Rainstorm early warning method based on multi-source forecasting product dynamic fusion

Abstract

The invention relates to the technical field of disaster early warning, and discloses a rainstorm early warning method based on multi-source forecast product dynamic fusion, which comprises the steps of collecting multi-mode rainfall forecast data, constructing a rainfall cumulative distribution function, performing dynamic updating based on Kalman filtering, and obtaining frequency correction rainfall forecast products of each mode; historical forecast and live monitoring data samples are obtained, comprehensive weight coefficients corresponding to all levels of rainfall are calculated, and a fusion initial rainfall distribution field is obtained; performing matching reconstruction on the corrected rainfall forecast field to generate a reconstructed rainfall distribution field; collecting latest real-time monitoring data, and carrying out space-time dynamic correction on the reconstructed rainfall distribution field to generate a real-time rainstorm potential field; and comparing the real-time rainstorm potential field with the multistage rainstorm early warning threshold, generating rainstorm early warning information, and issuing the rainstorm early warning information. According to the invention, the timeliness and accuracy of rainstorm early warning response are improved, the public personal safety is guaranteed, and the city emergency disposal capability is improved.

Description

Technical Field

[0001] This invention relates to the field of disaster early warning technology, and more specifically, to a method for early warning of heavy rain based on the dynamic fusion of multi-source forecast products. Background Technology

[0002] Traditional rainstorm warning mechanisms mostly rely on static threshold triggering or manual judgment, lacking the ability to respond promptly to sudden and highly localized rainstorm events, thus failing to meet the needs. Therefore, developing a system capable of sensing rainstorm potential in real time and triggering warnings related to personal safety has become an important direction for smart meteorology and emergency response coordination.

[0003] However, existing rainstorm monitoring or early warning systems mostly focus on improving the accuracy of meteorological forecasts, lacking a deep integration of forecast results with early warning response mechanisms oriented towards personal safety. For example, current systems often cannot form dynamic judgments with real-time reliability among multi-model forecasts, resulting in low early warning accuracy. On the other hand, most early warning processes lack automatic grading and triggering mechanisms related to disaster intensity, making it difficult to promptly link alarm terminals, mobile device alerts, or public emergency systems.

[0004] Therefore, it is necessary to design a rainstorm early warning method based on the dynamic fusion of multi-source forecast products to solve the problems existing in the current technology. Summary of the Invention

[0005] In view of this, the present invention proposes a rainstorm early warning method based on dynamic fusion of multi-source forecast products, which aims to solve the problems of large magnitude error, inconsistent spatial structure, and low forecast accuracy caused by lack of real-time dynamic adaptation mechanism in the current technology.

[0006] This invention proposes a method for issuing rainstorm warnings based on the dynamic fusion of multi-source forecast products, comprising:

[0007] Global model precipitation forecast data and mesoscale model precipitation forecast data are collected. The cumulative precipitation distribution function between historical precipitation forecast data and corresponding real-time monitoring data of each model is constructed, and the cumulative precipitation distribution function is dynamically updated based on Kalman filtering.

[0008] Based on the updated precipitation cumulative distribution function, quantile linear interpolation is used to correct the frequency of the current lead time precipitation forecast values ​​of each model, and the frequency-corrected precipitation forecast products of each model are obtained.

[0009] Historical precipitation forecast data and actual monitoring data samples from 30 days prior to the target forecast date and 30 days after the same period last year are obtained. Threat scores and bias scores of each model at different precipitation levels are calculated according to the set precipitation classification thresholds. Comprehensive weight coefficients corresponding to each precipitation level are generated. Based on the comprehensive weight coefficients, the frequency correction precipitation forecast products of each model are weighted and fused to obtain the initial fused precipitation distribution field.

[0010] Using the precipitation grid intensity ranking of the fused initial precipitation distribution field as a reference, spatial probability matching is performed on the frequency correction precipitation forecast products of each model according to the ranking to generate a reconstructed precipitation distribution field.

[0011] Collect the latest real-time monitoring data, calculate the spatial structure deviation of precipitation between the latest real-time monitoring data and the reconstructed precipitation distribution field, perform spatiotemporal dynamic correction on the reconstructed precipitation distribution field, and generate a real-time rainstorm potential field;

[0012] A multi-level rainstorm warning threshold is set, the real-time rainstorm potential field is compared with the multi-level rainstorm warning threshold, rainstorm warning information is generated, and the rainstorm warning information is released to users through audio-visual equipment, mobile terminals or emergency broadcasts.

[0013] Furthermore, when constructing the cumulative precipitation distribution function between historical precipitation forecast data and corresponding real-time monitoring data for each model, and dynamically updating the cumulative precipitation distribution function based on Kalman filtering, the process includes:

[0014] The system collects global model precipitation forecast data and mesoscale model precipitation forecast data within a historical window set before the target forecast date, as well as real-time monitoring data for the corresponding time period.

[0015] Spatial registration and gridding of historical precipitation forecast data and real-time monitoring data are performed to establish grid-point-by-grid pairs of historical forecast and real-time monitoring data.

[0016] The historical precipitation forecast values ​​and actual precipitation values ​​for each model are sorted in ascending order to construct a cumulative precipitation distribution function;

[0017] During the construction process, the cumulative precipitation distribution function is updated recursively using Kalman filtering based on the latest real-time monitoring data and precipitation forecast values.

[0018] Furthermore, when recursively updating the cumulative precipitation distribution function based on Kalman filtering, the following steps are included:

[0019] The cumulative precipitation distribution function of each model is expressed as precipitation intensity values ​​corresponding to multiple precipitation locations;

[0020] The precipitation intensity value is used as the state variable for Kalman filtering;

[0021] Based on the deviation between the newly added precipitation forecast value and the actual precipitation value, the state variable is recursively corrected through the prediction and update steps of Kalman filtering.

[0022] Furthermore, when performing frequency correction based on the updated cumulative precipitation distribution function, it includes:

[0023] Calculate the current time-leading precipitation forecast value for each precipitation model, and calculate the cumulative probability value in the cumulative precipitation distribution function corresponding to that model;

[0024] The cumulative probability value is mapped to the corresponding actual precipitation cumulative distribution function, and the equivalent precipitation value in the actual space is determined by quantile linear interpolation.

[0025] The equivalent precipitation value is used as the result after frequency correction to generate the frequency-corrected precipitation forecast products for each model.

[0026] Furthermore, when calculating the threat score and bias score of each model at different precipitation levels according to the set precipitation classification thresholds, and generating the comprehensive weighting coefficients corresponding to each precipitation level, the following are included:

[0027] Several precipitation level thresholds are set, and the historical precipitation forecast data and corresponding real-time monitoring data of each model are converted into hit, miss and false alarm events at each level.

[0028] Based on the number of hits, the number of missed reports, and the number of false alarms, the threat score and bias score for each model at each precipitation level are calculated. The threat score is the ratio of the number of hit times to the total number of hit, missed, and false alarm events, and the bias score is the ratio of the total number of hits and false alarms to the total number of hits and missed reports.

[0029] The threat score and bias score are normalized, and weighting coefficients are set for different levels of precipitation. The weighting coefficient of the threat score is greater than that of the bias score when the precipitation level is heavy, and the weighting coefficient of the threat score is less than that of the bias score when the precipitation level is light, thereby generating a comprehensive weighting coefficient for each model at each precipitation level.

[0030] Furthermore, when obtaining the fused initial precipitation distribution field by weighting the frequency-corrected precipitation forecast products of each model according to the aforementioned comprehensive weighting coefficients, the process includes:

[0031] For each mode of frequency-corrected precipitation forecast product at the current time, the comprehensive weight coefficient for the corresponding level is selected from the comprehensive weight coefficients according to the precipitation level at which the current precipitation value is located.

[0032] The frequency-corrected precipitation value of each model is multiplied by its corresponding comprehensive weight coefficient, and the fused precipitation value of the current grid point is calculated by weighting all models.

[0033] A weighted fusion is performed on all grid points to obtain a fused initial precipitation distribution field containing spatial distribution information.

[0034] Furthermore, using the precipitation grid intensity ranking of the fused initial precipitation distribution field as a reference, spatial probability matching is performed on the frequency-corrected precipitation forecast products of each model according to the ranking to generate the reconstructed precipitation distribution field, including:

[0035] Sort all grid points of the fused initial precipitation distribution field from high to low precipitation intensity and record the sorting position;

[0036] The precipitation values ​​of each model frequency-corrected precipitation forecast product are sorted from high to low intensity to generate a precipitation intensity sequence.

[0037] The precipitation intensity sequence is reassigned to the corresponding grid points according to the grid point sorting position of the initial precipitation distribution field, thus forming a reconstructed precipitation distribution field with a unified spatial structure.

[0038] Furthermore, when calculating the spatial structure deviation of precipitation between the latest real-time monitoring data and the reconstructed precipitation distribution field, and performing spatiotemporal dynamic correction on the reconstructed precipitation distribution field to generate a real-time rainstorm potential field, the process includes:

[0039] The latest real-time monitoring data includes observation data from automatic weather stations, radar echo images, satellite cloud images, and wind profiler radar.

