Deep learning-based large-area high-resolution model precipitation correction method and system
By dividing the monitoring area into grids and applying deep learning models, significant coefficients and feature values are constructed. Combined with LSTM neural networks for rainfall prediction, the problem of error accumulation in traditional models is solved, and higher accuracy rainfall prediction is achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- 河南省气象台
- Filing Date
- 2026-03-19
- Publication Date
- 2026-06-12
Smart Images
Figure CN122194348A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of precipitation prediction and analysis technology, specifically to a method and system for correcting precipitation using a large-area high-resolution model based on deep learning. Background Technology
[0002] Precipitation is a core element in climate change research, and accurate correction of it is an extremely challenging task. This requires not only a deep understanding of complex meteorological processes but also efficient processing of massive amounts of multi-source data. The formation and evolution of precipitation are influenced by multiple factors, including atmospheric dynamics and thermodynamic processes; therefore, high-resolution model precipitation correction is crucial for disaster prevention and water resource management.
[0003] Traditional numerical weather prediction models rely primarily on observational data from meteorological satellites and radar for precipitation forecasting and analysis. While they can make long-term predictions, the highly complex nonlinear characteristics of atmospheric motion, the inherent randomness of precipitation, and the uncertainty of its spatiotemporal variations lead to errors in the forecast results. As the forecast lead time increases, these errors may gradually accumulate, becoming particularly pronounced during extreme weather events. This affects the spatiotemporal resolution of traditional weather forecasts, making it difficult to accurately capture small-scale precipitation changes and resulting in insufficient accuracy in high-resolution precipitation predictions. Consequently, corrections need to be made to the precipitation data obtained from traditional numerical weather prediction models. Summary of the Invention
[0004] To address the aforementioned technical problems, the purpose of this application is to provide a method and system for correcting precipitation in large-area high-resolution models based on deep learning. The specific technical solution adopted is as follows: In a first aspect, embodiments of this application provide a method for correcting precipitation in large-area high-resolution models based on deep learning, the method comprising the following steps: The monitored area is divided into grids, and time-series rainfall data within each grid is collected. The degree of disorder in rainfall variation within the grid at local time was analyzed, and the significance coefficients for each time point of the grid were constructed. The rainfall area is determined by the rainfall of the grid at the same time, and the edge degree of each grid position in the rainfall area is measured. Based on the difference in rainfall and the difference in significance coefficient between each grid and its spatial neighboring grids, the local prominence of the rainfall of each grid at the same time is analyzed. Combined with the edge degree, the first feature value of each grid at the same time is constructed. The second feature value of each grid at each time point is constructed based on the difference between the first feature value of each grid at each time point and the time point in its historical local time. The rainfall prediction model is trained by using the historical rainfall data and second feature values of each grid, and then the rainfall is corrected.
[0005] In one embodiment, the process of obtaining the significance coefficients of the grid at each time step is as follows: The differences in rainfall and the degree of disorder in the rate of change of rainfall within the time neighborhood of each time step are calculated, and the two are combined to obtain the significance coefficients of each time step of the grid.
[0006] In one embodiment, the differences in rainfall and the degree of disorder in the rate of change of rainfall within the time neighborhood grid are respectively: The difference in rainfall within the grid within the time neighborhood is: the difference between the average rainfall to the left and the average rainfall to the right of the center time within the time neighborhood of each grid at each time point; The disorder level of the rate of change of rainfall within the time neighborhood is defined as follows: by performing curve fitting on all rainfall amounts within the time neighborhood of the grid at each time step using a curve fitting algorithm, the standard deviation of the first derivative at all times on the fitted curve is taken as the disorder level.
[0007] In one embodiment, the significance coefficient is the product of the difference in rainfall within the grid in the time neighborhood and the degree of disorder.
[0008] In one embodiment, determining the rainfall area by measuring the rainfall in the grid at the same time specifically involves: The rainfall data of all grids at the same time are clustered using a clustering algorithm. All grids in the cluster with the smallest mean value in the clustering results are designated as non-rainfall areas, and all grids remaining outside the non-rainfall areas are designated as rainfall areas.
