Plateau precipitation prediction method, system and device based on intelligent similarity correction
By integrating multivariate information and using a bi-branch deep residual network and attention mechanism, the problem of difficulty in characterizing multivariate nonlinear coupling relationships in existing technologies is solved, thereby improving the accuracy of climate prediction and the precision of precipitation prediction.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- CHINESE ACAD OF METEOROLOGICAL SCI
- Filing Date
- 2026-03-02
- Publication Date
- 2026-07-03
AI Technical Summary
Existing technologies using single variables and linear measures cannot fully characterize the nonlinear coupling relationships between multiple variables, especially the nonlinear impact of topography on climate, leading to misjudgments and correction biases in similar years, which seriously affects the accuracy of climate prediction.
A method based on intelligent similarity correction is adopted. By acquiring multivariate historical factor data, a bi-branch deep residual network and attention mechanism are used to integrate multi-source variables such as wind field, height field and temperature to form a shared high-dimensional feature space. The interaction terms and high-order coupling features between variables are automatically learned to characterize the nonlinear synergistic relationship.
It improves the accuracy of similarity selection and seasonal precipitation forecasting, enhances the ability to characterize the driving mechanisms of precipitation anomalies in the forecast area, and reduces the bias of error correction.
Smart Images
Figure CN121765403B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of short-term climate prediction technology, and in particular to a method, system and device for predicting precipitation in high-altitude areas based on intelligent similarity correction. Background Technology
[0002] Current mainstream international climate dynamic models struggle to capture the nonlinear interactions between steep terrain and local circulation, resulting in significantly lower precipitation forecast skill. In the field of climate prediction, similarity error correction is a commonly used method to effectively improve the forecasting skill of dynamic models. Its basic principle is to find historical similarities to the current initial climate conditions and use the error distribution information from the model's historical predictions provided by these similar conditions to correct the model's current forecast results.
[0003] However, traditional error correction methods use single variables and linear measures, making it difficult to fully characterize the nonlinear coupling relationships between multiple variables, especially the nonlinear impact of topography on climate. This often leads to misjudgments and correction biases in similar years, seriously affecting the accuracy of climate predictions. Summary of the Invention
[0004] This invention provides a plateau precipitation prediction method, system, and device based on intelligent similarity correction. It addresses the shortcomings of existing technologies that use single variables and linear metrics, making it difficult to comprehensively characterize the nonlinear coupling relationships between multiple variables, especially the nonlinear impact of topography on climate. This often leads to misjudgments of similar years and correction biases, severely affecting the accuracy of climate prediction. The invention integrates multivariate information, more deeply explores their nonlinear synergistic relationships, and enhances the ability to characterize the driving mechanisms of precipitation anomalies in the predicted area. Compared to traditional single-variable or linear schemes, it effectively improves the accuracy of similarity selection and seasonal precipitation prediction.
[0005] This invention provides a plateau precipitation prediction method based on intelligent similarity correction, comprising the following steps:
[0006] Obtain model precipitation data for the month to be predicted, along with its corresponding multivariate historical factor data and global historical annual precipitation error data. The multivariate historical factor data includes gridded data of wind field, geopotential height field, and temperature.
[0007] The gridded data of wind field, altitude field and temperature are standardized to obtain standard gridded data of wind field, altitude field and temperature;
[0008] The effective historical annual precipitation error data corresponding to the effective geographic region is extracted from the global historical annual precipitation error data using a predefined regional mask; the effective geographic region includes plateau regions;
[0009] The standard grid data and the effective historical annual precipitation error data are input into a pre-trained error prediction model to obtain the prediction error field output by the error prediction model. The prediction error field is used to characterize the error between the model precipitation data and the observed precipitation data in the plateau region. The prediction error field is obtained by the error prediction model through a bi-branch deep residual network and an attention mechanism, based on the feature similarity between the multivariate historical factor data features corresponding to the month to be predicted and the multivariate historical factor data features corresponding to the historical year, as well as the weights corresponding to the feature similarity.
[0010] Based on the prediction error field, the model precipitation data is corrected to obtain the final predicted precipitation.