[0040] Calculate the spatial structure deviation of precipitation between the latest real-time monitoring data and the reconstructed precipitation distribution field. The spatial structure deviation of precipitation includes spatial offset distance, structural similarity, and difference in the location of heavy precipitation centers.

[0041] A spatially corrected weight field is constructed based on the aforementioned spatial structure deviation of precipitation, and the reconstructed precipitation distribution field is dynamically corrected in spatiotemporal manner on a grid-by-grid basis.

[0042] The real-time rainstorm potential field is obtained by combining the radar strong echo area, the wind field uplift motion area, and the high instability energy area.

[0043] Furthermore, a multi-level rainstorm warning threshold is set, and the real-time rainstorm potential field is compared with the multi-level rainstorm warning threshold. When generating rainstorm warning information, the process includes:

[0044] The rainstorm warning thresholds are divided into several levels based on the severity of the rainstorm, including blue, yellow, orange, and red levels.

[0045] The grid values ​​of each point in the real-time rainstorm potential field are compared with each threshold level step by step to identify the spatial location and level category that meet the warning triggering conditions, and to generate rainstorm warning information that includes the warning level, warning area, trigger time and expected duration.

[0046] Furthermore, after generating a rainstorm warning, it also includes:

[0047] Several rolling update time periods are set, including 1 hour, 3 hours and 6 hours;

[0048] The real-time rainstorm potential field is recalculated and compared in each rolling update period to generate rolling updated multi-time-effect rainstorm early warning information.

[0049] The warning level or warning area will be dynamically adjusted based on the development trend of heavy precipitation, the rate of change of potential value, and the degree of proximity of threshold for each time period.

[0050] Compared with existing technologies, the advantages of this invention are as follows: By constructing a multi-source precipitation forecasting system that integrates global and mesoscale models, combining historical data to build a dynamically updatable cumulative precipitation distribution function, and introducing a frequency correction mechanism based on quantile linear interpolation, accurate correction of precipitation forecast results from different models is achieved. A weighted fusion method driven by precipitation level threat scoring is introduced, and spatial probability matching and spatiotemporal dynamic correction further enhance the ability of the reconstructed precipitation field to restore the spatial structure of severe convective weather. Unlike traditional technologies that only output static warning information, this invention constructs a multi-level rainstorm warning threshold system and automatically compares it with the real-time rainstorm potential field to generate warning signals with clear triggering conditions and distinct levels. Finally, multi-channel alarm linkage is achieved through audio-visual equipment, mobile terminals, or emergency broadcasts, improving the timeliness, accuracy, and practicality of rainstorm warning response, effectively protecting public safety, and enhancing urban emergency response capabilities. Attached Figure Description

[0051] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings:

[0052] Figure 1 A flowchart of a rainstorm early warning method based on dynamic fusion of multi-source forecast products provided in an embodiment of the present invention. Detailed Implementation

[0053] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the disclosure to those skilled in the art. It should be noted that, unless otherwise specified, embodiments and features in the embodiments of the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0054] In traditional rainstorm warning systems, models relying on static threshold triggering mechanisms and manual experience-based judgment are ill-suited to the spatiotemporal dynamics of sudden heavy rainfall events. Systematic biases among multi-model numerical weather prediction products cannot be effectively eliminated, and the performance differences between models at different rainfall levels are not quantitatively assessed, leading to spatial structural distortions in the fusion results. Insufficient spatiotemporal matching between forecast products and real-time monitoring data prevents dynamic correction of the potential field, causing warning information to lag behind the actual evolution of the weather system.

[0055] For example, when rapidly developing convective clouds appear in the upper reaches of a river basin, the statistical relationships constructed by existing systems based on historical data within fixed time windows cannot reflect the abrupt changes in current atmospheric boundary layer conditions. Global and mesoscale models capture precipitation signals at different scales, but without a dynamic weighting mechanism, the location and intensity of heavy precipitation centers are smoothed during the fusion process. Forecasters need to manually compare radar echoes with model output fields, but due to limitations in data processing speed, it is difficult to identify gridded areas reaching warning thresholds in the early stages of strong convection development. At this point, the precipitation distribution field output by the system has a spatial offset of more than 20 kilometers from the measured data from automatic weather stations, and the magnitude of short-duration heavy precipitation is underestimated by more than 30%.

[0056] If the above problems are not addressed, the synergistic effect among multi-source data will not be fully realized, and systematic errors in models will be nonlinearly amplified during the fusion process, leading to a decrease in the spatial resolution of the rainstorm potential field. The spatiotemporal discrepancies between early warning information and the actual precipitation field will continue to accumulate, causing the missed reporting rate of orange-level and above warnings to rise to dangerous levels. Emergency response departments will struggle to obtain timely and accurate forecasts of heavy rainfall areas, making it impossible to complete the evacuation of people and traffic control in high-risk areas before disasters occur, increasing the risk of people being trapped when urban low-lying areas experience flooding.

[0057] For this, please refer to Figure 1 As shown, this application proposes a method for issuing rainstorm warnings based on the dynamic fusion of multi-source forecast products, including:

[0058] S100: Collect global model precipitation forecast data and mesoscale model precipitation forecast data, construct the precipitation cumulative distribution function between the historical precipitation forecast data of each model and the corresponding real-time monitoring data, and dynamically update the precipitation cumulative distribution function based on Kalman filtering;

[0059] S200: Based on the updated cumulative precipitation distribution function, quantile linear interpolation is used to correct the frequency of the current lead time precipitation forecast values ​​of each model, and frequency-corrected precipitation forecast products of each model are obtained.

[0060] S300: Acquire historical precipitation forecast data and actual monitoring data samples for 30 days prior to the target forecast date and 30 days after the same period last year. Calculate the threat score and bias score of each model under different precipitation levels according to the set precipitation classification threshold. Generate the comprehensive weight coefficient corresponding to each precipitation level. Perform weighted fusion of the frequency correction precipitation forecast products of each model based on the comprehensive weight coefficient to obtain the initial fused precipitation distribution field.

[0061] S400: Using the precipitation grid intensity ranking of the fused initial precipitation distribution field as a reference, spatial probability matching is performed on the frequency correction precipitation forecast products of each model according to the ranking to generate a reconstructed precipitation distribution field.

[0062] S500: Collect the latest real-time monitoring data, calculate the spatial structure deviation of precipitation between the latest real-time monitoring data and the reconstructed precipitation distribution field, perform spatiotemporal dynamic correction on the reconstructed precipitation distribution field, and generate a real-time rainstorm potential field;

[0063] S600: Set multi-level rainstorm warning thresholds, compare the real-time rainstorm potential field with the multi-level rainstorm warning thresholds, generate rainstorm warning information, and release the rainstorm warning information to users through audio-visual equipment, mobile terminals or emergency broadcasts.

[0064] Specifically, dynamic updates refer to the recursive correction of the cumulative precipitation distribution function based on Kalman filtering. This can be achieved by using quantile precipitation intensity as a state variable and implementing a prediction-update cycle with new samples. Dynamic updates can adapt to seasonal variations in precipitation models and sudden weather events, improving the timeliness of the distribution function. Quantile linear interpolation maps the cumulative probability of model forecasts to corresponding quantiles in the actual distribution function. This can be achieved by establishing a probability transfer relationship between historical forecasts and actual CDF curves, and processing discontinuous quantiles through linear interpolation. Quantile linear interpolation can eliminate systematic biases in models and achieve probability space calibration of forecast values. Threat scoring is a scoring index calculated based on hit rate, false alarm rate, and false alarm rate. The bias score is the ratio of the total number of hits / false alarms to the total number of hits / false alarms. This can be achieved using a binary classification event statistical method, dividing the test samples according to a preset precipitation level. Threat scoring can quantify the differences in forecasting capabilities of different models across various precipitation intensity ranges, providing an objective weighting basis for multi-source fusion. Spatial probability matching reconstruction refers to the redistribution of model forecast values ​​according to the intensity ranking of the fused field. Specifically, it can maintain spatial structural consistency by aligning grid points. Spatial probability matching reconstruction can eliminate spatial morphological differences between multiple models and preserve the spatial coherence of heavy precipitation signals. Spatiotemporal dynamic correction refers to adjusting the precipitation field using the deviation matching degree between real-time monitoring data and the reconstructed field. Specifically, it can use spatial offset calculation, structural similarity analysis, and comparison of the location of heavy precipitation centers. Spatiotemporal dynamic correction can fuse real-time observation data from radar, satellites, etc., to capture the evolution characteristics of sudden heavy precipitation systems. Multi-level rainstorm warning thresholds refer to graded trigger thresholds based on rainstorm intensity. Specifically, it can use the rainfall intensity thresholds corresponding to blue, yellow, orange, and red warnings in the national meteorological standards. Multi-level rainstorm warning thresholds can achieve disaster risk classification assessment and match the handling needs of different emergency response levels.

[0065] This application constructs a full-chain rainstorm early warning system, from data calibration and multi-model weighted fusion to spatial morphology optimization, by dynamically fusing and reconstructing multi-source forecast products and combining them with a spatiotemporal correction mechanism for real-time observation data. This enables rapid identification and graded early warning response to sudden heavy rainfall events.

