A storm forecasting method, device and system based on multi-mode integration

Abstract

The application discloses a storm forecast method, device and system based on multi-mode integration, and belongs to the technical field of weather forecast. The method comprises the following steps: acquiring and preprocessing multi-mode precipitation forecast data; calculating the dynamic weight of each mode through the Spearman correlation coefficient based on the historical data of the dynamic sliding window, and generating a precipitation spatial distribution field through weighted average; at the same time, screening an optimal forecast subset based on the symmetric average absolute percentage error, and generating a precipitation intensity field through the interval median of the subset forecast value sorting; finally, replacing the forecast value of the spatial distribution field with the same rank value in the intensity field according to the rank matching, and obtaining the final storm forecast product. The application realizes the double optimization of storm spatial falling area and intensity, significantly improves the forecast accuracy, and reduces the false alarm rate and the missing alarm rate.

Description

Technical Field

[0001] This invention relates to a method, device, and system for forecasting heavy rainfall based on multi-mode integration, belonging to the field of meteorological forecasting technology. Background Technology

[0002] In the field of meteorological forecasting, how to integrate and correct biases using multi-model ensemble forecast data to improve the accuracy of heavy rain event forecasts has become a current research hotspot and an important development direction.

[0003] Traditional multimode integration techniques mainly include frequency matching (FM) and probability matching (PM).

[0004] Frequency matching (FM, Li et al., 2014) is based on the statistical relationship between the occurrence frequencies of forecasts and actual precipitation at different precipitation thresholds, assuming that the frequency of model forecasts at each rainfall level should be consistent with that of the actual precipitation. On this basis, by performing polynomial fitting on the frequency curves of forecasts and actual precipitation, correction coefficients are obtained at each threshold. These coefficients are then used to scale the original precipitation data of the model, making its frequency distribution closer to the actual precipitation. This method can effectively improve the forecast frequency of heavy rainfall and significantly improve the overall bias structure of precipitation levels without changing the location of precipitation areas.

[0005] Probability matching (PM) (Ebert, 2001) is one of the most widely used techniques in multi-model precipitation ensemble methods. Based on precipitation rate samples from ensemble forecast members, it aggregates precipitation forecasts from all model members and constructs their probability distribution functions (PDFs). While maintaining the spatial structure of the ensemble mean field, it reassigns the precipitation intensity at each grid point according to its ranking to the corresponding quantile in the ensemble, thus reconstructing the distribution of forecasted rainfall. This method can effectively calibrate the rainfall distribution of the ensemble mean field, correct the underestimation of maximum rainfall and the overestimation of rain area caused by averaging, and significantly improve the accuracy and characterization of heavy rainfall events.

[0006] However, PM has significant limitations. First, it typically applies static or equal weights to all model members, failing to fully consider the differences in forecast performance and lead time decay characteristics of different models across different timeframes and regions. Second, when dealing with high-impact weather forecasts such as 0-3 day heavy precipitation, PM technology is insufficient in correcting spatiotemporal non-stationary issues commonly found in models, such as inaccurate spatial location and systematic intensity biases. This results in persistent pain points such as high false alarm rates or high missed forecast risks in operational scenarios.

[0007] Therefore, developing a novel integrated technology that can adaptively and dynamically optimize and correct biases of multi-model forecast members and incorporate key spatiotemporal features is of great significance for improving the accuracy and operational applicability of heavy precipitation forecasts. Summary of the Invention

[0008] The purpose of this invention is to provide a method, device, and system for heavy rain forecasting based on multi-model integration. This method aims to effectively solve the technical problems of weak heavy rain forecast intensity, inaccurate spatial location, and high false alarm and missed rate in traditional integration methods by dynamically evaluating model performance, adaptively allocating weights, and selecting forecast members, thereby significantly improving the accuracy and reliability of 0-3 day heavy rain forecasts.

[0009] To achieve the above objectives / to solve the above technical problems, the present invention is implemented using the following technical solution.

[0010] On the one hand, the present invention provides a method for forecasting heavy rainfall based on multi-mode integration, comprising:

[0011] Obtain the latest grid precipitation forecast data for each forecast period from multiple numerical weather prediction models for the target area at the start time of the forecast.

[0012] The grid precipitation forecast data is subjected to spatiotemporal standardization and quality control processing to obtain the latest forecast field of each numerical weather prediction model under the unified station coordinates of the target area;

[0013] A dynamic sliding window is constructed, and based on the historical forecast fields of each numerical weather prediction model and its corresponding historical observation fields within the sliding window, the dynamic sliding weights of each numerical weather prediction model for the current forecast lead time are calculated; the latest forecast fields of each numerical weather prediction model are weighted and averaged using the dynamic sliding weights to generate a precipitation spatial distribution field.