[0011] This invention also provides a plateau precipitation prediction system based on intelligent similarity correction, comprising the following modules:
[0012] The acquisition module is used to acquire model precipitation data for the month to be predicted, as well as its corresponding multivariate historical factor data and global historical annual precipitation error data. The multivariate historical factor data includes gridded data of wind field, geopotential height field and temperature.
[0013] The data preprocessing module is used to standardize the gridded data of the wind field, altitude field, and temperature to obtain standard gridded data of the wind field, altitude field, and temperature.
[0014] The extraction module is also used to extract effective historical annual precipitation error data corresponding to effective geographical regions from the global historical annual precipitation error data by applying a predefined regional mask; the effective geographical regions include plateau regions;
[0015] An error prediction module is used to input the standard grid data and the effective historical annual precipitation error data into a pre-trained error prediction model to obtain a prediction error field output by the error prediction model. The prediction error field is used to characterize the error between the model precipitation data and the observed precipitation data in the plateau region. The prediction error field is obtained by the error prediction model through a dual-branch deep residual network and an attention mechanism, based on the feature similarity between the multivariate historical factor data features corresponding to the month to be predicted and the multivariate historical factor data features corresponding to the historical year, as well as the weights corresponding to the feature similarity.
[0016] An error correction module is used to correct the model precipitation data based on the prediction error field to obtain the final predicted precipitation.
[0017] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the plateau precipitation prediction method based on intelligent similarity correction as described above.
[0018] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the plateau precipitation prediction method based on intelligent similarity correction as described above.
[0019] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the plateau precipitation prediction method based on intelligent similarity correction as described above.
[0020] The plateau precipitation prediction method, system, and device based on intelligent similarity correction provided by this invention acquires model precipitation data for the month to be predicted, along with its corresponding multivariate historical factor data and global historical annual precipitation error data. It then standardizes the gridded data of wind field, altitude field, and temperature to obtain standard gridded data for these fields. Using a predefined regional mask, it extracts effective historical annual precipitation error data corresponding to effective geographical regions from the global historical annual precipitation error data. The standard gridded data and effective historical annual precipitation error data are input into a pre-trained error prediction model to obtain the prediction error field output by the model. Based on this prediction error field, the model precipitation data is corrected to obtain the final predicted precipitation amount. Thus, by using an error prediction model, multiple source variables such as wind field, geopotential height, and temperature are uniformly mapped to a shared high-dimensional feature space. In this space, the interaction terms and high-order coupling features between variables are automatically learned through multi-layer nonlinear transformations to form a nonlinear feature vector that characterizes the overall climate state. This integrates multivariate information, explores their nonlinear synergistic relationships more deeply, and enhances the ability to characterize the driving mechanism of precipitation anomalies in the predicted area. Compared with traditional single-variable or linear schemes, this effectively improves the accuracy of similarity selection and seasonal precipitation prediction. Attached Figure Description
[0021] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.
[0022] Figure 1 This is a flowchart illustrating the plateau precipitation prediction method based on intelligent similarity correction provided by the present invention.
[0023] Figure 2This is a schematic diagram of the plateau precipitation prediction system based on intelligent similarity correction provided by the present invention.
[0024] Figure 3 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation
[0025] 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, of the embodiments of this 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.
[0026] The following is combined Figure 1 The present invention describes a plateau precipitation prediction method based on intelligent similarity correction. This method is applicable to precipitation prediction in any region, especially in high-altitude and complex terrain regions. The execution subject of this method can be an electronic device or a plateau precipitation prediction method based on intelligent similarity correction installed in the electronic device. The plateau precipitation prediction system based on intelligent similarity correction can be implemented by software, hardware, or a combination of both.
[0027] Figure 1 This is a flowchart illustrating the plateau precipitation prediction method based on intelligent similarity correction provided by the present invention, as shown below. Figure 1 As shown, the method includes the following:
[0028] Step 101: Obtain the model precipitation data for the month to be predicted, along with its corresponding multivariate historical factor data and global historical annual precipitation error data.
[0029] The multivariate historical factor data includes gridded data of wind field, geopotential height field, and temperature.
[0030] Here, the global historical annual precipitation error data is the error value between the monthly observed precipitation data and the model precipitation data in the historical year. The global historical annual precipitation error data can be calculated based on the observed precipitation data and the model precipitation data, or it can be downloaded from the database.