[0066] The working process and principle of this application are as follows: Precipitation forecast data from global and mesoscale models are collected, and a cumulative precipitation distribution function is constructed between historical precipitation forecast data and corresponding real-time monitoring data for each model. The cumulative precipitation distribution function is dynamically updated using Kalman filtering to reflect real-time changes in atmospheric conditions. Based on the updated cumulative precipitation distribution function, quantile linear interpolation is used to perform frequency correction on the current-leading precipitation forecast values ​​of each model, eliminating systematic biases.

[0067] Historical precipitation forecast data and actual monitoring data samples from the 30 days prior to the target forecast date and the 30 days following the same period last year were acquired. Threat scores and bias scores for each model under different precipitation levels were calculated according to set precipitation grading thresholds. A comprehensive weighting coefficient corresponding to each precipitation level was generated. Based on this comprehensive weighting coefficient, the frequency-corrected precipitation forecast products from each model were weighted and fused to obtain the initial fused precipitation distribution field. This enabled a quantitative evaluation of model performance under different precipitation levels.

[0068] Using the precipitation grid intensity ranking of the fused initial precipitation distribution field as a reference, spatial probability matching reconstruction is performed on the frequency-corrected precipitation forecast products of each model to generate a reconstructed precipitation distribution field. The spatial structure is unified through ranking and assignment, resolving the spatial offset problem between forecasts and actual data.

[0069] The latest real-time monitoring data is collected, and the reconstructed precipitation distribution field is dynamically corrected in time and space based on the spatial deviation matching degree between the latest real-time monitoring data and the reconstructed precipitation distribution field, generating a real-time heavy rainfall potential field. By introducing real-time monitoring data, the evolution trend of heavy precipitation is captured, thus improving the problem of early warning lag.

[0070] A multi-level rainstorm warning threshold is set, and the real-time rainstorm potential field is compared with the multi-level rainstorm warning threshold to generate rainstorm warning information. The rainstorm warning information is then disseminated to users through audio-visual equipment, mobile terminals, or emergency broadcasts, achieving timely delivery of warning information.

[0071] As a preferred embodiment, the solution of this application is specifically implemented as follows:

[0072] Precipitation forecast data from global and mesoscale models were collected. For each model, a cumulative precipitation distribution function was constructed between historical precipitation forecast data and corresponding real-time monitoring data. The cumulative precipitation distribution function was dynamically updated using a Kalman filter algorithm to adapt to real-time changes in atmospheric conditions.

[0073] Based on the updated cumulative precipitation distribution function, frequency corrections are performed on the current-leading precipitation forecasts for each model. Specifically, the cumulative probability of the current forecast value in the model's cumulative distribution function is calculated, and then the corresponding precipitation value is found in the actual cumulative distribution function, which is used as the result of the frequency correction.

[0074] Acquire historical precipitation forecast data and actual monitoring data samples for the 30 days prior to the target forecast date and the 30 days following the same period last year. Set multiple precipitation grading thresholds, such as 0.1mm, 1mm, 3mm, 5mm, 10mm, 15mm, 20mm, 30mm, and 50mm. For each model, calculate threat score and bias score at each precipitation level. The threat score is defined as the ratio of the number of hit events to the total number of hit, missed, and false alarm events, and the bias score is the ratio of the total number of hits and false alarms to the total number of hits and missed alarms.

[0075] A comprehensive weighting coefficient is generated based on the threat score and bias score at each level. A weighted average is then performed on the precipitation forecast products after frequency correction from each model to obtain the fused initial precipitation distribution field.

[0076] Using the precipitation grid intensity ranking of the fused initial precipitation distribution field as a reference, spatial probability matching reconstruction is performed on the frequency-corrected precipitation forecast products of each model. Specifically, all grid points in the fused initial field are sorted from highest to lowest precipitation intensity, and the ranking positions are recorded. The same operation is performed on the frequency-corrected precipitation forecast fields of each model. Then, values ​​are reassigned according to the grid point ranking positions of the fused initial field to generate the reconstructed precipitation distribution field.

[0077] The latest real-time monitoring data is collected, including automatic weather station observations, radar echo images, satellite cloud images, and wind profiler radar data. The spatial structure deviation between the latest real-time monitoring data and the reconstructed precipitation distribution field is calculated, including spatial offset distance, structural similarity, and differences in the location of heavy precipitation centers. A spatially corrected weight field is constructed based on the spatial structure deviation to dynamically correct the reconstructed precipitation distribution field on a grid-by-grid basis. A real-time heavy rainfall potential field is generated by combining radar strong echo areas, wind uplift motion areas, and high instability energy regions.

[0078] Set multi-level rainstorm warning thresholds, such as blue, yellow, orange, and red levels. Compare the real-time rainstorm potential field grid values ​​with each threshold level step by step to identify spatial locations and level categories that meet the warning trigger conditions. Generate rainstorm warning information including the warning level, warning area, trigger time, and expected duration. Disseminate the rainstorm warning information to users via audio-visual equipment, mobile terminals, or emergency broadcasts.

[0079] Through the above-described scheme, this application achieves dynamic fusion of multi-source forecast products, improving the accuracy and timeliness of heavy rain warnings. By dynamically updating the cumulative distribution function, it reflects real-time changes in atmospheric conditions, solving the problem that traditional static threshold mechanisms cannot adapt to the dynamic characteristics of sudden heavy precipitation. A graded threat scoring mechanism quantifies model performance under different precipitation levels, improving the reliability of the fusion results. Spatial probability matching reconstruction technology unifies the spatial structure, effectively solving the spatial offset problem between forecast and actual monitoring data. A dynamic correction module based on real-time monitoring data captures the evolution trend of heavy precipitation, reducing warning lag. The setting of multi-level warning thresholds and the multi-channel warning information dissemination mechanism improve the accuracy and coverage of warning information.

[0080] In some of the above-mentioned schemes of this application, a method for dynamic fusion of forecast data is proposed by constructing a cumulative precipitation distribution function. However, in the process of constructing the empirical cumulative distribution function, due to the uneven spatiotemporal distribution of historical data samples or insufficient data volume, the cumulative distribution function is prone to statistical instability in the low probability interval, which in turn affects the mapping accuracy of extreme precipitation values ​​in the frequency correction process.

[0081] This application further proposes collecting global and mesoscale model precipitation forecast data within a historical window range set before the target forecast date, along with corresponding real-time monitoring data. Spatial registration and gridding are performed on the historical precipitation forecast data and real-time monitoring data to establish grid-by-grid pairs of historical forecast and real-time monitoring data. Historical and real-time precipitation forecast values ​​for each model are sorted in ascending order to construct a cumulative precipitation distribution function (CDF). During the construction of the CDF, it is recursively updated using Kalman filtering based on the latest real-time monitoring data and precipitation forecast values. Precipitation forecast values ​​for a given model within a specific grid point are sorted in ascending order to form an empirical distribution; the corresponding real-time precipitation values ​​are also sorted in ascending order. The empirical CDF curve between model forecasts and real-time precipitation is obtained by statistically analyzing the quantile of each precipitation value. If the data is discontinuous or the sample size is insufficient, kernel density estimation can be used to smooth the CDF.

[0082] Spatial registration and gridding employ a bilinear interpolation algorithm to unify data of different resolutions into a standard grid within the same geographic coordinate system, ensuring spatial comparability of data from different models. During the construction of the empirical cumulative distribution function, a sliding time window mechanism dynamically adjusts the historical data sample size, automatically expanding the time range when the number of samples within the window falls below a preset threshold. Kernel density estimation uses a Gaussian kernel function for probability density smoothing, and the bandwidth parameter is automatically optimized and adjusted based on the sample standard deviation, effectively addressing the oscillation problem of the distribution curve under small sample conditions.

[0083] Specifically, when establishing historical precipitation forecast data and actual monitoring data pairs grid-by-grid, geocoding is used to map precipitation data from different sources to a unified geographic grid coordinate system, and spatial interpolation is used to eliminate resolution differences between different data sources. During the ascending order sorting of the cumulative distribution function, a separate sorting sequence is established for each grid point to ensure the effective preservation of local precipitation characteristics. When new forecast and actual monitoring data pairs arrive, the statistical characteristics of the new samples are integrated into the cumulative distribution function in real time through a Kalman filter state variable update mechanism. In sparse data regions, kernel density estimation generates a smooth cumulative distribution curve through probability density extrapolation, avoiding distribution function discontinuities caused by missing samples. For example, when a grid point has only 5 valid samples within the historical window, a kernel density estimator with a bandwidth of 1.06 times the sample standard deviation is used to generate a continuous probability distribution curve, ensuring the numerical stability of subsequent quantile mapping processes.

[0084] As a preferred embodiment, the solution of this application is specifically implemented as follows:

[0085] When constructing the cumulative precipitation distribution function between historical precipitation forecast data and corresponding real-time monitoring data for each model, and dynamically updating the cumulative distribution function based on Kalman filtering, the following steps are included:

[0086] Global and mesoscale model forecasts for the 30 days prior to the target forecast date were collected, along with real-time monitoring data for the corresponding period.