[0009] In one embodiment, the edge degree of each grid location in the rainfall area is measured as follows: The heavy rainfall area and the light rainfall area are determined by the rainfall amount in the grid of the rainfall area. The center point of the heavy rainfall area and the light rainfall area are obtained by the center point detection algorithm. The distance measurement algorithm is used to obtain the measurement distance from each grid in the rainfall area to the corresponding center point. The measurement distance is used as the edge degree of each grid position in the rainfall area.
[0010] In one embodiment, the local prominence of the rainfall in each grid at the same moment is specifically as follows: Calculate the mean difference in rainfall between each grid and the other grids in its spatial neighborhood at the same time, as well as the mean difference in significance coefficients; calculate the positive fusion value of the mean difference in rainfall and the mean difference in significance coefficients to obtain the local prominence of rainfall in each grid at the same time.
[0011] In one embodiment, the first feature value is the product of the inverse proportional mapping value of the edge degree and the local protrusion degree.
[0012] In one embodiment, the second feature value is the sum of the differences between the first feature values of each grid at each time point and all times in the historical local time period.
[0013] Secondly, embodiments of this application also provide a large-area high-resolution model precipitation correction system based on deep learning, including a memory, a processor, and a computer program stored in the memory and running on the processor, wherein the processor executes the computer program to implement the steps of any of the methods described above.
[0014] The embodiments of this application have at least the following beneficial effects: This application divides the monitored area into grids and deeply analyzes the fluctuation characteristics of the intensity and rate of change of rainfall at each grid location under the influence of complex and variable climate. It constructs significance coefficients for each grid at each time point, which has the advantage of accurately capturing short-term rainfall states at the grid scale. Furthermore, it considers the spatial non-uniformity and multi-scale characteristics of rainfall to construct the first feature value for each grid at each time point, and constructs the second feature value based on the temporal variation characteristics of the first feature value for each grid. This accurately reflects the spatiotemporal variation characteristics of rainfall states under the complex influence of various atmospheric movements. Using the historical rainfall and the second feature value of each grid, an LSTM neural network model is used for rainfall prediction analysis. Introducing the second feature value while using rainfall data for prediction enables the model to more directly and accurately reflect the rainfall change state. The prediction results of this model are used to correct rainfall data predicted by traditional methods. Therefore, facing the challenges of complex and variable climate, and the high non-stationarity and spatial non-uniformity of rainfall, it achieves more accurate and reliable rainfall prediction, helping to compensate for the insufficient accuracy of high-resolution rainfall prediction. Attached Figure Description
[0015] To more clearly illustrate the technical solutions and advantages in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0016] Figure 1 A flowchart illustrating the steps of a deep learning-based large-area high-resolution model precipitation correction method provided in one embodiment of this application; Figure 2 This is a schematic diagram illustrating the process of obtaining the first eigenvalue. Detailed Implementation
[0017] To further illustrate the technical means and effects adopted by this application to achieve the intended inventive objective, the following, in conjunction with the accompanying drawings and preferred embodiments, details the specific implementation, structure, features, and effects of the deep learning-based large-area high-resolution model precipitation correction method and system proposed in this application. In the following description, different "one embodiment" or "another embodiment" do not necessarily refer to the same embodiment. Furthermore, specific features, structures, or characteristics in one or more embodiments can be combined in any suitable form.
[0018] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains.
[0019] The following, in conjunction with the accompanying drawings, details the specific scheme of the deep learning-based large-area high-resolution model precipitation correction method and system provided in this application.
[0020] Please see Figure 1 The diagram illustrates a flowchart of a deep learning-based large-area high-resolution model precipitation correction method provided in an embodiment of this application. The method includes the following steps: Step S1: Divide the monitored area into grids and collect time-series rainfall data within the grids.
[0021] High resolution of precipitation data is reflected in both spatial and temporal dimensions. High temporal resolution can better capture the start and end times and duration of precipitation, more realistically reflecting its transient characteristics and intensity changes. High spatial resolution can better express the spatial characteristics of continuous distribution at different locations. Therefore, this application performs gridding on the monitored area, setting the grid size to [size missing]. The system acquires rainfall data for each grid location from the weather station every minute, and then normalizes all collected rainfall data using a normalization algorithm. Specifically, this embodiment uses the maximum-minimum normalization method. Many existing normalization methods exist, and implementers may use other normalization algorithms to normalize rainfall data; this application does not impose any specific limitations.
