Extreme rainfall intensity potential intelligent forecasting method based on optimized physical factors

Abstract

The invention discloses an extreme rainfall intensity potential intelligent forecasting method based on optimized physical factors. The method comprises the following steps: preparing numerical mode forecasting data for training an extreme rainfall intensity potential intelligent forecasting model; rainfall observation data used for training of the extreme rainfall intensity potential intelligent forecasting model are prepared; calculating a physical diagnostic quantity based on a numerical pattern; constructing forecasting factors of the extreme rainfall intensity intelligent forecasting model; carrying out sensitivity analysis on the forecasting factors and carrying out preliminary screening; carrying out dimension reduction processing on the forecasting factors; constructing a fuzzy matching observation label of a space-time neighborhood; constructing an optimal extreme rain intensity balance sample training set; performing data cleaning and standardized preprocessing; constructing an extreme rainfall intensity intelligent forecasting model and optimizing key parameters; intelligent forecasting and forecasting effect evaluation are carried out. According to the method, the problems of magnitude underestimation and falling area deviation in the extreme rainfall intensity forecast in the next three days in the existing numerical mode are solved.

Description

Technical Field

[0001] The present invention relates to a forecasting method, and particularly to an intelligent forecasting method for extreme rainfall intensity potential based on optimized physical factors. Background Art

[0002] Extreme rainfall intensity is characterized by local, sudden and strong disaster-causing properties. Its formation cause is very complex and involves the influence of multi-scale factors (weather scale systems, meso-scale and small-scale convective systems, terrain effects, etc.). Under the background of global warming, extreme rainfall intensity events occur frequently and the disaster impacts are becoming increasingly serious.

[0003] At present, numerical weather prediction models (hereinafter referred to as numerical models) are the main means for carrying out precipitation forecasting operations. However, due to many limitations of numerical models in aspects such as initial fields, physical processes, and dynamic frameworks, they show obvious deficiencies in forecasting heavy precipitation, especially in forecasting extreme rainfall intensity. Specifically, it is manifested as a significant underestimation of the extreme value of the extreme rainfall intensity center and a serious deviation in forecasting the falling area of extreme rainfall intensity. Especially for extreme rainfall intensity exceeding 24 hours, there is almost no forecasting skill.

[0004] Moreover, traditional bias correction methods, such as the "frequency matching" method, eliminate the systematic bias of numerical model precipitation forecasts by correcting the frequency of numerical model precipitation forecasts to the observed precipitation frequency. Its defect is that it can only correct the forecasting bias of medium and small precipitation levels, and cannot correct the forecasting bias of heavy precipitation, especially extreme precipitation. In addition, although the emerging artificial intelligence technology in recent years has made progress in non-linear bias correction of numerical model precipitation forecasts, the learning effect for small sample events (such as extreme precipitation) is still not good. Summary of the Invention

[0005] In order to solve the deficiencies existing in the above technologies, the present invention provides an intelligent forecasting method for extreme rainfall intensity potential based on optimized physical factors.

[0006] In order to solve the above technical problems, the technical solution adopted by the present invention is: an intelligent forecasting method for extreme rainfall intensity potential based on optimized physical factors, including the following steps:

[0007] Step 1: Prepare numerical model forecast data for training an extreme rainfall intensity intelligent forecasting model, and the data covers surface elements and upper air elements;

[0008] Step 2: Prepare precipitation observation data for training an extreme rainfall intensity intelligent forecasting model, and the precipitation observation data is hourly grid observation data;

[0009] Step 3: Calculate physical diagnostic quantities based on the numerical model based on the numerical model forecast data collected in Step 1;

[0010] Step 4: Based on the direct output of the numerical model and the physical diagnostic quantities calculated in Step 3, construct the predictors for the extreme rainfall intensity intelligent forecasting model to obtain a dataset of predictors.

[0011] Step 5: Conduct a sensitivity analysis on the predictors and perform a preliminary screening, and select the physical quantities with high sensitivity as the predictors for the extreme rainfall intensity intelligent forecasting model.

[0012] Step 6: Perform dimensionality reduction on the predictors of the extreme rainfall intensity intelligent forecasting model to optimize the screening.

[0013] Step 7: Construct fuzzy matching observation labels for the spatio-temporal neighborhood of precipitation observation data.