[0014] Within the dynamic sliding window, for each numerical weather prediction model's historical forecast field, the average error during the window period is calculated using the corresponding historical observation field. All numerical weather prediction models are classified according to their average error. Sort the patterns from smallest to largest and select the top K patterns to form the optimal forecast subset;

[0015] Extract the latest forecast field of all numerical weather prediction models in the optimal forecast subset, sort all the numerical values, and extract the median of the sorted values ​​one by one with the number of subset models K as the extraction interval to form the precipitation intensity field.

[0016] The precipitation forecast value of each station in the precipitation spatial distribution field is replaced with a value of the same rank in the precipitation intensity field according to its rank in the precipitation spatial distribution field, and the final integrated rainstorm forecast product is generated and output.

[0017] Furthermore, the spatiotemporal standardization specifically includes:

[0018] Step 1: Spatial Resampling: Select the locations of automatic weather stations within the target area as the spatial reference. Transform the latest forecast-lead time grid precipitation forecast data from each numerical weather prediction model to automatic weather station coordinates using bilinear interpolation. The expression is:

[0019] ;

[0020] in, Precipitation forecast data for the location of the target automatic weather station. For precipitation forecast data at the source grid location, Let be the coordinates of the target automatic weather station's location in the original grid local coordinate system of each numerical weather prediction model. To identify the four source grid points surrounding the target automatic weather station, , These represent the coordinate boundaries of the original mesh in the east-west direction. , These represent the coordinate boundaries of the original mesh in the north-south direction;

[0021] Step 2: Time Consolidation and Alignment: Spatially resampled precipitation forecast data from target automatic weather stations All forecasts should be aligned to the same forecast start time and the same forecast lead time sequence.

[0022] Step 3: Construct a multi-model forecast dataset with completely consistent spatiotemporal references, expressed as follows: ,

[0023] in: For the first m Precipitation forecast data from several numerical weather prediction models at various stations. This refers to the station number of a single automatic weather station within the target area. For the forecast lead time, M is the total number of numerical weather prediction models participating in the integration, S is the set of stations with unified automatic weather station coordinates in the target area, and T is the set of forecast lead time sequences within the forecast period.

[0024] Furthermore, the quality control process specifically includes:

[0025] Negative value correction steps: Replace precipitation values ​​with grid points less than 0 in the spatiotemporally standardized station precipitation forecast data with 0;

[0026] Outlier removal and replacement steps: In the spatiotemporally standardized station precipitation forecast data, precipitation values ​​with grid points exceeding the preset physical threshold are identified as outliers, and the values ​​of valid grid points around the outlier grid point are used for interpolation and replacement.

[0027] Missing value imputation steps: For missing grid points with no records, inverse distance weighted interpolation is used for imputation;

[0028] Wherein: the inverse distance weighted interpolation method searches for valid grid points within a preset radius around the missing grid point, assigns a weight to each valid grid point based on the inverse square of the distance, and calculates the filling value through a weighted average, as expressed in the following expression:

[0029] ;

[0030] ;

[0031] in, To fill in the value, For the first j The value of each valid grid point For missing grid points up to the 1st j The distance of each effective grid point For the first j The weight of the distance to each valid grid point.

[0032] Furthermore, the method for obtaining the dynamic sliding weight specifically includes:

[0033] Construct a dynamic sliding window: for the current forecast date D and forecast lead time Using the past L days as the sliding window length, extract the numerical weather prediction models within the sliding window that have the same start time and forecast lead time. The historical forecast fields and the historical observation fields that are strictly matched in time and space with each historical forecast field constitute the historical training sample set;

[0034] Based on the historical training sample set, the Spearman rank correlation coefficient between the historical forecast field and the historical actual observation field of each numerical weather prediction model is calculated, and used as the dynamic performance coefficient of each numerical weather prediction model.

[0035] Extract the sum of the maximum dynamic performance coefficients of each numerical weather prediction model within a window of L days, and calculate its dynamic sliding weight based on the sum of the maximum dynamic performance coefficients;

[0036] The expression for calculating the Spearman rank correlation coefficient is as follows:

[0037] ;

[0038] ;

[0039] ;

[0040] in: For the first The numerical weather forecast model is in the sliding window. Day, forecast timeframe Spearman rank-correlation dynamic performance coefficients under the following conditions It is a rank transformation function. For the first A numerical weather prediction model has a historical forecast field at the k-th station. For the first Number of valid sites per day For the first The arithmetic mean of the ranks of all historical forecast fields for a numerical weather forecast model. It is the arithmetic mean of the ranks of all observed values ​​at the historical observation site. This represents the historical observation field of the k-th station;

[0041] The expression for calculating the maximum value of the dynamic performance coefficient is as follows:

[0042] ;

[0043] in: M The total number of numerical weather prediction models participating in the integration. It is the first The maximum dynamic performance coefficient of all modes in the day. It is the first Heavenly M The dynamic performance coefficient values ​​of a numerical weather prediction model;

[0044] The expression for calculating the sum of the maximum values ​​of the dynamic performance coefficients is:

[0045] ;

[0046] in: For the first m The sum of the maximum dynamic performance coefficients of each mode within the sliding window L.