[0031] Step 102: Standardize the gridded data of the wind field, altitude field and temperature to obtain standard gridded data of the wind field, altitude field and temperature.
[0032] Here, standardization processing can include declimatization processing, normalization processing, etc.
[0033] Step 103: Apply a predefined region mask to extract the effective historical annual precipitation error data corresponding to the effective geographical region from the global historical annual precipitation error data.
[0034] The effective geographical area includes plateau regions.
[0035] Here, the predefined regional mask refers to masking invalid data such as precipitation error data and historical factor data outside the area to be predicted.
[0036] Step 104: Input the standard grid data and the effective historical annual precipitation error data into the pre-trained error prediction model to obtain the prediction error field output by the error prediction model.
[0037] The prediction error field is used to characterize the error between the model precipitation data and the observed precipitation data in the plateau region. The prediction error field is obtained by the error prediction model through a bi-branch deep residual network and an attention mechanism, which determines the feature similarity between the multivariate historical factor data features corresponding to the month to be predicted and the multivariate historical factor data features corresponding to the historical year, as well as the weights corresponding to the feature similarity.
[0038] Step 105: Based on the prediction error field, correct the model precipitation data to obtain the final predicted precipitation.
[0039] The methods for determining similarity include, but are not limited to, Euclidean distance, cosine similarity, Manhattan distance, and other methods.
[0040] In this embodiment of the invention, an error prediction model is used to uniformly map multiple source variables such as wind field, geopotential height, and temperature to a shared high-dimensional feature space. In this space, the interaction terms and high-order coupling features between variables are automatically learned through multi-layer nonlinear transformation to form a nonlinear feature vector that characterizes the "overall climate state". This integrates multivariate information, explores their nonlinear synergistic relationships more deeply, and enhances the ability to characterize the driving mechanism of precipitation anomalies in the predicted area. Compared with traditional single-variable or linear schemes, this effectively improves the accuracy of similarity selection and seasonal precipitation prediction.
[0041] Furthermore, the pre-trained error prediction model is obtained based on the following training steps: acquiring the first multivariate historical factor data corresponding to the target region in the target prediction month, the actual precipitation error data corresponding to the target prediction month, the precipitation error data of multiple historical years corresponding to the target region, and the second multivariate historical factor data of each historical year; inputting the first multivariate historical factor data and each second multivariate historical factor data into the bi-branch deep residual network in the initial error prediction model to obtain the first target feature corresponding to the first multivariate historical factor data and the second target feature corresponding to each second multivariate historical factor data; determining the feature similarity between the first target feature and each second target feature and the weight corresponding to each feature similarity based on the attention mechanism; determining the prediction error field corresponding to the target prediction month based on the feature similarity, the weight corresponding to each feature similarity, and the precipitation error data of the multiple historical years; and training the model parameters of the initial error prediction model based on the prediction error field corresponding to the target prediction month and the actual precipitation error data to obtain the error prediction model.
[0042] Here, the initial error prediction model can be a pre-trained model or an untrained model.
[0043] Here, the target prediction month refers to the month in which the precipitation prediction error is calculated. For example, if the precipitation error for May, June, and July is predicted in February, then the factor data for February is the first multivariate historical factor data.
[0044] It should be noted that the prediction error of the target prediction month will be different depending on the month corresponding to the first multivariate historical factor data, i.e., the starting month. Therefore, multiple sets of training data and training datasets can be constructed using the starting month to train the initial error prediction model.
[0045] Here, the deep feature extraction module is used to extract the deep features of historical factor data samples. The network architecture of the deep feature extraction module can be ResNet.
[0046] Here, the attention module is used to identify second multivariate historical factor data that are similar to the first multivariate historical factor data. It should be noted that when determining feature similarity, a nonlinear feature similarity measure between feature vectors is obtained because features of a high-dimensional nonlinear feature space are used.
[0047] Weights are used to characterize the similarity between second multivariate historical factor data that are similar to the first multivariate historical factor data. The larger the weight, the greater the feature similarity; conversely, the smaller the feature similarity.