[0014] Step 8: Construct the best extreme rainfall intensity balanced sample training set.

[0015] Step 15: Conduct data cleaning and standardization preprocessing on the best extreme rainfall intensity balanced sample training set.

[0016] Step 18: Construct an extreme rainfall intensity intelligent forecasting model and optimize the key parameters to obtain the optimal extreme rainfall intensity intelligent forecasting model.

[0017] Step 21: Apply the optimal extreme rainfall intensity intelligent forecasting model for forecasting, output the potential forecast of extreme rainfall intensity every 3 hours from 0 to 72 hours, and conduct a forecast effect evaluation.

[0018] Preferably, in Step 1, the prepared numerical model forecast data are the output variables of the numerical model with a forecast period of 0 - 72 hours in the recent 5 years, covering surface element data and upper-air element data, including total precipitation, convective precipitation, large-scale precipitation, sea-level pressure, surface pressure, surface 2m air temperature, surface 2m dew point, atmospheric temperature field, geopotential height field, and three-dimensional wind field; the spatial resolution of the data is 0.125°, and the time resolution is every 3 hours.

[0019] Preferably, in Step 2, the precipitation observation data are hourly grid observation data, with a spatial resolution of 1 km and a time resolution of 1 hour.

[0020] Preferably, in Step 3, the calculated physical diagnostic quantities are vorticity, divergence, water vapor flux divergence, and pseudo-equivalent potential temperature at each pressure level, as well as the K-index and A-index describing the stability of the atmospheric stratification.

[0021] Preferably, in Step 4, the process of constructing the predictor dataset is as follows:

[0022] Based on the meteorological elements directly output at each forecast lead time in the numerical model collected in Step 1 and the calculation results of the physical diagnostic quantities calculated in Step 3, a forecast factor dataset for the extreme rainfall intensity intelligent forecast model at each forecast lead time is constructed. Each forecast lead time is 3 hours, 6 hours, 9 hours,..., 72 hours.

[0023] Preferably, in Step 5, the forecast factor sensitivity analysis and preliminary screening process are as follows:

[0024] Based on the hourly precipitation grid observation data collected in Step 2 and the numerical model physical quantity zero-field forecast data collected in Step 3, the probability density distribution of the physical quantities corresponding to the occurrence of extreme rainfall intensity and non-extreme rainfall intensity events is calculated grid by grid; the physical quantities with a mean difference exceeding one standard deviation are used as the forecast factors of the extreme rainfall intensity intelligent forecast model.

[0025] Preferably, in Step 6, the dimensionality reduction processing of the forecast factors creates a set of uncorrelated predictors through principal component analysis, and the influence of retaining different numbers of principal components on the model score is tested to select the final number of principal components to be retained.

[0026] Preferably, in Step 7, the spatio-temporal neighborhood fuzzy matching method for precipitation observation data is as follows:

[0027] First, based on the hourly precipitation grid observation data, a first-level original observation label for hourly extreme rainfall intensity is constructed, that is, it is judged grid by grid and hour by hour whether there is observed precipitation > 50 mm / h. If there is, the original observation label at this hour and this grid point is recorded as 1, and if not, it is recorded as 0.

[0028] Then, a second-level observation label for extreme rainfall intensity is constructed, that is, under the condition that the first-level original observation label is 1, it is judged whether there is observed precipitation > 50 mm / h within 3 hours covered by each lead time according to the forecast lead time of the numerical model. If there is, the second-level observation label is recorded as 1, and if not, it is recorded as 0.

[0029] Finally, a third-level observation label for extreme rainfall intensity is constructed, that is, under the condition that the second-level label is 1, it is judged whether there is observed precipitation > 50 mm / h within 3 hours covered by each lead time at the target grid point and within a range of 0.25° around it. If there is, the third-level observation label is recorded as 1, and if not, it is recorded as 0.

[0030] Preferably, in Step 8, the construction method of the best extreme rainfall intensity balanced sample training set is as follows:

[0031] Using the random undersampling method, samples 1 time, 2 times, and 3 times the number of extreme rainfall intensity samples are randomly drawn from the non-extreme rainfall intensity samples respectively to construct a new non-extreme rainfall intensity sample set. The non-extreme samples are those with the observation label recorded as 0.