[0047] Furthermore, the expression for calculating the dynamic sliding weight is as follows:

[0048] ;

[0049] in: For the first m Dynamic sliding weights for numerical weather prediction models m This refers to the number of the numerical weather prediction model;

[0050] The calculation expression for the spatial distribution field of precipitation is as follows:

[0051] ;

[0052] in, F distribution For the spatial distribution field of precipitation, F m For the first The latest forecast field of a numerical weather prediction model.

[0053] Furthermore, the average error The calculation expression is:

[0054] ;

[0055] in, SMAPE m For the first m The average error of a numerical weather forecast model within a sliding window L. For the first A numerical weather prediction model has a historical forecast field at the k-th station. O k For the first k Historical observation field of each station.

[0056] Furthermore, the multiple numerical weather prediction models include at least two of the following: China Meteorological Administration Operational Guidance Report BABJ, European Centre for Medium-Range Weather Forecasts (ECMWF), National Center for Environmental Prediction (NCEP-GFS), China Meteorological Administration Guangdong Mesoscale Numerical Prediction Model (CMA-GD), China Meteorological Administration Shanghai Mesoscale Numerical Prediction Model (CMA-SH9), and China Meteorological Administration Mesoscale Numerical Prediction Model (CMA-MESO).

[0057] Secondly, the present invention provides a rainstorm forecasting device based on multi-mode integration, comprising:

[0058] The acquisition module is used to acquire the latest grid precipitation forecast data for each forecast period of multiple numerical weather prediction models at the start time of the forecast for the target area, as well as the historical observation fields of the corresponding ground meteorological observation stations.

[0059] Preprocessing module: used to perform spatiotemporal standardization and quality control on the grid precipitation forecast data to obtain the latest forecast field of each numerical weather prediction model under the unified station coordinates of the target area;

[0060] A dynamic sliding window is constructed, and based on the historical forecast fields of each numerical weather prediction model and its corresponding historical observation fields within the sliding window, the dynamic sliding weights of each numerical weather prediction model for the current forecast lead time are calculated; the latest forecast fields of each numerical weather prediction model are weighted and averaged using the dynamic sliding weights to generate a precipitation spatial distribution field.

[0061] The precipitation spatial distribution field module is used to calculate the average error of each numerical weather prediction model's historical forecast field within the window period, using the corresponding historical observation field. All numerical weather prediction models are classified according to their average error. Sort the patterns from smallest to largest and select the top K patterns to form the optimal forecast subset;

[0062] Extract the latest forecast field of all numerical weather prediction models in the optimal forecast subset, sort all the numerical values, and extract the median of the sorted values ​​one by one with the number of subset models K as the extraction interval to form the precipitation intensity field.

[0063] The forecast output module is used to replace the precipitation forecast value of each station in the precipitation spatial distribution field with a value of the same rank in the precipitation intensity field according to its rank in the precipitation spatial distribution field, and generate and output the final integrated rainstorm forecast product.

[0064] Thirdly, the present invention provides a rainstorm forecasting system based on multi-mode integration, comprising:

[0065] Memory, used to store computer programs / instructions;

[0066] A processor is used to execute the computer program / instructions to implement the steps of the above-described multi-mode integrated rainstorm forecasting method.

[0067] Fourthly, the present invention provides a computer-readable storage medium having a computer program / instructions stored thereon, which, when executed by a processor, implements the steps of the above-described method for forecasting heavy rain based on multi-mode integration.

[0068] Fifthly, the present invention provides a computer program product, including a computer program / instruction, which, when executed by a processor, implements the steps of the above-described multi-mode integrated rainstorm forecasting method.