[0048] In this embodiment of the invention, by constructing a feature extraction module and an attention mechanism, the complex nonlinear characteristics of the error field are accurately modeled. This comprehensively integrates data from multiple climate variables such as wind field, geopotential height, air temperature, and sea surface temperature, capturing various nonlinear interactions in the precipitation process. This overcomes the limitations of traditional linear methods, thus more accurately correcting model prediction errors. Simultaneously, a deep learning network is used to perform multi-dimensional feature analysis on factor data. The attention mechanism dynamically calculates the nonlinear weight of each similarity, intelligently allocating weights accordingly to optimize the estimation of prediction errors. This process is entirely data-driven, adaptively strengthening the contribution of similar samples closest to the current state while suppressing the influence of similar samples with large deviations. The intelligent similarity measurement of multi-dimensional factor data optimizes the selection of similar years, quickly and accurately filtering states with high similarity to the current prediction year from a large amount of historical data. This reduces the workload and subjectivity of manual selection, providing more reliable similar samples for error correction and effectively addressing the shortcomings of traditional methods in precipitation prediction in complex terrain areas.
[0049] Furthermore, the dual-branch deep residual network includes a basic feature extraction branch and an adaptive branch. The step of inputting the first multivariate historical factor data and each of the second multivariate historical factor data into the dual-branch deep residual network in the initial error prediction model to obtain the first target feature corresponding to the first multivariate historical factor data and the second target feature corresponding to each of the second multivariate historical factor data includes: inputting the first multivariate historical factor data into the basic feature extraction branch and the adaptive branch to obtain the global feature and the local feature corresponding to the first multivariate historical factor data; fusing the global feature and the local feature to obtain the first target feature; inputting each of the second multivariate historical factor data into the basic feature extraction branch and the adaptive branch to obtain the global feature and the local feature corresponding to each of the second multivariate historical factor data; and fusing the global feature and the local feature to obtain the second target feature.
[0050] Here, the basic feature extraction branch is used to extract global features corresponding to multivariate historical factor data, while the adaptive branch is used to extract local features corresponding to multivariate historical factor data.
[0051] Among them, the depth of the dual-branch deep residual network is pluggable, including but not limited to 15 layers, 18 layers or 34 layers.
[0052] In this embodiment of the invention, global features and local features are extracted and fused through a basic feature extraction branch and an adaptive branch to obtain target features, thereby improving the accuracy of factor data features and further improving the accuracy of feature similarity.
[0053] Furthermore, determining the prediction error field corresponding to the target prediction month based on the similarity of each feature, the weight corresponding to each feature similarity, and the precipitation error data of the multiple historical years includes: filtering the weights corresponding to each feature similarity based on an attention filtering mechanism to obtain at least one target weight; and weighting and summing the target weights and the precipitation error data of the multiple historical years corresponding to each target weight to obtain the prediction error field corresponding to the target prediction month.
[0054] Optionally, the attention filtering mechanism is configured to support at least one of the following modes: full attention, selecting the top K highest attentions in order of attention (Top-K), and threshold attention.
[0055] In this embodiment of the invention, target weights are determined through multiple attention filtering mechanisms. The target weights and the corresponding historical precipitation error data are weighted and summed to obtain the prediction error field, thereby improving the accuracy of model prediction.
[0056] Furthermore, the step of training the model parameters of the initial error prediction model based on the prediction error field corresponding to the target prediction month and the actual precipitation error data to obtain the error prediction model includes:
[0057] Based on the training loss function, the prediction error field corresponding to the target prediction month, and the actual precipitation error data, the difference between the prediction error field corresponding to the target prediction month and the actual precipitation error data is determined; the training loss function is the mean square error function, and its calculation formula is:
[0058]
[0059] in, Indicates mean square error. Indicates the number of region masks. This represents the true error corresponding to the region mask. This represents the prediction error corresponding to the region mask.
[0060] Based on the difference value, the model parameters of the initial error prediction model are trained to obtain the error prediction model.
[0061] In this embodiment of the invention, the initial error prediction model is subjected to gradient descent using a loss function, and the error prediction model is obtained when the loss value converges.
[0062] Furthermore, the training steps of the error prediction model adopt a training strategy of gradual expansion by year, which includes gradually adding data from recent years in chronological order for training.
