A method, medium and program product for predicting long-range extreme precipitation of typhoons in inland areas based on deep learning and integrating IDII index

Abstract

The present invention discloses a method, medium and program product for predicting long-distance extreme precipitation of typhoon in inland areas based on deep learning and integration of IDII index, constructs a whole-layer dry intrusion index, comprehensively considers the overall effect of dry cold air at different heights; obtains historical TRP precipitation events, typhoons and reanalysis meteorological data, performs preprocessing and integrates IDII index, and constructs a training data set; uses deep learning technology to construct a prediction model including a feature extraction layer, an IDII feature fusion layer and a TRP prediction output layer; uses TRP sample data sets to train, verify and test the model; uses the trained model to predict future TRP events, and outputs prediction results such as TRP precipitation, occurrence time, location and duration. The present invention integrates the dry intrusion physical quantity IDII and deep learning technology, improves the prediction accuracy of TRP events, and provides technical support for the forecast of inland extreme precipitation and disaster prevention and mitigation.

Description

Technical Field

[0001] The present invention belongs to the technical field of machine learning and disastrous weather prediction, and relates to a prediction method for extreme precipitation events under the long-range influence of typhoons. Specifically, it is a prediction method, medium and program product for typhoon remote extreme precipitation (TRP) in inland areas based on a deep learning algorithm and integrating the integrated dry intrusion index (IDII). By introducing the three-dimensional structural characteristics of dry intrusion and big data analysis technology, the refined forecasting capability of typhoon remote precipitation in inland areas is improved. Background Art

[0002] Typhoon is one of the important disastrous weather systems that affects my country and even East Asia. Under the background of global warming, the probability of typhoons rapidly strengthening near my country's coast increases. The position of typhoons at their maximum intensity tends to move northward, and they are getting closer and closer to the coastline, increasing the probability of typhoons interacting with mid-latitude systems. In addition to the direct precipitation caused by the typhoon itself and its spiral rain belt, typhoons also work together with mid-latitude weather systems to cause extreme rainstorms in inland areas hundreds or even thousands of kilometers away from their centers. This precipitation is called typhoon remote precipitation (TRP). The occurrence and development mechanism of TRP is complex, involving the path, intensity and interaction of typhoons with multi-scale weather systems, as well as atmospheric dynamics, thermodynamics and other factors. Therefore, under the complex weather situation of typhoon remote influence, making accurate and precise predictions of extreme rainstorms in inland areas at a fixed time, fixed point and quantitative level has always been a difficult problem in the field of meteorological forecasting.

[0003] Existing TRP prediction methods are mainly based on numerical weather prediction models and statistical methods. Numerical weather prediction models can simulate atmospheric movement and physical processes and provide forecasts of future weather conditions. However, due to the limitations of the model itself and its sensitivity to initial conditions, as well as the interaction between dry and cold air and mid-latitude systems in the long-range effects of typhoons, its parameterization scheme of physical processes is difficult to accurately depict complex atmospheric dynamics, especially in the reflection of phenomena such as dry intrusion. As a result, its ability to predict TRP is still limited, especially in quantitative precipitation forecasts. There are large errors. Statistical methods are based on historical observation data and establish statistical models to predict TRP. However, due to the low probability characteristics of TRP and the complexity of influencing factors, the prediction accuracy of statistical methods is also limited, and traditional methods are difficult to effectively capture nonlinear meteorological processes under the long-range influence of typhoons.

[0004] In the study of typhoon long-distance extreme precipitation events, the dry intrusion phenomenon is considered to be an important factor affecting precipitation intensity. Dry intrusion usually refers to the dry and cold air in the middle and high layers sinking along the isentropic surface to the middle and lower layers of the troposphere, invading the warm and humid air area, forming an obvious cold and warm tropospheric instability, which may induce heavy precipitation. Existing studies have shown that dry intrusion not only affects the manifestation of typhoon long-distance effects, but may even directly trigger extreme precipitation events in inland areas. Previous studies have mostly used indicators such as relative humidity, temperature advection, and potential vorticity to characterize the activity of dry and cold air, but these indicators are often difficult to fully reflect the three-dimensional structural characteristics of dry and cold air activities and their impact on TRP. Especially under the influence of typhoons, the main body of dry and cold air slows down, and its impact on inland areas changes from "fast invasion" to "slow penetration". The existing Dry Intrusion Index (DII) is difficult to accurately describe this feature. The main reason is that the DII index fails to fully integrate the vertical distribution characteristics of dry and cold air. In particular, at different altitudes, the intensity, humidity and potential vorticity characteristics of dry and cold air are significantly different. The existing DII index is mainly based on a single-layer isentropic surface, ignoring the evolution of dry and cold air in the multi-layer atmosphere during its sinking process, and cannot fully reflect its comprehensive impact on local atmospheric dynamics and thermal processes, thus affecting the accuracy of precipitation forecasts. In addition, although the DII index can reflect the meteorological characteristics of typhoons under long-distance influence to a certain extent, existing research mainly focuses on the analysis of typical cases, lacking systematic application under large-scale and big data conditions.

