic bias. Specifically: S31, under the forecast lead time LD at the time point T, the model prediction anomaly field is projected into the principal mode space of the observed anomaly field in sequence to obtain N projection coefficients, which quantify the "weight" or "similarity" of the anomaly field of the model prediction in each principal spatial distribution of the observed anomaly field. The model prediction anomaly field is reconstructed based on the principal mode space of the observed anomaly field and the projection coefficients, and the calculation formula is: ; wherein represents the th principal mode space of the observed anomaly field, represents the corresponding projection coefficient; S32, since in the EOF analysis, the first N modes do not represent 100% of the original variance (for example, in this embodiment, the first N modes cumulatively explain more than 75% of the variance Var_sum), the field The amplitude will be attenuated. The variance adjustment step is to amplify the amplitude to make it match the magnitude of the original observed anomaly field. Perform variance adjustment, the calculation formula is: ; in, is the coefficient of explained variance, , For observing abnormal fields The explained variance of the principal mode space; S33, This step aims to correct the system mean state bias of the model. Replacing the model's own climate mean field fundamentally eliminates the model's deviation in understanding the "normal state".

[0038] The model forecast anomaly field after variance adjustment Compared with the observed climate mean field Superposition to obtain the reconstructed forecast field , expressed as: ; S34. Calculate and reconstruct the forecast field Corresponding observation field The residual between ; The residual term represents the remaining error that is not explained by the reconstruction process. It may come from the information contained in the discarded EOF mode (with a small variance), nonlinear errors, random errors, etc.

[0039] S35. Calculate the reconstructed forecast field at each time point and forecast time and the residual , we get the reconstructed forecast field set with linear relationship and the residual term set representing the nonlinear change part .

[0040] S4. Residual correction model construction and training.

[0041] This step aims to use a deep learning model (U-Net) to learn the residuals generated during the above-mentioned forecast field reconstruction process, thereby capturing the nonlinear components of the model forecast error. The core process is to independently train a model for each forecast time, using the original model forecast anomaly field as input and the reconstructed residual field as the output target.

[0042] The constructed U-Net network adopts an encoder-decoder structure, contains 4 times of downsampling and 4 times of upsampling operations, each convolution uses a 3x3 convolution kernel and a ReLU activation function, and the feature maps of the corresponding levels of the encoder and the decoder are fused through a jump connection; Using the constructed U-Net network, an independent residual correction model is trained for each prediction lead time LD. During training, the set of abnormal fields predicted by the model is used as the input feature The residual term set is used as the input feature As the prediction target, it is expressed as: ; 60 residual correction models for different prediction lead times are trained.

[0043] During training, the mean square error (MSE) is used as the loss function, the Adam optimizer (learning rate set to 0.001) is used for parameter optimization, and Dropout (ratio 0.2) is introduced to prevent overfitting.

[0044] Specifically, 2008-2019 is used as the training set and 2020-2023 is used as the validation set. (1) Loop one epoch: input all data in the training set into the model in batches, calculate the loss, and update the model parameters through back propagation. (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 the training loss and the validation loss continue to decrease, the training is complete; otherwise, if the training loss continues to decrease but the validation loss starts to rise or no longer decreases, overfitting occurs and further adjustment is needed. Optionally, when the validation set loss no longer decreases for consecutive epochs (such as 10), stop training immediately and roll back to the model parameters of the epoch with the best performance on the validation set, or increase the Dropout ratio.

[0045] S5, final correction of the prediction field.

[0046] As Fig. 2 shown, at any time T in the future, the model-predicted abnormal field and the reconstructed prediction field are calculated for each prediction lead time LD, and the corresponding residual term is predicted based on the residual correction model for the corresponding prediction lead time. The predicted residual term is superimposed on the reconstructed prediction field to obtain the corrected precipitation prediction field .