[0063] It should be noted that after determining the month to be predicted, an error prediction model is trained using recent years' data before the month to be predicted, and then used to predict the current month to be predicted. For example, when predicting the error between model precipitation data and observed precipitation data in June 2018, the initial prediction model is trained using data from all or some months before 2018; when predicting the error between model precipitation data and observed precipitation data in June 2019, the initial prediction model is trained using data from all or some months before 2019.
[0064] In this embodiment of the invention, a training strategy of gradually expanding the model year by year is adopted to address the precipitation error of different months to be predicted. The training strategy of gradually expanding the model year by year includes gradually adding data from recent years in chronological order for training, which improves the accuracy of model prediction.
[0065] The following are application scenarios of the plateau precipitation prediction method based on intelligent similarity correction provided by this invention.
[0066] Data Collection and Processing: CRA observation data and coupled dynamical model data (GEM5-NEMO) from 1982 to 2020 were collected, including surface air temperature, 200 hPa wind field, 850 hPa wind field, 500 hPa geopotential height field, and precipitation data. The collected observation data and model data were declimatized, and historical precipitation error data of the model precipitation were calculated. The precipitation error data and feature data were further standardized, and corresponding data masks were generated to mask invalid data areas in subsequent processing. Data from June to August 2006 to 2020 were selected as validation data, and data from 1982 to 2005 were selected as training dataset.
[0067] Constructing the fine-tuning dataset: Based on the loaded feature data and precipitation error data, a fine-tuning dataset is constructed. The factor months are set to January to February and the prediction months to June to August, generating data samples such as input sequence features, model errors, masks, current features, and current errors for model training.
[0068] Build a deep feature extraction model: Use the Residual Network (ResNet18) architecture to build a basic model, map the input climate feature data to a feature space of specified dimensions, and set the output feature dimension.
[0069] Design a multi-attention mechanism module: Construct an attention model that includes multiple variants such as full attention, Top-K attention, and threshold attention. Full attention calculates the similarity between the factor features of the target starting month and the factor data of all historical months to obtain attention weights; Top-K attention selects the Top-K historical months most similar to the target starting month, highlighting their error contribution; threshold attention sets a threshold to filter out historical months with less impact on error correction, further focusing on key similarity error information.
[0070] Model Training: During model training, target factor data samples, historical factor data samples for historical months, and historical precipitation error samples corresponding to each historical month are input into the deep feature extraction model to obtain feature vectors. Then, attention weights are calculated using an attention model. Based on these weights, the errors of historically similar months are weighted and summed to obtain the prediction error for the month to be predicted. The average root mean square error of the target region is determined as the loss function, and the model parameters are optimized through backpropagation to complete the model training for the prediction error.
[0071] Error correction: Using a trained model, the target factor data is input into the prediction error model, which outputs the prediction error value. This prediction error value is then applied to the first model precipitation data to correct the error and generate prediction data that more closely reflects actual observations.
[0072] The plateau precipitation prediction system based on intelligent similarity correction provided by the present invention will be described below. The plateau precipitation prediction system based on intelligent similarity correction described below can be referred to in correspondence with the plateau precipitation prediction method based on intelligent similarity correction described above.
[0073] Figure 2 This is a schematic diagram of the plateau precipitation prediction system based on intelligent similarity correction provided by the present invention, as shown below. Figure 2 As shown, the plateau precipitation prediction system 200 based on intelligent similarity correction includes the following:
[0074] The acquisition module 210 is used to acquire the model precipitation data of the month to be predicted and its corresponding multivariate historical factor data and global historical annual precipitation error data. The multivariate historical factor data includes gridded data of wind field, geopotential height field and temperature.
[0075] The data preprocessing module 220 is used to standardize the gridded data of the wind field, altitude field and temperature to obtain standard gridded data of the wind field, altitude field and temperature.
[0076] Extraction module 230 is further configured to apply a predefined region mask to extract effective historical annual precipitation error data corresponding to effective geographical regions from the global historical annual precipitation error data; the effective geographical regions include plateau regions;
[0077] Error prediction module 240 is used to input the standard grid data and the effective historical annual precipitation error data into a pre-trained error prediction model to obtain the prediction error field output by the error prediction model; the prediction error field is used to characterize the error between the model precipitation data and the observed precipitation data in the plateau region; the prediction error field is obtained by the error prediction model through a dual-branch deep residual network and an attention mechanism, based on the feature similarity between the multivariate historical factor data features corresponding to the month to be predicted and the multivariate historical factor data features corresponding to the historical year, as well as the weights corresponding to the feature similarity;
[0078] Error correction module 250 is used to correct the model precipitation data based on the prediction error field to obtain the final predicted precipitation.