[0069] Compared with existing technologies, the beneficial effects achieved by this invention are as follows: This invention fully explores the differences in historical performance of different models by dynamically allocating weights through a dynamic sliding window and Spearman correlation coefficient, optimizing the forecast of precipitation spatial structure and making the forecast of heavy rainfall areas more accurate. By introducing symmetric mean absolute percentage error as a cost function, the optimal subset that is more skillful in forecasting heavy rainfall is selected from all models. This subset is used to reconstruct the precipitation intensity field, effectively correcting the systematic weakness of heavy rainfall intensity caused by traditional ensemble averaging. By separating and processing spatial distribution (weighted average) and precipitation intensity (preferably median), and finally performing rank matching fusion, this invention significantly improves the forecast of heavy rainfall magnitude while maintaining a reasonable spatial structure, thus demonstrating a lower false alarm rate and missed alarm rate in operational evaluation. Attached Figure Description

[0070] Figure 1 A flowchart illustrating a rainstorm forecasting method based on multi-mode integration provided in an embodiment of the present invention;

[0071] Figure 2 This is a schematic diagram of the dynamic calculation of the sliding weight in an embodiment of the present invention;

[0072] Figure 3 This is a schematic diagram of the generated spatial distribution field of precipitation in an embodiment of the present invention;

[0073] Figure 4 This is a schematic diagram illustrating the generation of the final forecast product through rank matching in an embodiment of the present invention. Detailed Implementation

[0074] It should be noted that:

[0075] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the embodiments of the present invention and the specific features in the embodiments are detailed descriptions of the technical solution of the present invention, rather than limitations thereof. In the absence of conflict, the embodiments of the present invention and the technical features in the embodiments can be combined with each other.

[0076] The term "and / or" simply describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, or B alone. Additionally, the character " / " generally indicates that the preceding and following related objects have an "or" relationship.

[0077] Example 1

[0078] like Figure 1 As shown, a method for forecasting heavy rainfall based on multi-model integration specifically includes the following steps:

[0079] Step 1: Data Acquisition, specifically including:

[0080] The system establishes a connection with the data source through the preset data interface of the "Tianqing Cloud Platform" and acquires the latest 0-3 day 12-hour grid precipitation forecast data of multiple numerical weather prediction models for the target area (range 108°E-115°E, 24°N-30°N) at the start time. The models used in this embodiment include, but are not limited to: China Meteorological Administration Operational Guidance Report BABJ, European Centre for Medium-Range Weather Forecasts ECMWF, National Center for Environmental Prediction NCEP-GFS, China Meteorological Administration Guangdong Mesoscale Numerical Prediction Model CMA-GD, China Meteorological Administration Shanghai Mesoscale Numerical Prediction Model CMA-SH9, and China Meteorological Administration Mesoscale Numerical Prediction Model CMA-MESO.

[0081] Step 2: Perform spatiotemporal standardization and quality control on the data to obtain the latest forecast fields for each numerical weather prediction model under unified station coordinates in the target area. Specifically, this includes:

[0082] Step 21: Spatiotemporal Standardization

[0083] Spatial resampling: The locations of approximately 2000 automatic weather stations within the target area (range 108°E-115°E, 24°N-30°N) are used as a unified spatial reference; the gridded precipitation forecast data from various numerical weather prediction models are interpolated onto the coordinates of these stations using a bilinear interpolation algorithm, expressed as follows:

[0084] ;

[0085] in, Precipitation forecast data for the location of the target automatic weather station. For precipitation forecast data at the source grid location, Let be the coordinates of the target automatic weather station's location in the original grid local coordinate system of each numerical weather prediction model. To identify the four source grid points surrounding the target automatic weather station, , These represent the coordinate boundaries of the original mesh in the east-west direction. , These represent the coordinate boundaries of the original grid in the north-south direction.

[0086] Time alignment: Align all forecast products to the same forecast start time (e.g., 08:00, 20:00) and the same forecast lead time sequence (e.g., the next 12, 24, 36, ..., 72 hours);

[0087] Constructing a multi-model forecast dataset with completely consistent spatiotemporal references, the expression is: ;

[0088] in: For the first m Precipitation forecast data from several numerical weather prediction models at various stations. This refers to the station number of a single automatic weather station within the target area. To provide a forecast timeframe, M S represents the total number of numerical weather prediction models participating in the integration, S represents the set of stations with unified automatic weather station coordinates within the target area, and T represents the set of forecast-by-forecast time-leading sequences within the forecast period.

[0089] Step 22: Quality Control

[0090] Negative value correction steps: Replace precipitation values ​​with grid points less than 0 in the spatiotemporally standardized station precipitation forecast data with 0;

[0091] Outlier removal and replacement steps: Precipitation values ​​in the spatiotemporally standardized station precipitation forecast data where grid points exceed a preset physical threshold are identified as outliers. These outliers are then replaced using interpolated values ​​from valid grid points surrounding the outlier. Outliers are: precipitation forecast values ​​exceeding... 1000mm / 12h (Exceeding my country's record for the highest 12-hour rainfall).