[0047] Embodiment 2: The embodiment provides a non-transitory computer readable storage medium having stored thereon computer instructions for causing a computer to execute a precipitation forecast correction method combining EOF projection and a U-Net network, the method comprising the following steps: S1, obtaining historical observed precipitation data as an observation field, performing climate state calculation according to calendar days to obtain a daily observation climate average field, and calculating a daily observation anomaly field based on the observation climate average field and the historical observed precipitation data; performing empirical orthogonal function decomposition on the observation anomaly field, extracting the first N principal modes to constitute a principal mode space of the observation anomaly field; S2, obtaining numerical model historical back-calculation prediction data, calculating a climate average field for each prediction time, and extracting a model prediction anomaly field; S3, projecting the model prediction anomaly field to the principal mode space of the observation anomaly field to obtain projection coefficients, and reconstructing the model prediction anomaly field based on the principal mode space of the observation anomaly field and the projection coefficients; performing variance adjustment on the reconstructed model prediction anomaly field, superimposing the adjusted model prediction anomaly field and the observation climate average field to obtain a reconstructed prediction field; calculating the residual between the reconstructed prediction field and the corresponding observation field to obtain a residual term; S4, constructing a U-Net network, taking the model prediction anomaly field as input, and taking the residual term as output target, training the U-Net network to obtain independent residual correction models under different prediction times; S5, for any time in the future, obtaining the model prediction anomaly field and the reconstructed prediction field under each prediction time, predicting the residual term based on the corresponding residual correction model, and superimposing the predicted residual term and the reconstructed prediction field to obtain a corrected precipitation prediction field.

[0048] Embodiment 3: The embodiment provides an electronic device, which can include a processor, a communications interface, a memory and a communications bus, wherein the processor, the communications interface and the memory complete mutual communication through the communications bus. The processor can invoke logical instructions in the memory to execute a precipitation forecast correction method combining EOF projection and a U-Net network, the method comprising the following steps: S1, obtaining historical observed precipitation data as an observation field, performing climate state calculation according to calendar days to obtain a daily observation climate average field, and calculating a daily observation anomaly field based on the observation climate average field and the historical observed precipitation data; performing empirical orthogonal function decomposition on the observation anomaly field, extracting the first N principal modes to constitute a principal mode space of the observation anomaly field; S2, acquire the forecast data of numerical mode history back-calculation, calculate the climate average field for each forecast time, and extract the mode forecast anomaly field; S3, project the mode forecast anomaly field to the principal modal space of the observation anomaly field to obtain the projection coefficient, reconstruct the mode forecast anomaly field based on the principal modal space of the observation anomaly field and the projection coefficient, perform variance adjustment on the reconstructed mode forecast anomaly field, superimpose the adjusted mode forecast anomaly field and the observation climate average field to obtain the reconstructed forecast field, and calculate the residual between the reconstructed forecast field and the corresponding observation field to obtain the residual term; S4, construct a U-Net network, input the mode forecast anomaly field, and take the residual term as the output target to train the U-Net network to obtain an independent residual correction model under different forecast times; S5, for any time in the future, obtain the mode forecast anomaly field and the reconstructed forecast field under each forecast time, predict the residual term based on the corresponding residual correction model, and superimpose the predicted residual term and the reconstructed forecast field to obtain the corrected precipitation forecast field.

[0049] In addition, the logical instructions in the above-mentioned memory can be realized in the form of a software functional unit and sold or used as an independent product, and can be stored in a computer readable storage medium. Based on such understanding, the technical solutions of the present application essentially or the parts that contribute to the prior art or parts of the technical solutions can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes a plurality of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the method described in the embodiments of the present application. The aforementioned storage medium includes: a U disk, a mobile hard disk, a read-only memory (ROM, Read-Only Memory), a random access memory (RAM, Random Access Memory), a magnetic disk or an optical disk, and various media that can store program codes.