[0079] In another embodiment, the plateau precipitation prediction system 200 based on intelligent similarity correction further includes a training module, specifically used for: acquiring first multivariate historical factor data of the target region corresponding to the target prediction month, actual precipitation error data corresponding to the target prediction month, precipitation error data of multiple historical years corresponding to the target region, and second multivariate historical factor data of each of the historical years; inputting the first multivariate historical factor data and each of the second multivariate historical factor data into a bi-branch deep residual network in the initial error prediction model to obtain a first target feature corresponding to the first multivariate historical factor data and a second target feature corresponding to each of the second multivariate historical factor data; determining the feature similarity between the first target feature and each of the second target features and the weights corresponding to each feature similarity based on an attention mechanism; determining the prediction error field corresponding to the target prediction month based on each feature similarity, the weights corresponding to each feature similarity, and the precipitation error data of the multiple historical years; and training the model parameters of the initial error prediction model based on the prediction error field corresponding to the target prediction month and the actual precipitation error data to obtain the error prediction model.
[0080] In another embodiment, the dual-branch deep residual network includes a basic feature extraction branch and an adaptive branch. The training module is further specifically used for: inputting the first multivariate historical factor data into the basic feature extraction branch and the adaptive branch respectively to obtain global features and local features corresponding to the first multivariate historical factor data; fusing the global features and local features corresponding to the first multivariate historical factor data to obtain a first target feature corresponding to the first multivariate historical factor data; inputting each second multivariate historical factor data into the basic feature extraction branch and the adaptive branch respectively to obtain global features and local features corresponding to each second multivariate historical factor data; and fusing the global features and local features corresponding to each second multivariate historical factor data to obtain a second target feature corresponding to each second multivariate historical factor data.
[0081] In another embodiment, the training module is further specifically used to: filter the weights corresponding to the similarity of each feature based on the attention filtering mechanism to obtain at least one target weight; and perform a weighted summation of each target weight and the precipitation error data of the multiple historical years corresponding to each target weight to obtain the prediction error field corresponding to the target prediction month.
[0082] In another embodiment, the attention filtering mechanism is configured to support at least one of full attention, Top-K attention, and threshold attention.
[0083] In another embodiment, the training module is further specifically configured to: determine the difference between the prediction error field corresponding to the target prediction month and the actual precipitation error data based on the training loss function, the prediction error field corresponding to the target prediction month, and the actual precipitation error data; the training loss function is a mean squared error function, and its calculation formula is:
[0084]
[0085] in, Indicates mean square error. Indicates the number of region masks. This represents the true error corresponding to the region mask. The difference value represents the prediction error corresponding to the region mask; based on the difference value, the model parameters of the initial error prediction model are trained to obtain the error prediction model.
[0086] Figure 3 This is a schematic diagram of the structure of the electronic device provided by the present invention, such as... Figure 3As shown, the electronic device may include: a processor 810, a communications interface 820, a memory 830, and a communications bus 840, wherein the processor 810, the communications interface 820, and the memory 830 communicate with each other through the communications bus 840. The processor 810 can call logic instructions in the memory 830 to execute a plateau precipitation prediction method based on intelligent similarity correction. This method includes: acquiring model precipitation data for the month to be predicted, along with corresponding multivariate historical factor data and global historical annual precipitation error data, wherein the multivariate historical factor data includes gridded data of wind field, geopotential height, and temperature; standardizing the gridded data of wind field, geopotential height, and temperature to obtain standard gridded data of wind field, geopotential height, and temperature; applying a predefined regional mask to extract effective historical annual precipitation error data corresponding to effective geographical regions from the global historical annual precipitation error data; wherein the effective geographical regions include plateau regions; and inputting the standard gridded data and the effective historical annual precipitation error data into a pre-trained error prediction model to obtain the prediction error field output by the error prediction model.