[0092] Missing value imputation steps: For missing grid points with no records, inverse distance weighted interpolation is used for imputation;

[0093] Wherein: the inverse distance weighted interpolation method searches for valid grid points within a preset radius around the missing grid point, assigns a weight to each valid grid point based on the inverse square of the distance, and calculates the filling value through a weighted average, as expressed in the following expression:

[0094] ;

[0095] ;

[0096] in, To fill in the value, For the first j The value of each valid grid point For missing grid points up to the 1st j The distance of each effective grid point For the first j The weight of the distance to each valid grid point.

[0097] Step 3: As Figure 2 As shown, the dynamic sliding weight calculation and spatial distribution field generation specifically include:

[0098] Step 31: Construct a dynamic sliding window: for the current forecast date D and forecast lead time Using the past L days (10-30 days) as the sliding window length, extract data within the window that have the same start time and forecast lead time. The historical forecast fields of each numerical weather prediction model, and the historical observation fields that are strictly matched spatiotemporally and temporally with each historical forecast field, constitute the historical training sample set. ;

[0099] The dynamic sliding window updates daily with the forecast date, always maintaining historical data including the most recent L days. The forecast date D is 0-3 days, and the forecast lead time is... It lasts for 12 hours;

[0100] Step 32: Performance Quantitative Evaluation: Based on the historical training sample set The Spearman rank correlation coefficient (SCC) between the historical forecast fields and historical observed fields of each numerical weather prediction model is calculated and used as the dynamic performance coefficient of each model. The expression is as follows:

[0101] ;

[0102] ;

[0103] ;

[0104] in: For the first The numerical weather forecast model is in the sliding window. Day, forecast timeframe Spearman rank-correlation dynamic performance coefficients under the following conditions It is a rank transformation function. For the first A numerical weather prediction model has a historical forecast field at the k-th station. For the first Number of valid sites per day For the first The arithmetic mean of the ranks of all historical forecast fields for a numerical weather forecast model. It is the arithmetic mean of the ranks of all observed values ​​at the historical observation site. This represents the historical observation field of the k-th station;

[0105] Step 33: For the first m A numerical weather prediction model is selected, and its daily performance coefficient is taken within the dynamic sliding window L days. The sum of the maximum values ​​is denoted as . Specifically:

[0106] The expression for calculating the maximum value of the dynamic performance coefficient is as follows:

[0107] ;

[0108] in:M The total number of numerical weather prediction models participating in the integration. It is the first The maximum dynamic performance coefficient of all modes in the day. It is the first The dynamic performance coefficient value of the Mth numerical weather prediction model;

[0109] The expression for calculating the sum of the maximum values ​​of the dynamic performance coefficients is:

[0110] ;

[0111] in: For the first m The sum of the maximum dynamic performance coefficients of each mode within the sliding window L.

[0112] Step 34: Calculate the dynamic sliding weight: based on each mode Calculate its dynamic sliding weight w m The expression is:

[0113] ;

[0114] in: For the first m Dynamic sliding weights for numerical weather prediction models m This refers to the number of the numerical weather prediction model. This is the sum of the maximum values ​​of the dynamic performance coefficients. M The total number of numerical weather prediction models participating in the integration;

[0115] Step 35: Calculate the spatial distribution field of precipitation, the expression is:

[0116] ;

[0117] in, F distribution For the spatial distribution field of precipitation, F m For the first The latest forecast field of a numerical weather prediction model.

[0118] Step 4: Optimal forecast subset selection and precipitation intensity field generation, specifically including:

[0119] Step 41: Select the cost function

[0120] Within the dynamic sliding window, for each numerical weather prediction model's historical forecast field, the average error of the corresponding model within the window period is calculated using the 12-hour historical observation field from ground meteorological stations that perfectly match the forecast period and target area. The expression is:

[0121] ;

[0122] in, SMAPE m For the first m The average error of a numerical weather forecast model within a sliding window L. For the first A numerical weather prediction model has a historical forecast field at the k-th station, O k This represents the historical observation field of the k-th station;

[0123] Step 42: Member Sorting and Filtering

[0124] All numerical weather prediction models are classified according to their average error SMAPE m Sort the models from smallest to largest and select the top K models to form the optimal forecast subset. Preferably, K is 10% to 30% of the total number of numerical weather prediction models. For example, when n=10, K=3.

[0125] Step 43: Determine the rainfall intensity

[0126] The latest forecast fields from all models within the optimal forecast subset are sorted by magnitude, and the median is taken every K values ​​to obtain the final precipitation intensity field. F intensity The field formed by selecting the median values ​​is the precipitation intensity field. F intensity This represents the information on the intensity of heavy rainfall that is agreed upon by the optimal subset.