[0032] Construct multiple extreme rainfall intensity balanced sample sets by combining extreme rainfall intensity samples and new non - extreme rainfall intensity samples in different ratios. The different ratios include 1:1, 1:2, and 1:3. Then conduct model training experiments to determine the optimal ratio of the number of extreme rainfall intensity samples to non - extreme rainfall intensity samples, and finally construct the optimal extreme rainfall intensity balanced sample set.

[0033] Preferably, in step 11, based on the binary classification idea, by constructing a confusion matrix at a specific threshold of 50 mm / hour, calculate different precipitation forecast evaluation indicators, including the threat score TS, probability of detection POD, false alarm rate FAR, and miss rate MAR. The calculation formulas are shown in equations 7 - 10:

[0034] Threat score TS = number of correct positives / (number of correct positives + number of false alarms + number of misses) (7)

[0035] Probability of detection POD = number of correct positives / (number of correct positives + number of misses) (8)

[0036] False alarm rate FAR = number of false alarms / (number of correct positives + number of false alarms) (9)

[0037] Miss rate MAR = number of misses / (number of correct positives + number of misses) (10).

[0038] The present invention discloses an intelligent forecast method for extreme rainfall intensity potential based on optimized physical factors, which solves the significant deviation problems existing in the extreme rainfall intensity forecast for the next 3 days by existing numerical models, especially the underestimation of precipitation levels and the deviation of the forecast of the rainfall area. By introducing the optimized physical factors of numerical weather prediction models for the forecast site and high - spatiotemporal - resolution precipitation observation data, and combining machine - learning modeling, the effective improvement of the hit rate of the extreme rainfall intensity potential forecast every 3 hours for 0 - 72 hours (the next three days) is realized. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 It is a schematic flow chart of the technical solution of the present invention.

[0040] Figure 2 It is a comparison chart of the precipitation observation from 20:00 to 23:00 on August 5, 2023, the 63 - hour - time - limit forecast of the EC model, and the 63 - hour - time - limit forecast result of the model of the present invention in the embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0041] The present invention will be further described in detail below with reference to the drawings and specific embodiments.

[0042] The present invention utilizes the meteorological element forecast data of the numerical weather prediction model and the high spatio-temporal resolution multi-source fusion grid precipitation observation data, and constructs an intelligent prediction model for extreme rainfall intensity potential based on a machine learning model, realizing the intelligent prediction of the extreme rainfall intensity potential every 3 hours within 0 - 72 hours.

[0043] Among them, extreme rainfall intensity refers to the precipitation intensity where the cumulative precipitation within one hour exceeds 50 mm; extreme rainfall intensity potential refers to the possibility that a certain place will have a one-hour rainfall exceeding 50 mm within a specific future time period.

[0044] As Figure 1 shown, the technical solution process of the present invention is specifically as follows:

[0045] Step 1: Prepare the numerical model forecast data for training the intelligent prediction model of extreme rainfall intensity: The numerical model forecast data covers surface elements and upper-air elements;

[0046] Specifically: Collect the output variables of the numerical model (such as the EC model) with a forecast period of 0 - 72 hours in the past 5 years, covering surface element data and upper-air element data, including total precipitation, convective precipitation, large-scale precipitation, sea-level pressure, surface pressure, surface 2m temperature, surface 2m dew point, atmospheric temperature field, geopotential height field, three-dimensional wind field, etc., with a spatial resolution of 0.125° (about 13 km) and a time resolution of every 3 hours;

[0047] Step 2: Prepare the precipitation observation data for training the intelligent prediction model of extreme rainfall intensity. The precipitation data is hourly grid observation data;

[0048] Collect the hourly precipitation grid observation data (CMPAS) of the multi-source fusion real-time analysis product of the National Meteorological Information Center in the past 5 years, with a spatial resolution of 1 km and a time resolution of 1 hour;

[0049] Step 3: Calculate the physical diagnostic quantities based on the numerical model: Calculate the physical diagnostic quantities based on the numerical model forecast data collected in Step 1;

[0050] Specifically: Based on the surface element and upper-air element data forecasted by the numerical model (such as the EC model) collected in Step 1, calculate the vorticity, divergence, water vapor flux divergence, pseudo-equivalent potential temperature of each pressure layer, and physical diagnostic quantities such as the K index and A index describing the stability of the atmospheric stratification. The physical meanings and calculation formulas of the above physical diagnostic quantities can be specifically seen in Formulas (1 - 6):

[0051] 1) Vorticity is used to characterize the rotation degree of the atmosphere, and is represented by relative vorticity ζ:

[0052]

[0053] where u and v are the zonal and meridional wind speed components respectively; x and y represent the meridional distance and zonal distance of a unit grid respectively, represents the partial differential operation.