[0087] Furthermore, the logical instructions in the aforementioned memory 830 can be implemented as software functional units and, when sold or used as independent products, 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.
[0088] On the other hand, the present invention also 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 the plateau precipitation prediction method based on intelligent similarity correction provided by the above methods. The method includes: acquiring model precipitation data for the month to be predicted and its corresponding multivariate historical factor data and global historical annual precipitation error data, wherein the multivariate historical factor data includes gridded data of wind field, geopotential height, and temperature; standardizing the gridded data of wind field, geopotential height, and temperature to obtain standard gridded data of wind field, geopotential height, and temperature; applying a predefined regional mask to extract effective historical annual precipitation error data corresponding to effective geographical regions from the global historical annual precipitation error data; wherein the effective geographical regions include plateau regions; and inputting the standard gridded data and the effective historical annual precipitation error data into a pre-trained error prediction model to obtain the prediction error field output by the error prediction model.
[0089] In another aspect, the present invention also provides a non-transitory computer-readable storage medium storing a computer program thereon. When executed by a processor, the computer program implements the plateau precipitation prediction method based on intelligent similarity correction provided by the above methods. The method includes: acquiring model precipitation data for the month to be predicted, along with its corresponding multivariate historical factor data and global historical annual precipitation error data, wherein the multivariate historical factor data includes gridded data of wind field, geopotential height, and temperature; standardizing the gridded data of wind field, geopotential height, and temperature to obtain standard gridded data of wind field, geopotential height, and temperature; applying a predefined regional mask to extract effective historical annual precipitation error data corresponding to effective geographical regions from the global historical annual precipitation error data; wherein the effective geographical regions include plateau regions; and inputting the standard gridded data and the effective historical annual precipitation error data into a pre-trained error prediction model to obtain the prediction error field output by the error prediction model.
[0090] The system 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.
[0091] 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.
[0092] 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 plateau precipitation prediction method based on intelligent similarity correction, characterized in that, include: Obtain model precipitation data for the month to be predicted, along with its corresponding multivariate historical factor data and global historical annual precipitation error data. The multivariate historical factor data includes gridded data of wind field, geopotential height field, and temperature. The gridded data of wind field, altitude field and temperature are standardized to obtain standard gridded data of wind field, altitude field and temperature; The effective historical annual precipitation error data corresponding to the effective geographic region is extracted from the global historical annual precipitation error data using a predefined regional mask; the effective geographic region includes plateau regions; The standard grid data and the effective historical annual precipitation error data are input into a pre-trained error prediction model to obtain the prediction error field output by the error prediction model. The prediction error field is used to characterize the error between the model precipitation data and the observed precipitation data in the plateau region. The prediction error field is obtained by the error prediction model through a bi-branch deep residual network and an attention mechanism, based on the feature similarity between the multivariate historical factor data features corresponding to the month to be predicted and the multivariate historical factor data features corresponding to the historical year, as well as the weights corresponding to the feature similarity. Based on the prediction error field, the model precipitation data is corrected to obtain the final predicted precipitation. The pre-trained error prediction model is obtained based on the following training steps: Acquire the first multivariate historical factor data of the target area in the target prediction month, the actual precipitation error data of the target prediction month, the precipitation error data of multiple historical years corresponding to the target area, and the second multivariate historical factor data of each historical year; The first multivariate historical factor data and each of the second multivariate historical factor data are respectively input into the bi-branch deep residual network in the initial error prediction model to obtain the first target feature corresponding to the first multivariate historical factor data and the second target feature corresponding to each of the second multivariate historical factor data. The dual-branch deep residual network includes a basic feature extraction branch and an adaptive branch. The basic feature extraction branch is used to extract global features corresponding to multivariate historical factor data, and the adaptive branch is used to extract local features corresponding to multivariate historical factor data. Based on the attention mechanism, the feature similarity between the first target feature and each of the second target features and the weights corresponding to each feature similarity are determined; Based on the similarity of each feature, the weight corresponding to each feature similarity, and the precipitation error data of the multiple historical years, the prediction error field corresponding to the target prediction month is determined; Based on the prediction error field corresponding to the target prediction month and the actual precipitation error data, the model parameters of the initial error prediction model are trained to obtain the error prediction model.