[0127] Step 5: As Figure 3 As shown, rank matching and the generation of the final forecast product

[0128] Calculate the rank: calculate separately F distribution field and F intensity The rank of each station (or grid point) value within its own field (i.e., its position after sorting by numerical value).

[0129] Rank substitution: traversal F distribution For each point in the field, find its numerical rank within the field, and then... F intensity Find values ​​with the same rank in the field. If there are values ​​with the same rank, take their average rank.

[0130] Generate the final product: using F intensity Replace the strength value corresponding to the ranking rank in the game.F distribution The original forecast value for that point in the field; after performing this operation on all points, the resulting new field is the final integrated rainstorm forecast product; this process ensures that the final product inherits the spatial distribution pattern optimized by the weighted average field, and also has a more accurate rainstorm intensity magnitude determined by the optimal subset;

[0131] like Figures 2 to 4 As shown, within the dynamic sliding window, the daily Spearman rank correlation coefficient of each numerical weather prediction model is calculated to obtain the maximum value of the daily dynamic performance coefficient; the sum of the maximum values ​​of the dynamic performance coefficients of each model is calculated, and the dynamic sliding weight is calculated based on the sum (1 / 3 for model 1 and 2 / 3 for model 2).

[0132] like Figure 3 As shown, the latest forecast field of each model is multiplied by the dynamic sliding weight to obtain a new spatial distribution field of precipitation.

[0133] like Figure 4 As shown, the rank distribution field is obtained based on the new spatial distribution field of precipitation. The rank distribution is then replaced with an intensity field composed of the median to obtain the final forecast product.

[0134] Product Output: The final product will be converted into the China Meteorological Administration's Micaps Category 3 (grid field) or PNG graphic drawing and other business standard formats for use by forecasters or for subsequent business processes; at the same time, the reserved real-time data can be used to conduct real-time verification of the product using indicators such as TS score and BIAS score.

[0135] The Micaps output format specifications are as follows: File header format (must include the following fields): grid direction, grid type, starting latitude and longitude, ending latitude and longitude, resolution, number of latitude and longitude grid points, ground layer, start time, forecast lead time, feature name, unit, and data type; data are arranged in order from west to east and from south to north, with each value retained to 2 decimal places and separated by spaces.

[0136] Real-time detection and evaluation:

[0137] The TS score and BIAS score for heavy rainfall are calculated as follows:

[0138] ;

[0139] BIAS= ;

[0140] Where: NA represents the number of stations with correct forecasts, NB represents the number of stations with false reports, NC represents the number of stations with missed reports, and ND represents the number of stations with correct forecasts of no heavy precipitation. The heavy rainfall assessment table is shown in Table 1 below:

[0141] Table 1 Rainstorm Inspection and Evaluation Form

[0142] .

[0143] Example 2

[0144] This embodiment provides a rainstorm forecasting device based on multi-mode integration, comprising:

[0145] The acquisition module is used to acquire the latest grid precipitation forecast data for each forecast period of multiple numerical weather prediction models at the start time of the forecast for the target area, as well as the historical observation fields of the corresponding ground meteorological observation stations.

[0146] Preprocessing module: used to perform spatiotemporal standardization and quality control on the grid precipitation forecast data to obtain the latest forecast field of each numerical weather prediction model under the unified station coordinates of the target area;

[0147] A dynamic sliding window is constructed, and based on the historical forecast fields of each numerical weather prediction model and its corresponding historical observation fields within the sliding window, the dynamic sliding weights of each numerical weather prediction model for the current forecast lead time are calculated; the latest forecast fields of each numerical weather prediction model are weighted and averaged using the dynamic sliding weights to generate a precipitation spatial distribution field.

[0148] The precipitation spatial distribution field module is used to calculate the average error of each numerical weather prediction model's historical forecast field within the window period, using the corresponding historical observation field. All numerical weather prediction models are classified according to their average error. Sort the patterns from smallest to largest and select the top K patterns to form the optimal forecast subset;

[0149] Extract the latest forecast field of all numerical weather prediction models in the optimal forecast subset, sort all the numerical values, and extract the median of the sorted values ​​one by one with the number of subset models K as the extraction interval to form the precipitation intensity field.

[0150] The forecast output module is used to replace the precipitation forecast value of each station in the precipitation spatial distribution field with a value of the same rank in the precipitation intensity field according to its rank in the precipitation spatial distribution field, and generate and output the final integrated rainstorm forecast product.

[0151] Example 3

[0152] This embodiment provides a rainstorm forecasting system based on multi-mode integration, including:

[0153] Memory, used to store computer programs / instructions;

[0154] A processor is used to execute the computer program / instructions to implement the steps of the rainstorm forecasting method based on multi-mode integration in Embodiment 1 above.