[0054] 2) Divergence D, which represents the degree of air current convergence or divergence within a unit volume:

[0055]

[0056] where u and v are the zonal and meridional wind speed components respectively; x and y represent the meridional distance and zonal distance of a unit grid respectively, represents the partial differential operation.

[0057] 3) Moisture flux divergence MFD, which represents the convergence or divergence of moisture transport:

[0058]

[0059] where V is the horizontal wind vector and q is the specific humidity.

[0060] 4) The pseudo-equivalent potential temperature θ se , which represents the potential temperature considering the latent heat effect of water vapor:

[0061]

[0062] where θ e represents the potential temperature, r is the mixing ratio, R d is the dry air gas constant, and c p is the specific heat capacity at constant pressure.

[0063] 5) The K-index, which characterizes the stability of the atmospheric stratification. The larger the K-index, the more unstable the stratification.

[0064] K = (T 850 - T 500 ) + T d500 + (T 700 - T d700 ) (5)

[0065] where T 850 , T 700 and T 500 are the temperatures at the 850 hPa, 700 hPa, and 500 hPa pressure levels respectively, and T d700 and T d500 are the dew point temperatures at the 700 hPa and 500 hPa pressure levels respectively.

[0066] 6) The A-index, which characterizes the stability of the atmospheric stratification. The larger the A-index, the more unstable the stratification.

[0067] A = (T 850+T d850 -2T 500 ) (6)

[0068] where T 850 and T 500 are the temperatures at the 850 hPa and 500 hPa pressure levels respectively, and T d850 is the dew point temperature at the 850 hPa pressure level.

[0069] Step 4: Construct the predictors of the extreme rainfall intensity intelligent forecasting model;

[0070] Based on the surface elements and upper-air elements predicted by the numerical model (such as the EC model) collected in Step 1, and the physical diagnostic quantities (such as vorticity, divergence, water vapor flux divergence, pseudo-equivalent potential temperature, and K-index, A-index, etc.) calculated in Step 3, construct the predictor datasets for each forecasting time period (3 hours, 6 hours, 9 hours,..., 72 hours) of the extreme rainfall intensity intelligent forecasting model.

[0071] Step 5: Sensitivity analysis and preliminary screening of the predictors;

[0072] Based on the hourly precipitation grid observation data (CMPAS) collected in Step 2 and the numerical model zero field (0-hour forecast) collected in Step 3, calculate the physical quantities (such as the meteorological elements directly output by the EC model and vorticity, divergence, water vapor flux divergence, pseudo-equivalent potential temperature, K-index, A-index, etc.) corresponding to the occurrence of extreme rainfall intensity and non-extreme rainfall intensity events at each grid point, and use the physical quantities with a mean difference exceeding one standard deviation as the predictors of the extreme rainfall intensity forecasting model.

[0073] Step 6: Dimension reduction and optimized screening of the model predictors;

[0074] Since there are many model predictors (about 70) and there is a strong correlation between the predictors, which is likely to cause overfitting of the forecasting model, it is necessary to carry out dimension reduction of the model predictors.

[0075] The present invention creates a set of uncorrelated predictors through principal component analysis, and determines the final number of principal components to be retained by testing the influence on the model score when different numbers of principal components are retained (Formulas 7-10).

[0076] Step 7: Construct the fuzzy matching observation labels in the spatio-temporal neighborhood;

[0077] Since the grid observation precipitation data has a higher spatio-temporal resolution, in order to retain as much observation information of extreme rainfall intensity as possible to better match the output results of the numerical model with a lower resolution, construct the fuzzy matching observation labels in the spatio-temporal neighborhood of the precipitation observation.

[0078] First, based on the hourly precipitation grid observation data, construct the first-level original observation labels for extreme rainfall intensity, that is, judge hour by hour and grid by grid whether there is observed precipitation greater than 50 mm / h. If there is, record the original observation label as 1 at this hour and this grid point; if not, record it as 0.