[0022] It should be noted that implementers can set the grid division method and data acquisition frequency according to the actual situation, and this application does not impose specific restrictions.
[0023] Step S2: Analyze the degree of disorder in the variation of rainfall within the grid at local time, and construct the significance coefficients for each time point of the grid.
[0024] Precipitation forecasting primarily uses numerical statistical models to predict the occurrence, intensity, location, and timing of precipitation in the future. However, atmospheric changes involve extremely complex physical processes and exhibit highly nonlinear characteristics. Under the influence of multiple factors, there is a significant discrepancy between predicted and actual precipitation data. Therefore, the following analysis addresses this issue.
[0025] First, influenced by complex and variable climate, such as the instantaneous adjustment of atmospheric stratification, rainfall intensity may experience drastic changes in a short period of time, manifesting as abrupt changes in precipitation data at a certain moment, such as a sudden shift from light rain to heavy rainfall. Furthermore, the formation, dissipation, and movement of large-scale convective systems can also easily trigger intermittent and rapid fluctuations. Therefore, short-term rainfall data exhibits highly non-stationary characteristics and abrupt changes.
[0026] Starting from the occurrence of rainfall in the monitored area, any grid within the monitored area is taken as the current grid. Taking the i-th time at the current grid position as an example, the greater the difference in rainfall data within a short period, the more significant the abrupt change in rainfall data at that time, indicating a more drastic change in rainfall intensity. A local time window of size N*1 is set with the i-th time as the center. In this embodiment, the value of N is set to 11. In other embodiments of this application, the implementer can set the value of N according to the actual situation. Then, the difference between the mean of rainfall data at all times to the left of the center time and the mean of rainfall data at all times to the right of the center time within this local time window is calculated and denoted as . The result This reflects the intensity of the sudden change in rainfall data at time i. The difference can be the absolute value of the difference, the square of the difference, a ratio, etc. In this embodiment, the difference is the absolute value of the difference between the mean of all rainfall data on the left and the mean of all rainfall data on the right.
[0027] Furthermore, the overall non-stationary variation characteristics of rainfall data are analyzed. Since direct calculation using rainfall data is easily affected by noise, the rainfall data is first smoothed through curve fitting. In this embodiment, the least squares method is used to fit all rainfall data within the local time window of the current grid at time i, resulting in a fitted curve. The least squares method is a well-known technique, and its specific process will not be elaborated further. Then, the first derivative of the fitted curve at each time step is calculated. The magnitude of the first derivative reflects the rate of change of the rainfall data at each time step. The standard deviation of the first derivative of the fitted curve at all times is also calculated and denoted as . The standard deviation reflects the degree of disorder in the data distribution of the first derivative. The larger the value, the more pronounced the non-stationary characteristics of rainfall in the local time period at that current grid point.
[0028] It should be noted that this application only provides one curve fitting method for curve fitting of rainfall data within a local time window. There are many existing curve fitting methods, and implementers may also use other curve fitting algorithms to perform curve fitting of rainfall data within a local time window. This application does not impose any specific restrictions.
[0029] Therefore, the significance coefficients of the sudden changes and non-stationarity of rainfall data at each time point in the current grid are calculated. Specifically, the significance coefficients are the forward fusion results of the differences in rainfall data and the degree of disorder in the rate of change of rainfall within the local time window. The forward fusion refers to combining two or more indicators by addition or multiplication. Preferably, in this embodiment, the expression for the significance coefficients of the current grid at each time point is: In the formula, Let be the significance coefficient of the current grid at time i. This represents the difference between the mean of all rainfall data to the left of the center time and the mean of all rainfall data to the right of the center time within the local time window of the current grid at time i. Let be the standard deviation of the first derivative of the fitted curve of rainfall data within the local time window at time i of the current grid, across all times. The resulting... This reflects the sudden change in rainfall intensity and its non-stationary characteristics within a short period of time at that moment.