[0155] Example 4

[0156] This embodiment provides a computer-readable storage medium storing a computer program / instruction thereon, characterized in that, when the computer program / instruction is executed by a processor, it implements the steps of the rainstorm forecasting method based on multi-mode integration in Embodiment 1 above.

[0157] Example 5

[0158] This embodiment provides a computer program product, including a computer program / instruction, which, when executed by a processor, implements the steps of the rainstorm forecasting method based on multi-mode integration described in Embodiment 1 above.

[0159] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0160] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0161] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0162] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0163] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of the present invention without departing from the spirit and scope of the claims. All of these forms are within the protection scope of the present invention.

Claims

1. A method for forecasting heavy rainfall based on multi-model integration, characterized in that, include: Obtain the latest grid precipitation forecast data for each forecast period from multiple numerical weather prediction models for the target area at the start time of the forecast. The grid precipitation forecast data is subjected to spatiotemporal standardization and quality control processing to obtain the latest forecast field of each numerical weather prediction model under the unified station coordinates of the target area; A dynamic sliding window is constructed, and based on the historical forecast fields of each numerical weather prediction model and its corresponding historical observation fields within the sliding window, the dynamic sliding weights of each numerical weather prediction model for the current forecast lead time are calculated; the latest forecast fields of each numerical weather prediction model are weighted and averaged using the dynamic sliding weights to generate a precipitation spatial distribution field. Within the dynamic sliding window, for each numerical weather prediction model's historical forecast field, the average error during the window period is calculated using the corresponding historical observation field. All numerical weather prediction models are classified according to their average error. Sort the patterns from smallest to largest and select the top K patterns to form the optimal forecast subset; Extract the latest forecast field of all numerical weather prediction models in the optimal forecast subset, sort all the numerical values, and extract the median of the sorted values ​​one by one with the number of subset models K as the extraction interval to form the precipitation intensity field. The precipitation forecast value of each station in the precipitation spatial distribution field is replaced with a value of the same rank in the precipitation intensity field according to its rank in the precipitation spatial distribution field, and the final integrated rainstorm forecast product is generated and output.

2. The rainstorm forecasting method based on multi-mode integration according to claim 1, characterized in that, The spatiotemporal standardization specifically includes: Step 1: Spatial Resampling: Select the locations of automatic weather stations within the target area as the spatial reference. Transform the latest forecast-lead time grid precipitation forecast data from each numerical weather prediction model to automatic weather station coordinates using bilinear interpolation. The expression is: ； in, Precipitation forecast data for the location of the target automatic weather station. For precipitation forecast data at the source grid location, Let be the coordinates of the target automatic weather station's location in the original grid local coordinate system of each numerical weather prediction model. To identify the four source grid points surrounding the target automatic weather station, , These represent the coordinate boundaries of the original mesh in the east-west direction. , These represent the coordinate boundaries of the original mesh in the north-south direction; Step 2: Time Consolidation and Alignment: Spatially resampled precipitation forecast data from target automatic weather stations All forecasts should be aligned to the same forecast start time and the same forecast lead time sequence. Step 3: Construct a multi-model forecast dataset with completely consistent spatiotemporal references, expressed as: , Where: represents the station precipitation forecast data of the m-th numerical weather prediction model. This refers to the station number of a single automatic weather station within the target area. For the forecast lead time, M is the total number of numerical weather prediction models participating in the integration, S is the set of stations with unified automatic weather station coordinates in the target area, and T is the set of forecast lead time sequences within the forecast period.

3. The rainstorm forecasting method based on multi-mode integration according to claim 1, characterized in that, The quality control process specifically includes: Negative value correction steps: Replace precipitation values ​​with grid points less than 0 in the spatiotemporally standardized station precipitation forecast data with 0; Outlier removal and replacement steps: In the spatiotemporally standardized station precipitation forecast data, precipitation values ​​with grid points exceeding the preset physical threshold are identified as outliers, and the values ​​of valid grid points around the outlier grid point are used for interpolation and replacement. Missing value imputation steps: For missing grid points with no records, inverse distance weighted interpolation is used for imputation; Wherein: the inverse distance weighted interpolation method searches for valid grid points within a preset radius around the missing grid point, assigns a weight to each valid grid point based on the inverse square of the distance, and calculates the filling value through a weighted average, as expressed in the following expression: ； ； in, To fill in the value, For the first j The value of each valid grid point For missing grid points up to the 1st j The distance of each effective grid point For the first j The weight of the distance to each valid grid point.