[0079] Then, construct the second-level observation labels for extreme rainfall intensity. Under the condition that the first-level original observation label is 1, according to the forecast lead time of the numerical model, judge whether there is observed precipitation greater than 50 mm / h within the 3 hours covered by each lead time (for example, the observations corresponding to the 6-hour lead time on July 29, 2024 should be the observations at 04:00, 05:00, and 06:00 on July 29, 2024). If there is, record the second-level observation label as 1; if not, record it as 0.

[0080] Finally, construct the third-level observation labels for extreme rainfall intensity. Under the condition that the second-level label is 1, judge whether there is observed precipitation greater than 50 mm / h within the 3 hours covered by each lead time in the target grid point and the surrounding area within 0.25°. If there is, record the third-level observation label as 1; if not, record it as 0.

[0081] Step 8: Construction of the optimal extreme rainfall intensity balanced sample training set;

[0082] Since the sample size of extreme rainfall intensity is extremely scarce (that is, the samples with the observed label recorded as 1 only account for 1% of the total samples), there is a problem of extremely unbalanced extreme rainfall intensity samples and non-extreme rainfall intensity samples in model training, which is not conducive to the construction of the extreme rainfall intensity forecast model.

[0083] Based on steps 5 and 6, use the random undersampling method to randomly extract 1 time, 2 times, and 3 times the number of extreme rainfall intensity samples from the non-extreme rainfall intensity samples (that is, the samples with the observed label recorded as 0) respectively to construct a new non-extreme rainfall intensity sample set.

[0084] Through multiple extreme rainfall intensity balanced sample sets composed of different ratios (1:1, 1:2, 1:3) of extreme rainfall intensity samples and non-extreme rainfall intensity samples, carry out model training experiments to determine the optimal ratio of extreme rainfall intensity samples to non-extreme rainfall intensity samples, and finally construct the optimal extreme rainfall intensity balanced sample set.

[0085] Step 9: Data cleaning and standardization preprocessing;

[0086] Perform data cleaning (including removing missing values, etc.) and normalization preprocessing on the optimal extreme rainfall intensity balanced sample set determined in step 8 to improve data quality and enhance model performance.

[0087] Step 10: Construction of the extreme rainfall intensity intelligent forecast model and optimization of key parameters;

[0088] Using the optimal extreme rainfall intensity balanced sample set constructed in steps 8-9, divide the sample set into a training set, a validation set, and a test set according to a ratio (such as 7:2:1). Establish an intelligent extreme rainfall intensity prediction model based on machine learning methods such as random forest, and use the training set and validation set of the optimal extreme rainfall intensity balanced sample set to carry out model training and validation. Through the grid search method, determine the optimal values of the parameters of the intelligent extreme rainfall intensity prediction model (the number of decision trees, the maximum tree depth, etc.), and use the cross-validation method to verify the simulation effect of the key parameters.

[0089] Step 11: Evaluate the prediction effect of the intelligent extreme rainfall intensity prediction model.

[0090] Use the test set input of the optimal extreme rainfall intensity balanced sample set to input the constructed intelligent extreme rainfall intensity prediction model, and output the extreme rainfall intensity potential prediction every 3 hours from 0 to 72 hours (0-1 prediction, 0 indicates that extreme rainfall intensity does not occur, and 1 indicates that extreme rainfall intensity occurs). According to the model output results, calculate the evaluation indexes of the model prediction effect, as shown in Table 1. According to Formulas 7-10, calculate the false alarm rate (FAR), the miss rate (MAR), the probability of detection (POD), and the threat score (TS) to evaluate the overall prediction performance of the model.

[0091] The specific evaluation indexes of the intelligent extreme rainfall intensity prediction model are as follows:

[0092] Different from the continuous variable of temperature, precipitation is a typical discontinuous variable. Therefore, for the accuracy test of precipitation prediction in meteorology, it is usually based on the binary classification idea. By constructing a confusion matrix under a specific threshold (50 mm / hour), different precipitation prediction evaluation indexes are calculated.