[0087] Spatial registration and gridding were performed on historical forecast data and actual monitoring data to establish grid-point-by-grid pairs of historical forecast and actual monitoring data. Specifically, forecast data and actual monitoring data were uniformly interpolated onto a 0.1×0.1 grid.

[0088] For each model, historical precipitation forecasts and actual precipitation values ​​are sorted in ascending order to construct a cumulative precipitation distribution function. For example, for a given grid point, the forecast and actual values ​​over 30 days are sorted separately to obtain their respective cumulative distribution functions.

[0089] In constructing the cumulative distribution function, based on the latest real-time monitoring data and precipitation forecast values, the cumulative distribution function is recursively updated using Kalman filtering. Specifically, each time a new set of forecast-real-time monitoring data pairs is added, the parameters of the cumulative distribution function are updated using the Kalman filtering algorithm, so that the cumulative distribution function can dynamically reflect the latest forecast-real-time relationship.

[0090] Through the above technical solution, this application enables dynamic updating of the cumulative precipitation distribution function, improving its ability to represent the latest forecast-actual relationship. Therefore, frequency correction based on the dynamically updated cumulative distribution function can more accurately correct forecast biases in various models, thereby improving the accuracy of heavy rain warnings. Furthermore, employing a Kalman filter algorithm for recursive updates maintains computational efficiency while effectively utilizing historical data, avoiding instability in distribution function estimation due to insufficient sample size.

[0091] In some of the above-mentioned schemes in this application, Kalman filtering is proposed to dynamically update the cumulative distribution function. However, in practical applications, there is a problem of dynamic bias accumulation in the process of updating the intensity values ​​of the quantiles of the cumulative distribution function. This causes the precipitation intensity values ​​corresponding to the quantiles to fail to accurately reflect the statistical relationship between the latest forecast and the actual monitoring data, affecting the accuracy of the subsequent frequency correction process.

[0092] This application further proposes to represent the cumulative distribution function of each precipitation forecast model as a set of precipitation intensity values ​​corresponding to precipitation quantiles, and to use the quantile intensity values ​​as the state variables of the Kalman filter. Based on the deviation between the newly added precipitation forecast values ​​and the actual precipitation values, the state variables are recursively corrected through the prediction update of the Kalman filter.

[0093] In this system, precipitation quantiles are set with fixed probability intervals, such as 0.05, 0.1, 0.2 to 0.95, with each quantile corresponding to a precipitation intensity value as an independent state variable. The Kalman gain matrix is ​​dynamically adjusted based on the historical error covariance, and the observation mapping matrix maps the bias of new samples to the corresponding quantiles. During the recursive update process, the precipitation intensity value of each quantile is predicted and corrected independently to avoid cross-quantile interference.

[0094] Specifically, the precipitation intensity value corresponding to a fixed probability quantile is defined as a state vector, for example, Q_p represents the precipitation value at the p-th quantile. During each update, newly added forecast and actual monitoring data pairs are mapped to the nearest neighbor quantiles using interpolation or extrapolation methods, generating observation vectors. The Kalman filter's prediction step calculates the current predicted value based on the state estimation error covariance of the previous moment, while the update step adjusts the state variables using the deviation between the observation vector and the predicted value. For example, when the actual precipitation value of a newly added sample is near the 0.3 quantile, only the state values ​​of that quantile and its adjacent quantiles are updated. During the recursive process, the covariance matrix of the state variables uses an exponential decay factor to control the weighting of historical data, ensuring the system's response speed to the latest data. Through independent point-by-point updates, the cumulative distribution function dynamically adapts to changes in the relationship between forecasts and actual data while maintaining its overall shape.

[0095] As a preferred embodiment, the solution of this application is specifically implemented as follows:

[0096] The cumulative distribution function (CDF) curve is represented as a mapping relationship between a fixed set of probability quantiles and corresponding precipitation values. For example, 11 quantiles are selected, namely 0.05, 0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9, and 0.95, and the precipitation value Q_p corresponding to each quantile is used as a filter state variable.

[0097] During each update, the impact of newly added forecast-actual sample pairs is mapped to the corresponding quantiles. Specifically, for a new sample (x_new, y_new), the cumulative probability p_new corresponding to x_new is found in the forecast CDF, and then y_new is compared with the original precipitation value at p_new in the actual CDF to calculate the difference δ.

[0098] The prediction-update formula based on Kalman filtering performs point-by-point recursive updates. In the prediction step, the state is assumed to remain unchanged; in the update step, the observation influence δ is distributed to neighboring quantiles through the observation mapping matrix H, where H can employ a triangular weighting function. The Kalman gain K_t is dynamically adjusted based on the prediction error covariance and the observation error covariance.

[0099] Finally, the updated quantile precipitation values ​​form a new CDF curve. In this way, the CDF can be dynamically adjusted with the addition of new samples, maintaining historical statistical characteristics while reflecting the latest changes in the forecast-actual relationship.

[0100] Through the above technical solution, this application enables dynamic updating of the cumulative distribution function, improving the adaptability and accuracy of frequency correction. The CDF curve can adaptively adjust with the addition of new samples, maintaining the stability of historical statistical characteristics while promptly reflecting the latest changes in the forecast-actual relationship. This dynamic updating mechanism allows the frequency correction process to better adapt to changes in weather models and improvements in forecast systems, thereby improving the accuracy and timeliness of heavy rain warnings.

[0101] In some of the schemes described above in this application, when performing frequency correction based on the cumulative distribution function, there may be situations where extreme precipitation values ​​exceed the range of historical samples, leading to distorted interpolation results. At the same time, discontinuous historical data or insufficient samples may cause probability mapping bias, affecting the correction accuracy.

[0102] This application further proposes to construct two cumulative distribution functions based on historical samples: the cumulative distribution function of the model's historical forecast values ​​and the cumulative distribution function of the corresponding actual precipitation values. A probability mapping relationship is established, and quantile linear interpolation correction is performed on the current forecast values. When dealing with boundary problems and outliers, boundary extension, threshold pruning, or robust interpolation algorithms are adopted.

[0103] The cumulative distribution function of historical forecasts is generated by arranging historical forecast data in ascending order, while the cumulative distribution function of actual precipitation is generated by arranging actual monitoring data for the corresponding time period in ascending order. The probability mapping process maps the cumulative probability values ​​of forecasts to the quantile values ​​corresponding to the actual distribution function. When the cumulative probability of a forecast value lies between two known quantiles, linear interpolation is used to determine the equivalent precipitation value. For cumulative probability values ​​exceeding the domain of the actual distribution function, boundary extension or threshold pruning is used to limit their range. A sliding window extrapolation algorithm is introduced in extreme precipitation regions, using the distribution trends of neighboring data points for extrapolation interpolation.

[0104] Specifically, model forecast values ​​are input into the historical cumulative distribution function (CFD) to calculate their corresponding cumulative probability values. These probability values ​​are then used as a mapping benchmark and passed to the actual cumulative distribution function (ECD). If the probability value in the ECD lies between two adjacent quantiles, the equivalent precipitation value is calculated using linear interpolation. When the cumulative probability corresponding to the forecast value exceeds the maximum defined range of the ECD, a threshold is used to limit it to the maximum actual precipitation value. For low-probability, high-intensity extreme precipitation forecasts, a sliding window algorithm is used to extract the distribution characteristics of adjacent high-value intervals, and the equivalent precipitation value is estimated using extrapolation. This process eliminates systematic biases in the model through probability mapping of the dual distribution functions, ensures data continuity through linear interpolation, prevents distortion of correction results through boundary handling mechanisms, and enhances the robustness of extreme precipitation corrections through sliding window extrapolation.

[0105] As a preferred embodiment, the solution of this application is specifically implemented as follows:

[0106] Based on the cumulative distribution function, quantile linear interpolation is used to perform frequency correction on the current-lead time precipitation forecasts of each model, obtaining frequency-corrected precipitation forecast products for each model. Specifically, for the current-lead time forecast value of each precipitation model, the cumulative probability value is calculated in the precipitation cumulative distribution function corresponding to that model. Further, the cumulative probability value is mapped to the corresponding observed precipitation cumulative distribution function, and the equivalent precipitation value in the observed space is determined by quantile linear interpolation. Thus, the equivalent precipitation value is used as the result of frequency correction to generate the frequency-corrected precipitation forecast products for each model.

[0107] For example, if a global model forecasts 50 mm of precipitation at a specific grid point, the first step is to find the cumulative probability value corresponding to 50 mm in the model's historical cumulative distribution function, let's say 0.85. Then, in the observed precipitation cumulative distribution function, the precipitation value corresponding to a probability of 0.85 is found. If 0.85 falls between two known probability points, the corresponding precipitation value is calculated using linear interpolation, let's say 45 mm. Finally, 45 mm is used as the frequency-corrected precipitation forecast value for that grid point. Repeating this process for all grid points yields the complete frequency-corrected forecast field.

[0108] Through the above technical solutions, this application achieves frequency bias correction for various numerical model forecast products, improving the accuracy of the forecast products. Quantile linear interpolation ensures the smoothness and continuity of the correction process, avoiding the discontinuity problems that may arise from step correction. Furthermore, correction based on a dynamically updated cumulative distribution function allows the correction results to adapt to dynamic changes in model performance in a timely manner, improving the timeliness and adaptability of the correction.