[0030] This application assesses the degree of abrupt changes in rainfall by analyzing the data difference characteristics on both sides of the central time point, and analyzes the degree of non-stationarity by calculating the fluctuation of the rate of change of rainfall data at each time point.
[0031] Step S3: Determine the rainfall area by measuring the rainfall of each grid at the same time, and measure the edge degree of each grid position in the rainfall area; based on the difference in rainfall and the difference in significance coefficient between each grid and its spatial neighboring grids, analyze the local prominence of rainfall of each grid at the same time, and construct the first feature value of each grid at the same time by combining the edge degree.
[0032] Influenced by macroscopic weather systems and small-scale convective activity, rainfall exhibits significant spatial heterogeneity and multi-scale characteristics. This typically manifests as rainfall zones of varying shapes and sizes, with different locations within each zone receiving varying amounts of rainfall. For example, rainfall may be extremely high in the central area, while decreasing towards the edges, resulting in non-uniform coverage. This combined influence from macroscopic systems and small-scale convection leads to strong spatial heterogeneity in rainfall. This heterogeneity is more pronounced in areas with significant topographic relief, as changes in topography directly affect atmospheric motion, causing different rainfall mechanisms to interact within a smaller geographical area. Ultimately, this results in a more complex and pronounced spatial distribution of rainfall compared to flat areas.
[0033] First, the percentage of grid cells with rainfall greater than 0 within the monitored area at time i is calculated. If this percentage is less than a threshold, subsequent clustering and calculation of the first feature value are skipped. In this embodiment, the threshold is set to 5%. In other embodiments of this application, the implementer can set the threshold according to the actual situation.
[0034] If the proportion is less than the proportion threshold, it indicates that there is rainfall in the monitored area at time i. To further analyze the spatial distribution characteristics of rainfall, taking time i as an example, in actual weather analysis, rainfall usually has a certain degree of spatial continuity. This application uses the K-means clustering algorithm to segment the rainfall data of all grids in the monitored area at time i, where each grid cell corresponds to one rainfall data point. The number of clusters is set to 3. The input of the algorithm is the rainfall data of all grids at that time, and the output is each cluster. The mean of all rainfall data in each cluster is calculated, and the magnitude of these means is compared. The area corresponding to all grids in the cluster with the largest mean is designated as the heavy rainfall area, the area corresponding to all grids in the cluster with the middle mean is designated as the light rainfall area, and the area corresponding to all grids in the cluster with the smallest mean is designated as the non-rainfall area. K-means clustering is a well-known technique, and the specific process will not be described in detail.
[0035] It should be noted that this application provides only one clustering method for clustering all grid rainfall at each time point. There are many existing clustering fitting methods, and implementers may also use other clustering algorithms to cluster all grid rainfall at each time point. This application does not impose any specific restrictions.
[0036] Further analysis is conducted on the identified areas of heavy and light rainfall. Specifically, the centroid calculation method is used to obtain the center points of the heavy and light rainfall areas, respectively. The Manhattan distance between each grid cell in each rainfall area and its corresponding center point grid cell is then calculated and denoted as the edge degree of that grid cell location. This edge degree reflects the distance of the grid cell location from the rainfall center at time i. Generally, the farther away from the rainfall center, the more stable the rainfall variation. Both the centroid calculation method and the Manhattan distance are well-known techniques, and their specific processes will not be elaborated upon.
[0037] It should be noted that this application provides only one distance measurement method for the distance between each grid in the rainfall area and its corresponding center point grid. There are many existing distance measurement fitting methods, and implementers may also use other distance measurement algorithms to calculate the distance between each grid in the rainfall area and its corresponding center point grid. This application does not impose any specific restrictions.
[0038] Then, taking the j-th grid in the rainfall zone at time i as an example, we calculate the mean of the difference in rainfall amount and the mean of the difference in significance coefficient between the j-th grid and all its neighboring grids at time i, and denot them as follows: , Preferably, in this embodiment, the rainfall difference and the significance coefficient difference are respectively the absolute value of the rainfall difference between the j-th grid and its surrounding adjacent rainfall areas at that moment, and the absolute value of the significance coefficient difference. In this embodiment, the adjacent grids of each grid refer to the remaining grids within the 3*3 neighborhood centered on each grid, excluding the central grid.