[0093] Table 1 Confusion matrix of extreme rainfall intensity observation and prediction

[0094]

[0095] Threat score (TS) = number of correct positives / (number of correct positives + false alarms + misses) (7)

[0096] False alarm rate (FAR) = number of false alarms / (number of correct positives + number of false alarms) (8)

[0097] Miss rate (MAR) = number of misses / (number of correct positives + number of misses) (9)

[0098] Probability of detection (POD) = number of correct positives / (number of correct positives + number of misses) (10)

[0099] Formulas 7-10 give the specific calculation formulas for the threat score (TS), false alarm rate (FAR) score, miss rate (MAR) score, and probability of detection (POD) score.

[0100] Application example:

[0101] In this example, the extreme rainfall intensity events in the North China and Huanghuai regions from May to September 2023 are taken as the research object. The forecasts of the EC model (a mainstream numerical model) with a forecast period of 0 - 72 hours from May to September 2019 - 2022 (time resolution every 3 hours, spatial resolution 0.125°) and the multi - source fusion grid precipitation observation data of the National Meteorological Information Center (time resolution every 1 hour, spatial resolution 1 km) are used as the input data for the extreme rainfall intensity intelligent forecasting model.

[0102] Through the sensitivity analysis of the forecast factors, 72 forecast factors are selected as the model input, including the model forecast variables such as total precipitation, horizontal wind field, vertical velocity, and physical diagnostic variables such as vorticity and divergence calculated from the above - mentioned model forecast variables (see Table 2). The key parameters of the model, the training window, and the number of principal components retained in the principal component analysis are calibrated. At each grid point, the extreme rainfall intensity intelligent prediction model based on the optimized physical factors (constructed in the present invention) is used to forecast the extreme rainfall intensity potential.

[0103] Table 3 presents the evaluation results of the back - calculation of the extreme rainfall intensity potential forecast with a forecast period of 0 - 72 hours from May to September 2023 by the model of the present invention and its comparison with the EC model forecast during the same period. It can be seen that the hit rate of the EC model's 0 - 72h extreme rainfall intensity forecast is only 0.001, and the TS score is only 0.001; while the model of the present invention has greatly improved the hit rate of extreme rainfall intensity (0.101) compared with the EC model, reduced the false alarm rate, and the overall TS score has been improved by one order of magnitude (0.036) compared with the EC model's forecast result.

[0104] Figure 2 Shown is the comparison of precipitation observations from 20:00 to 23:00 on August 5, 2023 with the forecast results of the model of the present invention and the EC model forecast. Figure 2 (a) shows the grid points (dark purple grid points) with an extreme rainfall intensity of >50 mm / hour during the actual situation from 20:00 to 23:00 on August 5, 2023; (b) shows the 3 - hour precipitation forecast (unit: mm) of the EC model with a 63 - hour forecast period starting from 08:00 on August 3; (c) shows the potential forecast of the extreme rainfall intensity intelligent forecasting model with a 63 - hour forecast period starting from 08:00 on August 3. The value of the dark purple grid point is 1, indicating the potential of extreme rainfall intensity; the value of the blue - green grid point is 0, indicating no potential of extreme rainfall intensity.

[0105] In this example, from 20:00 to 23:00 on August 5, 2023, extreme rainfall intensity occurred in northern Henan and northern Shandong. The 3 - hour precipitation forecast of the EC model with a 63 - hour forecast period was significantly weaker than the actual situation, and the precipitation extreme center was only about 20 mm. Therefore, both the hit rate and the TS score were 0; while the model of the present invention successfully predicted the extreme rainfall intensity belt from northern Henan to northern Shandong, such as Figure 2(c) As shown, the hit rate reaches 0.26 and the TS score reaches 0.11.

[0106] In this example, the intelligent extreme rainfall intensity prediction model is closer to the actual observation than the EC model in both the overall performance from May to September and the prediction results of typical cases.