[0109] In some of the schemes described above in this application, during the process of generating an initial precipitation distribution field by dynamically fusing multi-model precipitation forecast data, due to the differences in forecast performance of each numerical model within different precipitation intensity ranges, it is difficult to accurately reflect the differences in forecast reliability of the models under extreme precipitation levels when using fixed weight coefficients for fusion, resulting in insufficient ability of the fusion results to capture heavy precipitation events.

[0110] This application further proposes setting several precipitation level thresholds, converting historical forecasts and corresponding real-time monitoring data of each model into hit, missed, and false alarm events at each level; based on the number of hits, missed alarms, and false alarm events, calculating the threat score and bias score of each model at each precipitation level, where the threat score is the ratio of the number of hit times to the total number of hit, missed, and false alarm events; and the bias score is the ratio of the total number of hits and false alarms to the total number of hits and missed alarms. The threat score and bias score are normalized, and weighting coefficients are set for different precipitation levels. The weighting coefficient of the threat score is greater than that of the bias score at the heavy precipitation level, and the weighting coefficient of the threat score is less than that of the bias score at the light precipitation level, thereby generating a comprehensive weighting coefficient for each model at each precipitation level.

[0111] The precipitation level thresholds are set according to meteorological industry standards, with nine gradient intervals ranging from 0.1 mm to 50 mm, each gradient corresponding to an independent event discrimination criterion. During the calculation of threat and bias scores, a binary classification discrimination matrix is ​​established for each precipitation level, and cross-validation results between statistical model forecasts and actual observations are used. Normalization employs a linear scaling method, converting the threat scores of each model at each level into a weighted distribution with a sum of 1.

[0112] Specifically, in the dynamic fusion process of heavy rain warnings, nine precipitation level thresholds (0.1 mm, 1 mm, 3 mm, 5 mm, 10 mm, 15 mm, 20 mm, 30 mm, and 50 mm) are first set according to meteorological industry standards. For each precipitation level, historical forecast data from the model is compared grid-by-grid with corresponding observed data. When both the forecast and observed values ​​exceed the current level threshold, it is recorded as a hit event; if only the forecast value exceeds the threshold but the observed value does not, it is recorded as a false alarm event; and if only the observed value exceeds the threshold but the forecast does not, it is recorded as a missed event. By statistically analyzing the frequency of events at each precipitation level for each model within a 30-day training period, the threat score TS is calculated as follows: TS = (Number of hits + Number of missed alarms + Number of false alarms), and the bias score is calculated as follows: TS = (Number of hits + Number of false alarms) / (Number of hits + Number of missed alarms). Subsequently, a range normalization method is used to map the threat score and bias score of each model at each precipitation level to the 0-1 interval.

[0113] As a preferred embodiment, the solution of this application is specifically implemented as follows:

[0114] Precipitation level thresholds were set at 0.1mm, 1mm, 3mm, 5mm, 10mm, 15mm, 20mm, 30mm, and 50mm. Historical forecasts and actual monitoring data for each model were compared, and the number of hits, missed reports, and false alarms was counted at each threshold level. A threat score was calculated for each model at each precipitation level, where the threat score equals the number of hits divided by the total number of hits, missed reports, and false alarms. The bias score was the ratio of the total number of hits and false alarms to the total number of hits and missed reports. The threat score and bias score were normalized to generate weighting coefficients for each model at each precipitation level.

[0115] For example, for a certain precipitation level, Model A has 80 hits, 20 missed reports, and 15 false alarms. Therefore, at this level, Model A's threat score is 80 / (80+15+20) = 0.696, and its bias score is (80+15) / (80+20) = 0.95. Weighting coefficients are then assigned to both scores for different precipitation levels. For heavy precipitation, the weighting coefficient for the threat score is greater than that for the bias score, while for light precipitation, the weighting coefficient is less. For example, if the precipitation level is heavy, the weighting coefficient for the threat score can be 0.7, and the weighting coefficient for the bias score can be 0.3, resulting in a comprehensive weighting coefficient of 0.696*0.7 + 0.95*0.3 = 0.77. This process is repeated to calculate the weighting coefficients for each model at all precipitation levels.

[0116] Specifically, the weighting coefficient combination for each precipitation level can be determined based on historical assessment data. For example, under the heavy precipitation level (e.g., 24-hour precipitation ≥ 50 mm), the threat score weighting coefficient is set to 0.7, and the deviation score weighting coefficient is set to 0.3; under the light precipitation level (e.g., 24-hour precipitation < 25 mm), the threat score weighting coefficient is set to 0.4, and the deviation score weighting coefficient is set to 0.6. For the intermediate level (e.g., 25 mm ≤ precipitation < 50 mm), a balance coefficient of 0.5 can be used.

[0117] Through the above technical solutions, this application can objectively evaluate the forecasting capabilities of each model under different precipitation intensities based on historical forecast performance, providing a reasonable weighting basis for subsequent fusion. By introducing multiple precipitation level thresholds, the forecasting capabilities of each model for precipitation of different intensities can be evaluated in detail, improving the sensitivity of the fusion results to heavy precipitation. Simultaneously, using threat scores and bias scores as evaluation indicators can comprehensively consider the model's hit, miss, and false alarm situations, fully reflecting the forecasting capabilities. The dynamic update mechanism of the weighting coefficients also enables the fusion results to adaptively adjust as the performance of each model changes, maintaining the real-time nature and accuracy of the fusion results.

[0118] In some of the schemes mentioned above in this application, threat scores and bias scores of each model under different precipitation levels are calculated by setting precipitation classification thresholds and generating comprehensive weight coefficients. However, in the actual fusion process, the forecast error distribution of different precipitation levels is different, and a single weight allocation is difficult to adapt to the dynamic changes of multi-level precipitation scenarios, resulting in insufficient spatial consistency and intensity accuracy of the initial precipitation distribution field of the fusion.

[0119] This application further proposes a method for frequency-corrected precipitation forecasts for each model at the current lead time. Based on the precipitation level of the current precipitation value, the model weights corresponding to the current precipitation level are selected from the weight coefficients. The frequency-corrected precipitation value of each model is multiplied by its corresponding weight coefficient, and the fused precipitation value of the current grid point is calculated by weighted sum of all models. Weighted fusion is performed on all grid points to obtain a fused initial precipitation distribution field containing spatial distribution information.

[0120] The weighting coefficients corresponding to precipitation levels are obtained by normalizing threat scores from historical samples, and these coefficients are dynamically adjusted according to the precipitation intensity level. During grid fusion calculations, each grid point independently undergoes weight matching and weighted summation to ensure the preservation of spatial distribution characteristics. The product of the weighting coefficients and precipitation values ​​is calculated using floating-point multiplication, and the weighted sum is calculated using an arithmetic mean or weighted average algorithm. Grid processing covers the entire spatial grid, including land, water, and areas with complex topography, maintaining the same spatial resolution as the input forecast data.

[0121] Specifically, in the frequency-corrected precipitation forecast product, the precipitation value at each grid point is first matched with a preset precipitation level threshold range based on its intensity to determine its level category. The weight values ​​corresponding to each model under that level are extracted from a pre-calculated weight coefficient table. The corrected precipitation value of each model at that grid point is multiplied by its respective weight and then summed to obtain the merged precipitation value. This process is executed in parallel on all spatial grid points, ultimately forming the initial merged precipitation distribution field. Through dynamic weight allocation, the merged results for areas of heavy precipitation rely more on models with high threat scores at that level, while areas of weak precipitation combine the advantages of multiple models, improving the rationality of spatial distribution. Grid-level independent computation avoids spatial smoothing effects, preserves local precipitation characteristics, and provides accurate input for subsequent spatial probability matching and reconstruction.

[0122] As a preferred embodiment, the solution of this application is specifically implemented as follows:

[0123] For each model's frequency-corrected precipitation forecast at the current lead time, the model weights are selected from the weighting coefficients based on the precipitation level of the current value. For example, for a certain grid point, the frequency-corrected precipitation forecast value of the global model is 25 mm / h, and the frequency-corrected precipitation forecast value of the mesoscale model is 30 mm / h. According to the pre-set precipitation level threshold, both values ​​fall within the 20-30 mm / h range. Therefore, the corresponding weights for the global model and the mesoscale model at this level are selected from the weighting coefficient table, assumed to be 0.4 and 0.6, respectively.

[0124] The frequency-corrected precipitation value for each model is multiplied by its corresponding weighting coefficient, and the merged precipitation value for the current grid point is calculated as a weighted sum of all models. Continuing with the example above, the merged precipitation value for this grid point is calculated as: 250.4 + 300.6 = 28 mm / h

[0125] A weighted fusion is performed on all grid points to obtain a fused initial precipitation distribution field containing spatial distribution information. Specifically, the above steps are repeated for each grid point within the forecast area to finally obtain a two-dimensional precipitation intensity field covering the entire forecast area, i.e., the fused initial precipitation distribution field.