[0039] income This reflects the rainfall differences between the corresponding locations of this grid and its surrounding grids. This reflects the degree of difference in rainfall events between the corresponding locations of the grid and its surrounding grids. Furthermore, the positive fusion value of the mean of the rainfall difference and the mean of the significance coefficient difference is used as the local prominence of the rainfall in that grid at that moment. Subsequently, the first characteristic value of the rainfall status and its degree of change for each grid in the heavy and light rainfall areas at each moment is calculated.
[0040] Preferably, in this embodiment, the expression for the first feature value is: In the formula, Let be the first characteristic value of the j-th grid in the rainfall area at time i; Let $\frac{j}{i}$ be the mean of the absolute values of the differences in rainfall data between the $j$ grid in the $i$-th rainfall area and all its neighboring grids in the rainfall area. It is the mean of the absolute values of the significance coefficient differences between the j-th grid in the rainfall area at time i and all its adjacent grids in the rainfall area; The edge degree of the j-th grid in the rainfall area at time i; These are preset adjustment parameters used to avoid a denominator of 0. Preferably, in this embodiment, [the parameter is set to...]. The value is set to 1. The result is... The larger the value, the greater the difference in the rainfall state and its changing characteristics at time i for the corresponding grid.
[0041] For a given grid, the closer it is to the rainfall center, and the greater the difference in local rainfall intensity between that grid and its surrounding grid locations, the better the results will be. The larger the area, the more pronounced the spatial unevenness of rainfall distribution may be; and The study compared and analyzed the spatial differences in the dynamic characteristics of rainfall processes (e.g., abrupt changes and non-stationarity). A larger value indicates that the rainfall variation at that grid point is significantly different from that of its surrounding grid points. This allows for the assessment of the spatial non-uniformity and multi-scale characteristics of rainfall distribution.
[0042] Step S4: Construct the second feature value of each grid at each time point based on the difference between the first feature value of each grid at each time point and the time point in its historical local time.
[0043] Furthermore, most precipitation areas move as a whole with changes in weather systems. This movement is not a simple linear translation, but rather the result of multiple factors, including convective dynamics, topographic forcing, and upper atmospheric structure, exhibiting complex characteristics of direction, velocity variations, and morphological evolution. For example, cyclones move relatively smoothly, while convective cells move faster and more irregularly, and the shape of precipitation areas changes accordingly as they advance. Thus, influenced by the time-varying characteristics of precipitation area movement, the precipitation state at the grid location exhibits more variable and complex features, with changes occurring in the location, shape, and internal intensity distribution of the precipitation area's center point.
[0044] Therefore, the absolute value of the difference between the first feature value at the current grid position at time i and the corresponding first feature value at the previous M consecutive times is calculated. The sum of all the absolute values at time i of the current grid is taken as the second feature value of the current grid affected by the spatial movement characteristics of rainfall at that time. Since some times are missing due to lack of rainfall, this embodiment uses edge filling to fill in the missing data. The edge filling algorithm is a known technique, and the specific process will not be described in detail. In this embodiment, the value of M is set to 20. In other embodiments of this application, the implementer can set the value of M according to the actual situation. If there are fewer than M times before time i, the data is filled with 0 values.
[0045] It should be noted that this application provides only one filling algorithm for missing values. There are many existing filling algorithms, and implementers may also use other filling algorithms for missing values. This application does not impose any specific restrictions.
[0046] The obtained second eigenvalue reflects the spatiotemporal variation characteristics of rainfall within the grid location. The obtained first eigenvalue reflects the spatial distribution characteristics of rainfall at a certain moment, and by comparing the spatial distribution differences at multiple consecutive moments, the degree of recent change in the grid location is reflected.
[0047] Step S5: Train the rainfall prediction model using the historical rainfall data and second feature values of each grid, and then correct the rainfall data.