[0005] In summary, the existing TRP prediction methods have the disadvantages of limited prediction accuracy, large quantitative precipitation forecast errors, and difficulty in capturing nonlinear meteorological processes. The existing dry intrusion index is also difficult to fully describe the activity characteristics of dry cold air under the long-range influence of typhoons. Therefore, how to develop a TRP prediction method that can effectively integrate dry intrusion physical quantities and make full use of big data and machine learning technology to improve the prediction accuracy of long-range extreme precipitation of typhoons in inland areas is a technical problem that needs to be solved urgently. Summary of the invention

[0006] 1. Purpose of the invention

[0007] In view of the above-mentioned defects and shortcomings of the prior art, and to solve at least one of the above-mentioned and other technical problems in the prior art, the present invention aims to provide a method, medium and program product for predicting long-distance extreme precipitation in inland areas due to typhoons based on deep learning and integrated with the IDII index. By constructing and introducing a whole-layer dry intrusion index IDII that reflects the three-dimensional structural characteristics of dry and cold air, and combining the nonlinear data mining capabilities of big data and machine learning technology, the nonlinear meteorological process under the long-distance influence of typhoons can be effectively captured, and the structural characteristics of dry and cold air in the vertical direction and its impact on precipitation can be fully considered, thereby enhancing the ability to accurately predict inland extreme precipitation, realizing the timely, quantitative and fixed-point prediction of extreme precipitation in inland areas under the long-distance influence of typhoons, and providing technical support and scientific basis for the forecast, warning and disaster prevention and mitigation of extreme precipitation in inland areas.

[0008] (II) Technical solution

[0009] In order to achieve the purpose of the invention and solve the technical problems, the present invention adopts the following technical solutions:

[0010] The first invention object of the present invention is to provide a method for predicting long-distance extreme precipitation of typhoons in inland areas based on deep learning and integrating the IDII index, which is used to improve the precise forecasting capability of extreme precipitation in inland areas in terms of timing, location and quantity, and provide technical support and scientific basis for forecasting, warning and disaster prevention and mitigation of extreme precipitation in inland areas. The prediction method comprises at least the following steps when implemented:

[0011] SS1. Construction of the whole-layer dry intrusion index IDII

[0012] By integrating the dry cold air movement at different height levels on the isobaric surface and the isentropic surface, the whole layer dry intrusion index is obtained. IDI p and IDI θ , to characterize the vertical penetration strength of dry cold air under the long-range effect of typhoon and its impact on extreme precipitation in inland areas:

[0013]

[0014] in, p is the air pressure, p blev and p tlev denote the bottom and top pressures of the isobaric surface integrals respectively; i is the potential temperature, i blev and i tlev denote the bottom potential temperature and top potential temperature of the isentropic surface integral, respectively; Vp and V θ represent the horizontal wind speed vector on the isobaric surface and the isentropic surface respectively; T is the temperature, and denote the horizontal temperature gradients on the isobaric and isentropic surfaces, respectively; PV p and PV θ represent the potential vorticity on the isobaric surface and isentropic surface respectively; i se is the pseudo-equivalent potential temperature and is used to characterize the temperature at which a moist air mass reaches saturation in an adiabatic process. ( i se ) P and( i se ) θ represent the pseudo equivalent potential temperature on the isobaric surface and isentropic surface respectively;

[0015] SS2. Acquisition of TRP-related multi-source meteorological data

[0016] Obtain multi-source meteorological data related to typhoon long-distance extreme precipitation in inland areas, including at least historical TRP precipitation event data in inland areas, corresponding typhoon data, and reanalysis meteorological data of the same period, where TRP precipitation event data is precipitation event data with 24-hour cumulative precipitation reaching the extreme precipitation threshold obtained based on the objective identification algorithm of typhoon long-distance precipitation;

[0017] SS3. Preprocessing of TRP-related multi-source meteorological data

[0018] The multi-source meteorological data obtained are preprocessed including time-space matching and quality control. Then, based on the preprocessed meteorological data and combined with the definition of the whole layer dry intrusion index, the isobaric layers and isentropic surfaces are calculated. IDI p and IDI θ Data, finally, the preprocessed multi-source meteorological data and the calculated IDII index field data are fused to form multi-source fusion data, and organized into training data sets according to the time series according to the occurrence time of TRP precipitation events;

[0019] SS4. Construction of TRP prediction model based on deep learning and integrating IDII index features

[0020] The TRP prediction model is constructed by using deep learning technology. Its network structure includes at least an input layer, a feature extraction layer, an IDII feature fusion layer, a fully connected layer, and a TRP prediction output layer. The input layer is used to receive multi-source fusion data in the training data set. The feature extraction layer is used to extract the spatiotemporal feature information of the typhoon field, the meteorological element field, and the IDII index field from the input data. The IDII feature fusion layer is used to perform multi-dimensional fusion of the extracted spatiotemporal feature information. The fully connected layer is used to perform nonlinear transformation on the fused feature vector and map it to the TRP precipitation prediction output layer. The TRP prediction output layer is used to predict TRP precipitation events in inland areas in the future.