[0126] Through the above technical solution, this application achieves dynamic fusion of forecast results from multiple numerical models. By introducing weighting coefficients based on historical performance, the forecasting advantages of different models under different precipitation levels are fully utilized, improving the accuracy of the fusion results. Simultaneously, the grid-scale weighted averaging method preserves the spatial distribution characteristics of the original forecast field, which is beneficial for subsequent spatial matching and dynamic correction processes. This fusion method considers both the historical performance of the models and the spatial characteristics of the current forecast, thereby improving the spatiotemporal accuracy of heavy rain warnings.

[0127] In some of the schemes described above in this application, the initial precipitation distribution field is generated by weighted averaging of multiple models. However, the spatial distribution structure of precipitation forecast fields from different models differs, which may lead to inconsistencies in the spatial morphology of the fusion results and affect the accuracy of the rainstorm potential field.

[0128] This application further proposes a scheme to sort the precipitation grid intensity of the fused initial precipitation distribution field and to reconstruct the precipitation forecast field after frequency correction of each model based on the sorting.

[0129] The sorting operation establishes a unified spatial intensity reference benchmark, the precipitation intensity sequence generation process preserves the precipitation distribution characteristics of each model, and the grid reassignment operation achieves spatial structure adaptation of precipitation fields from different models. In specific implementation, the grid sorting adopts a descending order to ensure priority matching of strong precipitation centers, the statistical distribution characteristics of the original precipitation values ​​of each model are preserved when generating the precipitation intensity sequence, and the spatial reconstruction process eliminates spatial offset errors between models through grid position mapping.

[0130] Specifically, firstly, all grid points in the initial fusion field are arranged from highest to lowest precipitation intensity, forming a standardized spatial intensity sequence. Then, the precipitation forecast fields of each model after frequency correction are independently sorted in descending order, generating their respective precipitation intensity sequences. Finally, the precipitation intensity sequences of each model are redistributed to their corresponding geographic coordinates according to the grid point ranking position in the initial fusion field. For example, the 5th highest precipitation value in a model's precipitation intensity sequence is assigned to the 5th grid point in the initial fusion field. This process ensures that the spatial distribution of precipitation fields from each model remains consistent with the initial fusion field, while preserving the statistical characteristics of precipitation intensity from each model, effectively eliminating spatial structural differences during multi-source data fusion.

[0131] As a preferred embodiment, the solution of this application is specifically implemented as follows:

[0132] All grid points in the merged initial precipitation distribution field are sorted from highest to lowest precipitation intensity, and their sorting positions are recorded. Specifically, the merged initial precipitation distribution field is first represented as a two-dimensional matrix, where each element represents the precipitation intensity value of a grid point. Then, this matrix is ​​converted into a one-dimensional array, and the array is sorted in descending order using the quicksort algorithm. Simultaneously, an index matrix of the same size as the original matrix is ​​created to record the position of each grid point after sorting.

[0133] Furthermore, the precipitation values ​​in each model's frequency-corrected precipitation forecast field are sorted from highest to lowest intensity to generate a precipitation intensity sequence. For example, the same sorting operation is performed on the frequency-corrected precipitation forecast fields of the global model and the mesoscale model, resulting in two precipitation intensity sequences.

[0134] Therefore, the precipitation intensity sequence is reassigned to the corresponding grid points according to the grid sorting position of the initial fusion field, forming a reconstructed precipitation distribution field with a unified spatial structure. In practice, the index matrix of the initial fusion field is traversed, and based on the sorting position of each grid point, the precipitation value at the corresponding position in the precipitation intensity sequence of each model is selected and assigned to the corresponding grid point in the reconstructed precipitation distribution field. In this way, the reconstructed precipitation distribution field maintains the spatial structure characteristics of the initial fusion field while incorporating precipitation intensity information from various models.

[0135] Through the above technical solution, this application achieves spatial probability matching reconstruction of multi-source forecast products. The reconstructed precipitation distribution field maintains the spatial structure characteristics of the fused initial field, while incorporating precipitation intensity information from various models, thus improving the spatial consistency and accuracy of precipitation forecasts. This method effectively solves the problem of inconsistency in the spatial distribution of multi-model forecast results, providing a more reliable precipitation distribution field for subsequent heavy rain warnings.

[0136] In some of the schemes mentioned above in this application, although the spatial structure of the reconstructed precipitation distribution field is optimized through multi-model fusion and probability matching, there are still spatial deviations between the reconstructed precipitation distribution field and the real-time monitoring data in practical applications. For example, the center position of heavy precipitation may shift or the intensity may differ, resulting in insufficient accuracy of the potential field and affecting the accuracy of the early warning.

[0137] This application further proposes a method to dynamically correct the reconstructed precipitation distribution field in time and space based on the spatial deviation matching degree between the actual monitoring data and the reconstructed precipitation distribution field, thereby generating a real-time rainstorm potential field. This includes: actual monitoring data such as automatic weather station observation data, radar echo images, satellite cloud images, and wind profiler radar; calculating the spatial structure deviation between the actual monitoring data and the reconstructed precipitation distribution field, including spatial offset distance, structural similarity, and differences in the location of heavy precipitation centers; constructing a spatial correction weight field based on the spatial structure deviation, and performing grid-by-grid dynamic correction on the reconstructed precipitation distribution field; and combining the radar strong echo area, wind uplift motion area, and high instability energy region to obtain the real-time rainstorm potential field.

[0138] The spatial offset distance is obtained by calculating the difference between the coordinates of the heavy precipitation center in the reconstructed field and the actual field. Structural similarity is measured using a structural similarity index to assess the consistency of the spatial morphology between the two fields. The difference in the location of the heavy precipitation center is evaluated using the grid intensity peak matching degree. The spatial correction weight field assigns correction coefficients according to the degree of deviation; the larger the deviation, the lower the weight of the region. The correction coefficients are mapped to the spatial field using a Gaussian kernel function. Spatiotemporal dynamic correction uses a sliding time window to analyze the evolution trend of the deviation and dynamically adjusts the correction intensity. Strong radar echo areas are identified using reflectivity thresholds, upward wind motion areas are calculated from the vertical velocity of the wind profile, and high unstable energy regions are delineated based on the convective effective potential energy index.

[0139] Specifically, automatic weather station observations provide real-time ground precipitation data, radar echo images reflect the real-time location of precipitation clouds, satellite cloud images monitor large-scale cloud system evolution, and wind profiler radar captures vertical wind field changes. When calculating spatial offset distance, the coordinates of the top 5% of heavy precipitation grid points in the reconstructed field and the actual field are extracted to calculate the centroid offset vector. The structural similarity index is calculated using the mean, variance, and covariance within a local window, with the window size set to 10km × 10km. The difference in the location of the heavy precipitation center is measured using the Hausdorff distance to determine the spatial matching degree of the peak point sets of the two fields. When constructing the corrected weight field, the deviation parameter is normalized to the 0-1 interval, and the weight coefficient has a negative exponential relationship with the deviation value. During the spatiotemporal dynamic correction process, Kalman filtering is used to recursively update the correction coefficients, and a deviation reanalysis is performed every 15 minutes. Strong radar echo areas are defined as regions with reflectivity ≥ 40dBZ, wind uplift motion areas are selected from regions with vertical velocity ≤ -1m / s, and high unstable energy areas are identified from regions with CAPE ≥ 1000J / kg. By fusing multi-source data, the corrected potential field reduced the positioning error of the center of heavy precipitation to within 3km, improved the structural similarity to over 0.85, and increased the warning hit rate by 12%.

[0140] As a preferred embodiment, the specific implementation of this application's solution is as follows: The generation process of the real-time rainstorm potential field includes three core steps. First, spatial matching is performed between minute-level precipitation observation data acquired by automatic weather stations and the reconstructed precipitation distribution field. The spatial morphological differences between the two at a 10km grid resolution are calculated using a structural similarity index, while the Hausdorff distance algorithm is used to identify the positional offset of the heavy precipitation center. Second, a two-dimensional Gaussian corrected weight field is constructed based on the spatial structural deviation, where the weight coefficients are calculated using an exponential decay function of the deviation distance, dynamically adjusting the grid values ​​of the reconstructed precipitation distribution field. Finally, areas with radar reflectivity factors greater than 40dBZ are marked as strong echo areas. Combined with vortex structure regions in the vertical wind field with upward velocities exceeding 2m / s, and superimposed with high-energy regions with atmospheric instability energy exceeding 1500J / kg, the core influence area of ​​the rainstorm potential field is determined through logical AND operations.

[0141] Through the above technical solutions, this application effectively solves the problem of delayed response of traditional rainstorm warning systems to sudden heavy rainfall events. By introducing a dynamic deviation correction mechanism between multi-dimensional real-time monitoring data and forecast fields, the false alarm rate caused by spatial location deviations is reduced; by combining radar strong echo characteristics and atmospheric dynamic parameters for comprehensive discrimination, accurate identification of rainstorm triggering conditions is achieved; and based on the continuous updating capability of the spatiotemporal dynamic weight field, the tracking accuracy of the potential field during the development and evolution of rainstorms is ensured, providing reliable technical support for the timely release of warning information.

[0142] In some of the schemes described above in this application, there is a spatial structure deviation between the reconstructed precipitation distribution field and the actual monitoring data, resulting in insufficient accuracy of the real-time rainstorm potential field and an inability to effectively reflect the dynamic evolution characteristics of the actual heavy precipitation area.