[0048] This application analyzes the fluctuation characteristics of rainfall intensity and rate of change at various grid locations under the influence of complex and variable climate. Furthermore, it considers the spatial non-uniformity and multi-scale characteristics of rainfall distribution, as well as the time-varying characteristics of rainfall zone spatial movement, to comprehensively evaluate the spatiotemporal variation of rainfall status, thereby improving the accuracy of rainfall prediction. Specifically, taking the current grid location as an example, this application uses a dual-channel fusion-based LSTM neural network model to fit and optimize rainfall data. Three months of historical data are selected, and the historical data is analyzed and calculated using the above steps to obtain the corresponding second feature values at each time point. The historical three-month rainfall data and the second feature values of the current grid are used as input to the LSTM neural network model, and the training set, test set, and validation set are divided in an 8:1:1 ratio. Then, the model is trained using well-known techniques, and the specific process is not detailed here. This yields the corresponding rainfall prediction model. Using historical rainfall data alone for predictive analysis primarily teaches the autocorrelation patterns over time. Introducing a second eigenvalue allows the model to more directly and accurately reflect rainfall changes, enabling more accurate and reliable rainfall predictions when facing challenges such as complex and variable climate, and the high non-stationarity and spatial heterogeneity of rainfall. In practical applications, the obtained rainfall prediction model is used for rainfall prediction analysis. The algorithm's input includes rainfall data for all times within the past hour and the corresponding second eigenvalues at each time point. The loss function is set to cross-entropy loss, and the optimizer is Adam. Rainfall data is input into one channel to capture strong temporal dependencies; the calculated second eigenvalue is input into the other channel. The hidden states of the two channels are concatenated and merged at each time step before being fed into subsequent LSTM or fully connected layers to obtain the corresponding rainfall prediction results. Furthermore, rainfall data predicted by traditional methods is replaced with the rainfall prediction results obtained by the method described in this application to correct the rainfall data. This corrected future rainfall data helps to compensate for the insufficient accuracy of high-resolution rainfall predictions. The implementation can set the selection period for historical data and the selection of neural network models according to the actual situation, and this application does not impose specific restrictions.
[0049] The process of obtaining the first eigenvalue is illustrated in the diagram below. Figure 2 As shown.
[0050] Based on the same inventive concept as the above methods, this application also provides a large-area high-resolution model precipitation correction system based on deep learning, including a memory, a processor, and a computer program stored in the memory and running on the processor. When the processor executes the computer program, it implements the steps of any one of the above-described large-area high-resolution model precipitation correction methods based on deep learning.
[0051] In summary, this application provides a large-area high-resolution model precipitation correction method based on deep learning. By dividing the monitored area into grids, it deeply analyzes the fluctuation characteristics of the intensity and rate of change of precipitation at each grid location under the influence of complex and variable climate, and constructs the significance coefficients of each grid at each time point. Its advantage lies in accurately capturing the short-term precipitation state at the grid scale. Furthermore, it considers the non-uniformity and multi-scale characteristics of precipitation in spatial distribution to construct the first feature value of each grid at each time point, and constructs the second feature value based on the temporal variation characteristics of the first feature value of each grid, which can accurately reflect the spatiotemporal variation characteristics of precipitation state under the complex influence of various atmospheric movements. Using the historical precipitation amount and the second feature value of each grid, the LSTM neural network model is used for precipitation prediction analysis. While using precipitation data for prediction, the second feature value is introduced, enabling the model to more directly and accurately reflect the precipitation change state. The prediction results of this model are used to correct the precipitation data predicted by traditional methods, thereby achieving more accurate and reliable precipitation prediction when facing the challenges of complex and variable climate, and the high non-stationarity and spatial non-uniformity of precipitation, which helps to make up for the shortcomings of insufficient accuracy in high-resolution precipitation prediction.
[0052] It should be noted that the order of the embodiments described above is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. Furthermore, the above description focuses on specific embodiments of this application. Additionally, the processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired results. In some implementations, multitasking and parallel processing are possible or may be advantageous.
[0053] The various embodiments in this application are described in a progressive manner. The same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on describing the differences from other embodiments.