[0050] Embodiment 4: The present embodiment provides a computer program product, which comprises a computer program, the computer program can be stored on a non-transient computer readable storage medium, and the computer program can be executed by a processor to enable a computer to execute a precipitation forecast correction method combined with EOF projection and U-Net network, the method comprising the following steps: S1, acquire historical observation precipitation data as an observation field, perform climate state calculation according to calendar days to obtain a daily observation climate average field, and calculate a daily observation anomaly field based on the observation climate average field and the historical observation precipitation data; perform empirical orthogonal function decomposition on the observation anomaly field, extract the first N principal modes to constitute the principal modal space of the observation anomaly field; S2, acquire the forecast data of the numerical mode history backcast, calculate the climate average field for each forecast lead time, and extract the mode forecast anomaly field; S3, project the mode forecast anomaly field to the principal modal space of the observation anomaly field to obtain the projection coefficient, reconstruct the mode forecast anomaly field based on the principal modal space of the observation anomaly field and the projection coefficient, perform variance adjustment on the reconstructed mode forecast anomaly field, superimpose the adjusted mode forecast anomaly field and the observation climate average field to obtain the reconstructed forecast field, and calculate the residual between the reconstructed forecast field and the corresponding observation field to obtain the residual term; S4, construct a U-Net network, input the mode forecast anomaly field, and take the residual term as the output target to train the U-Net network and obtain an independent residual correction model under different forecast lead times; S5, for any future time, obtain the mode forecast anomaly field and the reconstructed forecast field under each forecast lead time, predict the residual term based on the corresponding residual correction model, and superimpose the predicted residual term and the reconstructed forecast field to obtain the corrected precipitation forecast field.

[0051] The device embodiments described above are only schematic, wherein the units described as separate components can or can not be physically separate, and the components displayed as units can or can not be physical units, i.e., can be located in one place or distributed on multiple network units. Part or all of the modules can be selected according to actual needs to achieve the purpose of the present embodiment scheme. Those skilled in the art can understand and implement without creative labor.

[0052] From the above description of the embodiments, those skilled in the art can clearly understand that the embodiments can be realized by means of software plus necessary general hardware platforms, and of course can also be realized by hardware. Based on such understanding, the above technical solutions can be embodied in the form of software products, which can be stored in a computer readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and include a number of instructions to make a computer device (which can be a personal computer, a server, or a network device, etc.) execute the methods described in each embodiment or some parts of the embodiments.

[0053] It should be pointed out finally that the above embodiments are only used to illustrate the technical solutions of the present application, but not to limit the same; and although the present application has been described in detail with reference to the foregoing embodiments, it should be appreciated by those skilled in the art that the technical solutions recorded in the foregoing embodiments can be modified, or some technical features thereof can be replaced equivalently; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A precipitation forecast correction method combining EOF projection and U-Net network, characterized in that, Comprising the following steps: S1, obtaining historical observed precipitation data as an observation field, calculating a climate average field on a calendar day, obtaining a daily observation climate average field, and calculating a daily observation anomaly field based on the observation climate average field and the historical observed precipitation data; performing empirical orthogonal function decomposition on the observation anomaly field, extracting the first N principal modes to form a principal mode space of the observation anomaly field; S2, obtaining numerical model historical back-calculation prediction data, calculating a climate average field for each prediction time, and extracting a model prediction anomaly field; S3, projecting the model prediction anomaly field into the principal mode space of the observation anomaly field to obtain projection coefficients, reconstructing the model prediction anomaly field based on the principal mode space of the observation anomaly field and the projection coefficients, adjusting the variance of the reconstructed model prediction anomaly field, and superimposing the adjusted model prediction anomaly field and the observation climate average field to obtain a reconstructed prediction field; calculating the residual between the reconstructed prediction field and the corresponding observation field to obtain a residual term; S4, constructing a U-Net network, inputting the model prediction anomaly field, and outputting the residual term as a target, training the U-Net network to obtain an independent residual correction model under different prediction times; S5, for any future time, obtaining the model prediction anomaly field and the reconstructed prediction field under each prediction time, predicting the residual term based on the corresponding residual correction model, and superimposing the predicted residual term and the reconstructed prediction field to obtain a corrected precipitation prediction field.

2. The precipitation forecast correction method combining EOF projection and U-Net network according to claim 1, wherein, Step S1 is specifically: S11, collecting the observation field of all grid points on all dates in the modeling period , forming a data set of observation fields , wherein, represents the time, represents the latitude grid point, represents the longitude grid point; grouping the data in the modeling period according to the calendar day to form 365 independent data sets, performing an arithmetic average operation on the data in each calendar day grouping on the spatial grid point, that is, calculating the average value of the observation value on each geographical grid point corresponding to all years on this date to form a data set containing the observation climate average field on a daily basis ;​ S12. For any specific date in the historical sequence, the actual observation field of that date The corresponding observed climate mean field Perform grid-to-grid subtraction to obtain the daily observed anomaly field. , the observed anomaly fields of all dates will be obtained Arrange in chronological order and construct an observation anomaly field matrix ; S13, to the observed anomaly field matrix EOF decomposition is performed, the first N principal modes and the corresponding explained variance are retained, and the cumulative explained variance of the first N principal modes is ensured to exceed 75%.