4. The rainstorm forecasting method based on multi-mode integration according to claim 1, characterized in that, The method for obtaining the dynamic sliding weight specifically includes: Construct a dynamic sliding window: for the current forecast date D and forecast lead time Using the past L days as the sliding window length, extract the numerical weather prediction models within the sliding window that have the same start time and forecast lead time. The historical forecast fields and the historical observation fields that are strictly matched in time and space with each historical forecast field constitute the historical training sample set; Based on the historical training sample set, the Spearman rank correlation coefficient between the historical forecast field and the historical actual observation field of each numerical weather prediction model is calculated, and used as the dynamic performance coefficient of each numerical weather prediction model. Extract the sum of the maximum dynamic performance coefficients of each numerical weather prediction model within a window of L days, and calculate its dynamic sliding weight based on the sum of the maximum dynamic performance coefficients; The expression for calculating the Spearman rank correlation coefficient is as follows: ； ； ； in: For the first The numerical weather forecast model is in the sliding window. Day, forecast timeframe Spearman rank-correlation dynamic performance coefficients under the following conditions It is a rank transformation function. For the first A numerical weather prediction model has a historical forecast field at the k-th station. For the first Number of valid sites per day For the first The arithmetic mean of the ranks of all historical forecast fields for a numerical weather forecast model. It is the arithmetic mean of the ranks of all observed values ​​at the historical observation site. This represents the historical observation field of the k-th station; The expression for calculating the maximum value of the dynamic performance coefficient is as follows: ； in: M The total number of numerical weather prediction models participating in the integration. It is the first The maximum dynamic performance coefficient of all modes in the day. It is the first Heavenly M The dynamic performance coefficient values ​​of a numerical weather prediction model; The expression for calculating the sum of the maximum values ​​of the dynamic performance coefficients is: ； in: For the first m The sum of the maximum dynamic performance coefficients of each mode within the sliding window L.

5. The rainstorm forecasting method based on multi-mode integration according to claim 4, characterized in that, The expression for calculating the dynamic sliding weight is as follows: ； in: , where m is the number of the numerical weather prediction model; The calculation expression for the spatial distribution field of precipitation is as follows: ； in, F distribution For the spatial distribution field of precipitation, F m For the first The latest forecast field of a numerical weather prediction model.

6. The rainstorm forecasting method based on multi-mode integration according to claim 1, characterized in that, The average error The calculation expression is: ； in, SMAPE m For the first m The average error of a numerical weather forecast model within a sliding window L. For the first A numerical weather prediction model has a historical forecast field at the k-th station. O k For the historical observation field of the k-th station, For the first Number of valid sites per day.

7. A rainstorm forecasting device based on multi-mode integration, characterized in that, include: The acquisition module is used to acquire the latest grid precipitation forecast data for each forecast period of multiple numerical weather prediction models for the target area at the time of the initial report. Preprocessing module: used to perform spatiotemporal standardization and quality control on the grid precipitation forecast data to obtain the latest forecast field of each numerical weather prediction model under the unified station coordinates of the target area; A dynamic sliding window is constructed, and based on the historical forecast fields of each numerical weather prediction model and its corresponding historical observation fields within the sliding window, the dynamic sliding weights of each numerical weather prediction model for the current forecast lead time are calculated; the latest forecast fields of each numerical weather prediction model are weighted and averaged using the dynamic sliding weights to generate a precipitation spatial distribution field. The precipitation spatial distribution field module is used to calculate the average error of each numerical weather prediction model's historical forecast field within the window period, using the corresponding historical observation field. All numerical weather prediction models are classified according to their average error. Sort the patterns from smallest to largest and select the top K patterns to form the optimal forecast subset; Extract the latest forecast field of all numerical weather prediction models in the optimal forecast subset, sort all the numerical values, and extract the median of the sorted values ​​one by one with the number of subset models K as the extraction interval to form the precipitation intensity field. The forecast output module is used to replace the precipitation forecast value of each station in the precipitation spatial distribution field with a value of the same rank in the precipitation intensity field according to its rank in the precipitation spatial distribution field, and generate and output the final integrated rainstorm forecast product.

8. A rainstorm forecasting system based on multi-mode integration, characterized in that, include: Memory, used to store computer programs / instructions; A processor for executing the computer program / instructions to implement the steps of the rainstorm forecasting method based on multi-mode integration as described in any one of claims 1-6.

9. A computer-readable storage medium having a computer program / instructions stored thereon, characterized in that, When the computer program / instruction is executed by the processor, it implements the steps of the rainstorm forecasting method based on multi-mode integration as described in any one of claims 1-6.

10. A computer program product comprising a computer program / instructions, characterized in that, When the computer program / instruction is executed by the processor, it implements the steps of the rainstorm forecasting method based on multi-mode integration as described in any one of claims 1-6.

Images