2. The plateau precipitation prediction method based on intelligent similarity correction according to claim 1, characterized in that, The step of inputting the first multivariate historical factor data and each of the second multivariate historical factor data into the bi-branch deep residual network in the initial error prediction model to obtain the first target feature corresponding to the first multivariate historical factor data and the second target feature corresponding to each of the second multivariate historical factor data includes: The first multivariate historical factor data is input into the basic feature extraction branch and the adaptive branch respectively to obtain the global features and local features corresponding to the first multivariate historical factor data. The global features and local features corresponding to the first multivariate historical factor data are fused to obtain the first target feature corresponding to the first multivariate historical factor data. Each of the second multivariate historical factor data is input into the basic feature extraction branch and the adaptive branch respectively to obtain the global features and local features corresponding to each of the second multivariate historical factor data. The global features and local features corresponding to each of the second multivariate historical factor data are fused to obtain the second target features corresponding to each of the second multivariate historical factor data.
3. The plateau precipitation prediction method based on intelligent similarity correction according to claim 1, characterized in that, The step of determining the prediction error field corresponding to the target prediction month based on the similarity of each feature, the weight corresponding to each feature similarity, and the precipitation error data of the multiple historical years includes: Based on the attention filtering mechanism, the weights corresponding to the similarity of each feature are filtered to obtain at least one target weight; The prediction error field corresponding to the target prediction month is obtained by weighting and summing the precipitation error data of each target weight and the multiple historical years corresponding to each target weight.
4. The plateau precipitation prediction method based on intelligent similarity correction according to claim 3, characterized in that, The attention filtering mechanism is configured to support at least one of the following modes: full attention, Top-K attention, and threshold attention.
5. The plateau precipitation prediction method based on intelligent similarity correction according to claim 1, characterized in that, The step of training the model parameters of the initial error prediction model based on the prediction error field corresponding to the target prediction month and the actual precipitation error data to obtain the error prediction model includes: Based on the training loss function, the prediction error field corresponding to the target prediction month, and the actual precipitation error data, the difference between the prediction error field corresponding to the target prediction month and the actual precipitation error data is determined; the training loss function is the mean square error function, and its calculation formula is: Where MSE represents the mean squared error. Indicates the number of region masks. This represents the true error corresponding to the region mask. This represents the prediction error corresponding to the region mask. Based on the difference value, the model parameters of the initial error prediction model are trained to obtain the error prediction model.
6. The plateau precipitation prediction method based on intelligent similarity correction according to claim 1, characterized in that, The training steps of the error prediction model adopt a training strategy of gradual expansion by year, which includes gradually adding data from recent years in chronological order for training.
7. A plateau precipitation prediction system based on intelligent similarity correction for implementing the plateau precipitation prediction method based on intelligent similarity correction as described in any one of claims 1-6, characterized in that, include: The acquisition module is used to acquire model precipitation data for the month to be predicted, as well as its corresponding multivariate historical factor data and global historical annual precipitation error data. The multivariate historical factor data includes gridded data of wind field, geopotential height field and temperature. The data preprocessing module is used to standardize the gridded data of the wind field, altitude field, and temperature to obtain standard gridded data of the wind field, altitude field, and temperature. The extraction module is also used to extract effective historical annual precipitation error data corresponding to effective geographical regions from the global historical annual precipitation error data by applying a predefined regional mask; the effective geographical regions include plateau regions; An error prediction module is used to input the standard grid data and the effective historical annual precipitation error data into a pre-trained error prediction model to obtain a prediction error field output by the error prediction model. The prediction error field is used to characterize the error between the model precipitation data and the observed precipitation data in the plateau region. The prediction error field is obtained by the error prediction model through a dual-branch deep residual network and an attention mechanism, based on the feature similarity between the multivariate historical factor data features corresponding to the month to be predicted and the multivariate historical factor data features corresponding to the historical year, as well as the weights corresponding to the feature similarity. An error correction module is used to correct the model precipitation data based on the prediction error field to obtain the final predicted precipitation.
8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the plateau precipitation prediction method based on intelligent similarity correction as described in any one of claims 1-6.
9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the program implements the plateau precipitation prediction method based on intelligent similarity correction as described in any one of claims 1-6.