3. The precipitation forecast correction method combining EOF projection and U-Net network of claim 2, wherein, Step S2 is specifically: S21, obtaining the forecast data of historical back-calculation in the modeling period wherein, denotes the forecast lead time; S22, for the forecast age LD, extract from the forecast data backcasted from history all the forecast data for this forecast age, form the forecast field of the forecast age LD ;​ S23, aligning the prediction field of the prediction time limit LD in all the reported time in the modeling period according to the target date to obtain the climate average field corresponding to each day in a year under the prediction time limit LD, denoted as ; S24, under the forecast lead time LD, according to the obtained climate mean field The high-frequency disturbance component of the separable mode is obtained, and a mode prediction anomaly field is obtained .

4. The precipitation forecast correction method combining EOF projection and U-Net network of claim 3, wherein, In step S3, the model prediction anomaly field is projected into the principal mode space of the observation anomaly field to obtain projection coefficients, and the model prediction anomaly field is reconstructed based on the principal mode space of the observation anomaly field and the projection coefficients; specifically: S31, projecting the mode forecast anomaly field to the principal modal space of the observation anomaly field to obtain N projection coefficients, and reconstructing the mode forecast anomaly field based on the principal modal space of the observation anomaly field and the projection coefficients , and the calculation formula is:​ ； wherein, represents the first principal modal space of the observed anomaly field, represents the corresponding projection coefficients.

5. The precipitation forecast correction method combining EOF projection and U-Net network according to claim 4, characterized in that, In step S3, the variance of the reconstructed model prediction anomaly field is adjusted, and the adjusted model prediction anomaly field is superimposed with the observation climate average field to obtain a reconstructed prediction field; the residual between the reconstructed prediction field and the corresponding observation field is calculated to obtain a residual term; specifically: S32, predict an abnormal field for the reconstructed pattern Variance adjustment is performed, and the calculation formula is: ； in, is the coefficient of explained variance, , For observing abnormal fields The explained variance of the principal mode space; S33, the mode forecast anomaly field after variance adjustment is performed and the observed climate mean field superimposed to obtain a reconstructed forecast field is expressed as: ； S34, computing a reforecast field between the corresponding observation field obtaining a residual term ; S35. Calculate the reconstructed forecast fields at each time point and forecast lead time and the residual terms to obtain the set of reconstructed forecast fields with linear relationship and the set of residual terms representing the non-linearly changed parts .

6. The precipitation forecast correction method combining EOF projection and U-Net network according to claim 5, characterized in that, The U-Net network constructed in step S4 adopts an encoder-decoder structure, 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 the decoder are fused through a jump connection; Using the constructed U-Net network, an independent residual correction model is trained for each prediction lead time LD. During training, the set of abnormal fields predicted by the model is used as input features The residual term set is used as input features The prediction target is represented as: ； 60 residual correction models under different prediction times are trained.

7. The precipitation forecast correction method combining EOF projection and U-Net network of claim 5, wherein, Step S5 is specifically: At any time T in the future, the model prediction anomaly field is calculated at each lead time LD and the reconstructed prediction field and the corresponding residual term is predicted based on the residual correction model for the corresponding lead time The predicted residual term is superimposed on the reconstructed prediction field to obtain the corrected precipitation prediction field , which is expressed as: 。 8. A non-transitory computer-readable storage medium, comprising: It has computer instructions stored thereon, which enable the computer to execute the precipitation prediction correction method combining EOF projection and U-Net network according to any one of claims 1-7.

9. An electronic device, comprising: It comprises: 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 the logic instructions in the memory to execute the precipitation prediction correction method combining EOF projection and U-Net network according to any one of claims 1-7.

10. A computer program product, characterised in that, The computer program product comprises a computer program stored on a non-transitory computer readable storage medium, and the computer program, when executed by a processor, causes the computer to perform the precipitation forecast correction method combining EOF projection and U-Net network according to any one of claims 1-7.

Images