[0143] This application further proposes to dynamically correct the reconstructed precipitation distribution field in time and space based on the spatial deviation matching degree between the actual monitoring data and the reconstructed precipitation distribution field, thereby generating a real-time rainstorm potential field.

[0144] The calculation of spatial structure deviation includes spatial offset distance, structural similarity, and the difference in the location of the heavy precipitation center. Spatial offset distance is calculated using cross-correlation analysis to determine the displacement corresponding to the maximum correlation coefficient between the reconstructed field and the actual field. Structural similarity is measured using a structural similarity index to assess the difference in spatial texture between the two. The difference in the location of the heavy precipitation center is calculated by identifying the centroid coordinates of regions in the reconstructed and actual fields where precipitation intensity exceeds a preset threshold, and then calculating the Euclidean distance between the centroids. The spatial correction weight field constructed based on these deviations is generated using a Gaussian kernel function, and the weight coefficients are negatively correlated with the deviation values. Spatiotemporal dynamic correction is achieved by multiplying the grid values ​​of the reconstructed precipitation distribution field point by point with the correction weight field. Strong radar echo regions are identified using areas with a reflectivity factor exceeding 40 dBZ; wind uplift regions are defined by areas with a vertical velocity exceeding 0.5 m / s; and high unstable energy regions are determined by calculating the convective effective potential energy exceeding 1000 J / kg.

[0145] Specifically, after the reconstructed precipitation distribution field is generated, real-time automatic weather station observation data is used to generate a real-world precipitation field with a spatial resolution of 1 km through Kriging interpolation. After spatial registration of the reconstructed field and the real-world field, the spatial offset distance within each 5 km × 5 km window is calculated using the sliding window method. For example, the offset vector is determined by calculating the maximum cross-correlation coefficient between the two fields in the x and y directions. For structural similarity, a multi-scale structural similarity index is used, calculating the brightness, contrast, and structural similarity components of the two fields at three spatial scales, with weighting coefficients set to 0.6, 0.2, and 0.2, respectively. When calculating the difference in the location of the heavy precipitation center, grid areas with precipitation intensity exceeding 50 mm / 6 h are selected, and the longitude and latitude differences of their centroid coordinates are calculated. In the corrected weighted field constructed based on the above deviation indices, the weighting coefficient decreases exponentially with increasing offset distance, with an attenuation coefficient set to 0.1 / km. The corrected precipitation field is spatially superimposed with the radar strong echo area. When the overlapping area exceeds 30%, the precipitation intensity in that area is increased by 10%. The final real-time rainstorm potential field is formed by integrating the spatial intersection of the modified precipitation field, the wind field uplift motion region, and the high instability energy region, thus creating a rainstorm potential distribution with three-dimensional structural characteristics.

[0146] As a preferred embodiment, the solution of this application is implemented as follows: Rainstorm warning thresholds are divided into three levels: yellow, orange, and red, corresponding to critical values ​​of 50 mm, 100 mm, and 150 mm of cumulative 24-hour precipitation, respectively. Grid values ​​of the real-time rainstorm potential field are overlaid with an administrative division layer using a geographic information system. When the potential value of a grid point exceeds the red threshold, the administrative region to which that grid point belongs is automatically marked as a red warning area. The warning information generation module determines the trigger time as the forecast time when the potential value first exceeds the threshold, and the expected duration is calculated based on the evolution trend of the potential field over the next 6 hours. The generated warning information is stored in vector geographic data format, including a warning level code, administrative region code, trigger timestamp, and duration field, and is pushed to the emergency broadcasting system through a standard protocol interface.

[0147] Through the above technical solution, this application achieves multi-level automatic identification and dynamic dissemination of rainstorm warning information, solving the problem of the disconnect between warning level and precipitation intensity in traditional methods. By comparing the real-time potential field with preset thresholds grid-by-grid, the spatial distribution characteristics of rainstorms of different intensities can be accurately identified, avoiding response delays caused by manual judgment. The warning information incorporates trigger time and duration parameters, enabling emergency response measures to be dynamically adjusted according to the precipitation development trend, effectively improving the matching degree between warning instructions and disaster prevention needs.

[0148] In some of the schemes mentioned above in this application, the rainstorm warning information lacks a dynamic update mechanism after it is generated, which makes it impossible for the warning information to reflect the development trend of short-term heavy precipitation in a timely manner. The warning level may be delayed or the regional coverage may be biased due to changes in the potential field, which affects the timeliness and accuracy of the warning.

[0149] This application further proposes that after generating rainstorm warning information, several rolling update time periods are set, and the real-time rainstorm potential field is recalculated and compared in each rolling update time period to generate rolling updated multi-time-period rainstorm warning information. The warning level or warning area is dynamically adjusted according to the development trend of heavy precipitation, the rate of change of potential value and the degree of proximity of threshold corresponding to each time period.

[0150] The rolling update timeframe includes 1-hour, 3-hour, and 6-hour intervals, continuously optimizing the early warning information through periodic calculations at different time intervals. The recalculation process is based on the latest acquired real-time monitoring data and multi-model forecast products, employing a dynamic fusion algorithm to update the spatial distribution data of the real-time heavy rainfall potential field. The comparison phase matches the updated potential field values ​​with preset multi-level heavy rainfall warning thresholds grid-by-grid to identify spatial locations that meet the triggering conditions. The dynamic adjustment mechanism analyzes the gradient of potential value changes within adjacent timeframes to determine the development trend of heavy precipitation. When the potential value growth rate exceeds a set threshold, an upgrade of the warning level is triggered; when the potential value decays below the current level threshold, a downgrade or cancellation of the warning is triggered. The warning area adjustment uses a clustering algorithm to identify newly added or disappeared heavy precipitation core areas in the potential field, synchronously updating the warning range boundary.

[0151] Specifically, after generating the initial heavy rain warning, the early warning system initiates a periodic update process. Every 1, 3, or 6-hour interval, the system automatically retrieves the latest input radar echo, satellite-retrieved precipitation, and model forecast data, and re-executes data fusion and potential field calculations. For example, in a 1-hour update cycle, the system inputs real-time precipitation data collected within the last 60 minutes and model forecast products into the dynamic fusion module to update the precipitation intensity probability distribution at each grid point in the potential field. After comparing the updated potential field with the warning threshold, if the potential value of a certain area increases by more than 50% in two consecutive updates within 3 hours, the warning level for that area is upgraded from yellow to orange. Simultaneously, when a new core area of ​​heavy precipitation potential is detected at the edge of the original warning area using a spatial clustering algorithm, the warning area coverage is automatically expanded. This dynamic adjustment mechanism effectively solves the problem of insufficient timeliness caused by the rapid evolution of meteorological conditions in traditional static early warning information, ensuring the synchronization of early warning information with real-time developments through multiple rolling updates.

[0152] As a preferred embodiment, the specific implementation of this application is as follows: After the heavy rain warning information is generated, three rolling update time periods are set, namely 1 hour, 3 hours, and 6 hours. Based on real-time updated multi-source forecast data and actual monitoring data, the real-time heavy rain potential field is periodically recalculated. Within each 3-hour update period, the initial precipitation distribution field is reconstructed by dynamically fusing the latest model forecast products, and spatial probability matching reconstruction and spatiotemporal dynamic correction are performed to generate an updated real-time heavy rain potential field. The updated potential field is compared with the potential field of the same period in history to extract the heavy precipitation development trend parameters, including the potential value growth rate and spatial expansion speed. At the same time, the rate of change of the potential value in adjacent update periods is calculated, and its closeness to the preset warning threshold is analyzed. If the potential value growth rate of a certain area exceeds the preset threshold in two consecutive update periods, and the deviation from the orange warning threshold is less than the critical value, the warning level of the area is automatically upgraded from yellow to orange. Based on the 6-hour update cycle, and combining radar extrapolation data with the continuity of model forecasts, the warning area is dynamically adjusted. For example, the warning level of areas that have passed the peak of heavy precipitation is lowered, while newly developed heavy precipitation centers are included in the warning range.

[0153] Through the above technical solutions, this application enables the rolling update and dynamic optimization of rainstorm warning information, solving the problem of insufficient timeliness in traditional warning mechanisms. By periodically recalculating the potential field and correlating it with multiple time-dependent development trends, it avoids misjudgments caused by lag in forecast data; based on the dual judgment of the rate of change and the degree of proximity to the threshold, it effectively suppresses excessive fluctuations in warning levels; and through a multi-period collaborative update mechanism, it ensures that both short-term sudden heavy rainfall and continuous rainstorm processes can obtain matching warning responses, improving the accuracy of warning information in the spatiotemporal dimensions.