[0107] Table 2 Prediction factors input into the intelligent extreme rainfall intensity prediction model based on optimized physical factors

[0108]

[0109]

[0110] Table 3 Comparison of overall scores between the EC model and the intelligent extreme rainfall intensity prediction model from May to September 2023

[0111]

[0112] Thus, the intelligent extreme rainfall potential prediction method based on optimized physical factors disclosed by the present invention performs intelligent prediction of extreme rainfall potential based on numerical model output data and high spatiotemporal resolution grid precipitation observation data (such as steps 1-11). This prediction method has the following key points:

[0113] (1) Using sensitivity analysis, optimize physical diagnostic quantities directly related to extreme rainfall intensity as key prediction factors;

[0114] (2) Adopting the random undersampling method to alleviate the problem of extremely unbalanced samples between extreme rainfall intensity and non-extreme rainfall intensity;

[0115] (3) Construct fuzzy matching observation labels based on spatiotemporal neighborhoods to alleviate the problem of insufficient description of the spatial distribution characteristics of extreme rainfall intensity that may be caused by point-to-point modeling in machine learning.

[0116] Compared with the prior art, the present invention has the following technical advantages:

[0117] By innovatively constructing spatiotemporal fuzzy matching observation labels and sample balancing schemes to effectively solve the problem of extreme precipitation data imbalance, and integrating a large number of physical diagnostic quantities directly related to extreme rainfall intensity as prediction factors through sensitivity analysis, the problems of weak prediction of extreme rainfall intensity and large deviation in predicted area by numerical models are improved, and the prediction ability of extreme rainfall intensity for the next 3 days is enhanced.

[0118] The above embodiments are not limitations on the present invention, and the present invention is not limited to the above examples either. Changes, modifications, additions, or substitutions made by those skilled in the art within the scope of the technical solution of the present invention also fall within the protection scope of the present invention.

Claims

1. A method for intelligent forecasting of extreme rainfall intensity potential based on optimal physical factors, characterized by: The following steps are included: Step 1: Prepare numerical model forecast data for training the extreme rainfall intensity intelligent forecast model, which covers ground elements and high-altitude elements; Step 2: Prepare precipitation observation data for training the extreme rainfall intensity intelligent forecast model. The precipitation observation data is hourly grid observation data. Step 3: Calculate the physical diagnostic quantity based on the numerical model based on the numerical model prediction data collected in step 1; Step 4: Based on the direct output of the numerical model collected in step 1 and the physical diagnostic quantity calculated in step 3, the prediction factor of the extreme rainfall intensity intelligent prediction model is constructed to obtain a prediction factor data set; Step 5: Conduct sensitivity analysis on the forecast factors and make preliminary screening, and select highly sensitive physical quantities as forecast factors for the extreme rainfall intensity intelligent forecast model; Step 6: Perform dimensionality reduction processing on the prediction factors of the extreme rainfall intensity intelligent forecast model to optimize the screening; Step 7: construct fuzzy matching observation labels of the spatiotemporal neighborhood for precipitation observation data; Step 8: Construct the optimal extreme rainfall intensity balanced sample training set; Step 9: Perform data cleaning and standardization preprocessing on the optimal extreme rainfall intensity balance sample training set; Step 10: construct an extreme rainfall intensity intelligent forecasting model and optimize key parameters to obtain an optimal extreme rainfall intensity intelligent forecasting model; Step 11: Use the optimal extreme rainfall intensity intelligent forecasting model to make forecasts, output the potential forecast of extreme rainfall intensity every 3 hours from 0 to 72 hours, and evaluate the forecast effect.

2. The method for intelligent forecasting of extreme rainfall intensity potential based on optimal physical factors according to claim 1 is characterized in that: In step 1, the prepared numerical model forecast data are the output variables of the numerical weather forecast model with a forecast time of 0-72 hours in the past five years, covering ground element data and high-altitude element data, including total precipitation, convective precipitation, large-scale precipitation, sea level pressure, surface pressure, ground 2m air temperature, ground 2m dew point, atmospheric temperature field, potential height field, and three-dimensional wind field; the spatial resolution of the data is 0.125°, and the temporal resolution is 3 hours.

3. The method for intelligent forecasting of extreme rainfall intensity potential based on optimal physical factors according to claim 2 is characterized in that: In step 2, the precipitation observation data is hourly grid observation data with a spatial resolution of 1 km and a temporal resolution of 1 hour.

4. The method for intelligent forecasting of extreme rainfall intensity potential based on optimal physical factors according to claim 3 is characterized in that: In step 3, the physical diagnostic quantities calculated are the vorticity, divergence, water vapor flux divergence, pseudo-equivalent potential temperature of each pressure layer, and the K index and A index describing the stability of the atmospheric stratification.