[0054] The above description is only a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the principles of this application should be included within the protection scope of this application.
Claims
1. A method for correcting precipitation using a large-area high-resolution model based on deep learning, characterized in that: The method includes the following steps: The monitored area is divided into grids, and time-series rainfall data within each grid is collected. The degree of disorder in rainfall variation within the grid at local time was analyzed, and the significance coefficients for each time point of the grid were constructed. The rainfall area is determined by the rainfall of the grid at the same time, and the edge degree of each grid position in the rainfall area is measured. Based on the difference in rainfall and the difference in significance coefficient between each grid and its spatial neighboring grids, the local prominence of the rainfall of each grid at the same time is analyzed. Combined with the edge degree, the first feature value of each grid at the same time is constructed. The second feature value of each grid at each time point is constructed based on the difference between the first feature value of each grid at each time point and the time point in its historical local time. The rainfall prediction model is trained by using the historical rainfall data and second feature values of each grid, and then the rainfall is corrected.
2. The method for correcting precipitation in large-area high-resolution models based on deep learning as described in claim 1, characterized in that, The process of obtaining the significance coefficients of the grid at each time step is as follows: The differences in rainfall and the degree of disorder in the rate of change of rainfall within the time neighborhood of each time step are calculated, and the two are combined to obtain the significance coefficients of each time step of the grid.
3. The method for correcting precipitation in large-area high-resolution models based on deep learning as described in claim 2, characterized in that, The differences in rainfall and the degree of disorder in the rate of change of rainfall within the time neighborhood grid are as follows: The difference in rainfall within the grid within the time neighborhood is: the difference between the average rainfall to the left and the average rainfall to the right of the center time within the time neighborhood of each grid at each time point; The disorder level of the rate of change of rainfall within the time neighborhood is defined as follows: by performing curve fitting on all rainfall amounts within the time neighborhood of the grid at each time step using a curve fitting algorithm, the standard deviation of the first derivative at all times on the fitted curve is taken as the disorder level.
4. The method for correcting precipitation in large-area high-resolution models based on deep learning as described in claim 2, characterized in that, The significance coefficient is the product of the difference in rainfall within the grid in the time neighborhood and the degree of disorder.
5. The method for correcting precipitation in large-area high-resolution models based on deep learning as described in claim 1, characterized in that, The method of determining the rainfall area by measuring the rainfall in the grid at the same time is as follows: The rainfall data of all grids at the same time are clustered using a clustering algorithm. All grids in the cluster with the smallest mean value in the clustering results are designated as non-rainfall areas, and all grids remaining outside the non-rainfall areas are designated as rainfall areas.
6. The method for correcting precipitation in large-area high-resolution models based on deep learning as described in claim 1, characterized in that, The method for measuring the edge degree of each grid location in the rainfall area is as follows: The heavy rainfall area and the light rainfall area are determined by the rainfall amount in the grid of the rainfall area. The center point of the heavy rainfall area and the light rainfall area are obtained by the center point detection algorithm. The distance measurement algorithm is used to obtain the measurement distance from each grid in the rainfall area to the corresponding center point. The measurement distance is used as the edge degree of each grid position in the rainfall area.
7. The method for correcting precipitation in large-area high-resolution models based on deep learning as described in claim 1, characterized in that, The specific degree of local prominence of rainfall in each grid at the same moment is as follows: Calculate the mean difference in rainfall between each grid and the other grids in its spatial neighborhood at the same time, as well as the mean difference in significance coefficients; calculate the positive fusion value of the mean difference in rainfall and the mean difference in significance coefficients to obtain the local prominence of rainfall in each grid at the same time.
8. The method for correcting precipitation in large-area high-resolution models based on deep learning as described in claim 1, characterized in that, The first feature value is the product of the inverse proportional mapping value of the edge degree and the local protrusion degree.
9. The method for correcting precipitation in large-area high-resolution models based on deep learning as described in claim 1, characterized in that, The second feature value is the sum of the differences between the first feature values of each grid at each time point and all times in the historical local time period.
10. A large-area high-resolution model precipitation correction system based on deep learning, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the method as described in any one of claims 1-9.