[0154] The above embodiments construct a multi-source precipitation forecasting system integrating global and mesoscale models, build a dynamically updatable cumulative precipitation distribution function based on historical data, and introduce a frequency correction mechanism based on quantile linear interpolation to accurately correct precipitation forecasts from different models. A weighted fusion method driven by precipitation level threat scoring is introduced, and spatial probability matching and spatiotemporal dynamic correction further enhance the ability of the reconstructed precipitation field to restore the spatial structure of severe convective weather. Unlike traditional technologies that only output static warning information, this method constructs a multi-level rainstorm warning threshold system and automatically compares it with the real-time rainstorm potential field to generate warning signals with clear triggering conditions and distinct levels. Finally, multi-channel alarm linkage is achieved through audio-visual equipment, mobile terminals, or emergency broadcasts, improving the timeliness, accuracy, and practicality of rainstorm warning responses, effectively protecting public safety, and enhancing urban emergency response capabilities.

[0155] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program goods. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program goods 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.

[0156] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program goods according to embodiments of this application. 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... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0157] 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.

[0158] 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.

[0159] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A method for issuing rainstorm warnings based on dynamic fusion of multi-source forecast products, characterized in that, include: Global model precipitation forecast data and mesoscale model precipitation forecast data are collected. The cumulative precipitation distribution function between historical precipitation forecast data and corresponding real-time monitoring data of each model is constructed, and the cumulative precipitation distribution function is dynamically updated based on Kalman filtering. Based on the updated precipitation cumulative distribution function, quantile linear interpolation is used to correct the frequency of the current lead time precipitation forecast values ​​of each model, and the frequency-corrected precipitation forecast products of each model are obtained. Historical precipitation forecast data and actual monitoring data samples from 30 days prior to the target forecast date and 30 days after the same period last year are obtained. Threat scores and bias scores of each model at different precipitation levels are calculated according to the set precipitation classification thresholds. Comprehensive weight coefficients corresponding to each precipitation level are generated. Based on the comprehensive weight coefficients, the frequency correction precipitation forecast products of each model are weighted and fused to obtain the initial fused precipitation distribution field. Using the precipitation grid intensity ranking of the fused initial precipitation distribution field as a reference, spatial probability matching is performed on the frequency correction precipitation forecast products of each model according to the ranking to generate a reconstructed precipitation distribution field. Collect the latest real-time monitoring data, calculate the spatial structure deviation of precipitation between the latest real-time monitoring data and the reconstructed precipitation distribution field, perform spatiotemporal dynamic correction on the reconstructed precipitation distribution field, and generate a real-time rainstorm potential field; A multi-level rainstorm warning threshold is set, the real-time rainstorm potential field is compared with the multi-level rainstorm warning threshold, rainstorm warning information is generated, and the rainstorm warning information is released to users through audio-visual equipment, mobile terminals or emergency broadcasts.

2. The rainstorm early warning method based on dynamic fusion of multi-source forecast products according to claim 1, characterized in that, When constructing the cumulative precipitation distribution function between historical precipitation forecast data and corresponding real-time monitoring data for each model, and dynamically updating the cumulative precipitation distribution function based on Kalman filtering, the process includes: The system collects global model precipitation forecast data and mesoscale model precipitation forecast data within a historical window set before the target forecast date, as well as real-time monitoring data for the corresponding time period. Spatial registration and gridding of historical precipitation forecast data and real-time monitoring data are performed to establish grid-point-by-grid pairs of historical forecast and real-time monitoring data. The historical precipitation forecast values ​​and actual precipitation values ​​for each model are sorted in ascending order to construct a cumulative precipitation distribution function; During the construction process, the cumulative precipitation distribution function is updated recursively using Kalman filtering based on the latest real-time monitoring data and precipitation forecast values.

3. The rainstorm early warning method based on dynamic fusion of multi-source forecast products according to claim 2, characterized in that, When recursively updating the cumulative precipitation distribution function based on Kalman filtering, the following steps are included: The cumulative precipitation distribution function of each model is expressed as precipitation intensity values ​​corresponding to multiple precipitation locations; The precipitation intensity value is used as the state variable for Kalman filtering; Based on the deviation between the newly added precipitation forecast value and the actual precipitation value, the state variable is recursively corrected through the prediction and update steps of Kalman filtering.

4. The rainstorm early warning method based on dynamic fusion of multi-source forecast products according to claim 1, characterized in that, When performing frequency correction based on the updated cumulative precipitation distribution function, the following is included: Calculate the current time-leading precipitation forecast value for each precipitation model, and calculate the cumulative probability value in the cumulative precipitation distribution function corresponding to that model; The cumulative probability value is mapped to the corresponding actual precipitation cumulative distribution function, and the equivalent precipitation value in the actual space is determined by quantile linear interpolation. The equivalent precipitation value is used as the result after frequency correction to generate the frequency-corrected precipitation forecast products for each model.

5. The rainstorm early warning method based on dynamic fusion of multi-source forecast products according to claim 1, characterized in that, When calculating the threat score and bias score of each model at different precipitation levels according to the set precipitation classification thresholds, and generating the comprehensive weight coefficient corresponding to each precipitation level, the following are included: Several precipitation level thresholds are set, and the historical precipitation forecast data and corresponding real-time monitoring data of each model are converted into hit, miss and false alarm events at each level. Based on the number of hits, the number of missed reports, and the number of false alarms, the threat score and bias score for each model at each precipitation level are calculated. The threat score is the ratio of the number of hit times to the total number of hit, missed, and false alarm events, and the bias score is the ratio of the total number of hits and false alarms to the total number of hits and missed reports. The threat score and bias score are normalized, and weighting coefficients are set for different levels of precipitation. The weighting coefficient of the threat score is greater than that of the bias score when the precipitation level is heavy, and the weighting coefficient of the threat score is less than that of the bias score when the precipitation level is light, thereby generating a comprehensive weighting coefficient for each model at each precipitation level.

6. The rainstorm early warning method based on dynamic fusion of multi-source forecast products according to claim 5, characterized in that, When performing a weighted average of the frequency-corrected precipitation forecast products from each model based on the aforementioned comprehensive weighting coefficients to obtain the fused initial precipitation distribution field, the following steps are included: For each mode of frequency-corrected precipitation forecast product at the current time, the comprehensive weight coefficient for the corresponding level is selected from the comprehensive weight coefficients according to the precipitation level at which the current precipitation value is located. The frequency-corrected precipitation value of each model is multiplied by its corresponding comprehensive weight coefficient, and the fused precipitation value of the current grid point is calculated by weighting all models. A weighted fusion is performed on all grid points to obtain a fused initial precipitation distribution field containing spatial distribution information.

7. The rainstorm early warning method based on dynamic fusion of multi-source forecast products according to claim 1, characterized in that, Using the precipitation grid intensity ranking of the fused initial precipitation distribution field as a reference, spatial probability matching is performed on the frequency-corrected precipitation forecast products of each model according to the ranking to generate the reconstructed precipitation distribution field, including: Sort all grid points of the fused initial precipitation distribution field from high to low precipitation intensity and record the sorting position; The precipitation values ​​of each model frequency-corrected precipitation forecast product are sorted from high to low intensity to generate a precipitation intensity sequence. The precipitation intensity sequence is reassigned to the corresponding grid points according to the grid point sorting position of the initial precipitation distribution field, thus forming a reconstructed precipitation distribution field with a unified spatial structure.

8. The method for rainstorm early warning based on dynamic fusion of multi-source forecast products according to claim 1, characterized in that, When calculating the spatial structure deviation of precipitation between the latest real-time monitoring data and the reconstructed precipitation distribution field, and performing spatiotemporal dynamic correction on the reconstructed precipitation distribution field to generate a real-time rainstorm potential field, the following steps are included: The latest real-time monitoring data includes observation data from automatic weather stations, radar echo images, satellite cloud images, and wind profiler radar. Calculate the spatial structure deviation of precipitation between the latest real-time monitoring data and the reconstructed precipitation distribution field. The spatial structure deviation of precipitation includes spatial offset distance, structural similarity, and difference in the location of heavy precipitation centers. A spatially corrected weight field is constructed based on the aforementioned spatial structure deviation of precipitation, and the reconstructed precipitation distribution field is dynamically corrected in spatiotemporal manner on a grid-by-grid basis. The real-time rainstorm potential field is obtained by combining the radar strong echo area, the wind field uplift motion area, and the high instability energy area.

9. The method for rainstorm early warning based on dynamic fusion of multi-source forecast products according to claim 1, characterized in that, When generating rainstorm warning information by setting multi-level rainstorm warning thresholds and comparing the real-time rainstorm potential field with the multi-level rainstorm warning thresholds, the process includes: The rainstorm warning thresholds are divided into several levels based on the severity of the rainstorm, including blue, yellow, orange, and red levels. The grid values ​​of each point in the real-time rainstorm potential field are compared with each threshold level step by step to identify the spatial location and level category that meet the warning triggering conditions, and to generate rainstorm warning information that includes the warning level, warning area, trigger time and expected duration.

10. The method for rainstorm early warning based on dynamic fusion of multi-source forecast products according to claim 1, characterized in that, After generating a rainstorm warning, it also includes: Several rolling update time periods are set, including 1 hour, 3 hours and 6 hours; The real-time rainstorm potential field is recalculated and compared in each rolling update period to generate rolling updated multi-time-effect rainstorm early warning information. The warning level or warning area will be dynamically adjusted based on the development trend of heavy precipitation, the rate of change of potential value, and the degree of proximity of threshold for each time period.

Images