5. The method for intelligent forecasting of extreme rainfall intensity potential based on optimal physical factors according to claim 4 is characterized in that: In step 4, the process of constructing the predictor data set is: Based on the meteorological elements directly outputted at each forecast time of the numerical model collected in step 1 and the calculation results of the physical diagnostic quantities calculated in step 3, a forecast factor data set of the extreme rainfall intensity intelligent forecast model for each forecast time is constructed, and each forecast time is 3 hours, 6 hours, 9 hours, ..., 72 hours.

6. The method for intelligent forecasting of extreme rainfall intensity potential based on optimal physical factors according to claim 5 is characterized in that: In step 5, the sensitivity analysis and preliminary screening process of predictors is as follows: Based on the hourly precipitation grid observation data collected in step 2 and the zero-field forecast data of the numerical model physical quantities collected in step 3, the probability density distribution of the physical quantities corresponding to the occurrence of extreme rainfall intensity and non-extreme rainfall intensity events is calculated grid by grid; the physical quantities with mean differences exceeding one standard deviation are used as forecast factors for the extreme rainfall intensity intelligent forecast model.

7. The method for intelligent forecasting of extreme rainfall intensity potential based on optimal physical factors according to claim 6 is characterized in that: In step 6, the dimensionality reduction of the predictors is performed by creating a set of uncorrelated predictors through principal component analysis. The number of principal components to be retained is finally selected by testing the impact of retaining different numbers of principal components on the model score.

8. The method for intelligent forecasting of extreme rainfall intensity potential based on optimal physical factors according to claim 7 is characterized in that: In step 7, the spatiotemporal neighborhood fuzzy matching method of precipitation observation data is: First, based on the hourly precipitation grid observation data, the first-level original observation label of hourly extreme rainfall intensity is constructed, that is, whether there is observed precipitation >50 mm / h is determined hour by hour and grid by grid. If so, the original observation label is recorded as 1 at the grid point in that hour, otherwise, it is recorded as 0. Then, the secondary observation label of extreme rainfall intensity is constructed. That is, under the condition that the primary original observation label is 1, according to the forecast timeliness of the numerical model, it is judged whether there is >50 mm / h observed precipitation within the 3 hours covered by each timeliness. If so, the secondary observation label is recorded as 1, otherwise, it is recorded as 0. Finally, a third-level observation label for extreme rainfall intensity is constructed. That is, under the condition that the second-level label is 1, it is determined whether there is observed precipitation > 50 mm / h at the target grid point and the surrounding 0.25° range within the 3 hours covered by each time period. If so, the third-level observation label is recorded as 1, otherwise it is recorded as 0.

9. The method for intelligent forecasting of extreme rainfall intensity potential based on optimal physical factors according to claim 8, characterized in that: In step 8, the method for constructing the optimal extreme rainfall intensity balanced sample training set is: Using the random downsampling method, samples with 1, 2, and 3 times the number of extreme rainfall intensity samples are randomly selected from the non-extreme rainfall intensity samples to construct a new non-extreme rainfall intensity sample set, wherein the non-extreme rainfall intensity samples are samples with the observation label 0. The extreme rainfall intensity samples and new non-extreme rainfall intensity samples are divided into multiple extreme rainfall intensity balanced sample sets in different proportions, including 1:1, 1:2, and 1:

3. Model training experiments are carried out to determine the optimal ratio of the number of extreme rainfall intensity samples to non-extreme rainfall intensity samples, and finally the optimal extreme rainfall intensity balanced sample set is constructed.

10. The method for intelligent forecasting of extreme rainfall intensity potential based on optimal physical factors according to claim 9, characterized in that: In step 11, based on the binary classification idea, by constructing a confusion matrix under a specific threshold of 50 mm / h, different precipitation forecast evaluation indicators are calculated, including risk score TS, hit rate POD, false alarm rate FAR, and missed alarm rate MAR. The calculation formula is shown in formula 7-10: Risk score TS = number of positive hits / (number of positive hits + number of false alarms + number of missed alarms) (7) Hit rate POD = number of positive hits / (number of positive hits + number of missed reports) (8) FAR = false alarm number / (positive hit number + false alarm number) (9) Missing alarm rate MAR = number of missed alarms / (number of positive hits + number of missed alarms) (10).

Images