[0021] SS5. Training, validation and testing of TRP prediction model

[0022] The model training data set constructed in step SS3 is divided into a training set, a validation set, and a test set according to a preset ratio. The multi-source fusion data in the training set is used for model training through a deep learning algorithm. The validation set is used for model parameter adjustment and hyperparameter optimization. The test set is used to evaluate the generalization ability of the trained TRP prediction model under different typhoon, meteorological, and IDII index conditions.

[0023] SS6. Prediction and output of TRP events in inland areas

[0024] Acquire typhoon data and meteorological element field data of the target prediction area in real time, pre-process the acquired data and calculate the current moment IDI p and IDI θ The processed multi-source fusion data is input into the trained TRP prediction model to output the TRP prediction results of the target area in the future.

[0025] The second invention object of the present invention is to provide a computer program product, including computer instructions, for executing the above-mentioned method for predicting long-distance extreme precipitation of typhoons in inland areas based on deep learning and integrating the IDII index.

[0026] The third invention object of the present invention is to provide a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, it implements the above-mentioned inland area typhoon long-distance extreme precipitation prediction method based on deep learning and integration of the IDII index.

[0027] (III) Technical Effect

[0028] Compared with the prior art, the method, medium and program product for predicting long-distance extreme precipitation of typhoons in inland areas based on machine learning and integrating the IDII index of the present invention have the following significant technical effects:

[0029] (1) The present invention makes up for the shortcomings of the existing TRP prediction method by constructing a prediction model based on deep learning and integrating the whole-layer dry intrusion index IDII that can characterize the three-dimensional structural characteristics of dry cold air. It effectively improves the prediction accuracy of long-distance extreme precipitation of typhoons in inland areas, especially in the quantitative precipitation forecast of extreme rainstorms. Compared with traditional numerical prediction models and statistical methods, the method proposed in the present invention can more accurately predict the occurrence time, location, precipitation and duration of extreme precipitation, providing a more reliable scientific basis for meteorological disaster warning and disaster prevention and mitigation.

[0030] (2) The whole-layer dry intrusion index IDII constructed by the present invention comprehensively considers the temperature advection, potential vorticity and pseudo-equivalent potential temperature of dry cold air at different altitudes, and integrates these factors in the vertical direction, so as to more comprehensively reflect the three-dimensional structural characteristics of dry cold air and its impact on extreme rainstorms. It overcomes the deficiency of existing dry intrusion indices that it is difficult to accurately describe the activity characteristics of dry cold air under the influence of typhoons at a long distance, especially when the movement speed of the main body of dry cold air slows down and its impact on inland areas changes from "rapid invasion" to "slow penetration". The IDII index can more accurately capture the activity characteristics of dry cold air and provide a more reliable influencing factor for extreme precipitation prediction.

[0031] (3) The present invention integrates multi-source information such as historical precipitation event data, typhoon data, and reanalysis meteorological data through the acquisition and preprocessing of multi-source meteorological data, and ensures the quality and spatiotemporal consistency of input data through steps such as data cleaning, formatting, and spatiotemporal matching. Through multi-dimensional feature fusion, the model can make full use of data from different meteorological elements and IDII index fields, improving the effectiveness of model training and prediction. In addition, the TRP precipitation event feature data in the present invention is obtained based on the objective identification algorithm of typhoon long-distance precipitation, combined with multiple criteria such as rain belt separation, potential TRP discrimination, and water vapor conveyor belt identification, which significantly improves the accuracy and reliability of TRP event identification.

[0032] (4) The present invention uses deep learning technology to construct a TRP prediction model, which can effectively capture the nonlinear meteorological process under the long-distance influence of typhoons, and make full use of big data and machine learning technology. Compared with traditional numerical forecasting models and statistical methods, deep learning models have stronger nonlinear fitting and generalization capabilities, and can better learn and simulate complex meteorological processes, thereby improving the accuracy of TRP prediction. In addition, the present invention adopts a multi-task learning framework, which can not only predict the precipitation of long-distance extreme precipitation events of typhoons in inland areas, but also predict their occurrence time, location and duration, providing a precise forecast of precipitation at a fixed time, fixed point and quantitative, and providing strong technical support for the forecast and warning of extreme precipitation in inland areas and disaster prevention and mitigation. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] The accompanying drawings constituting a part of the present invention are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. The embodiments of the present invention are described in detail below in conjunction with the accompanying drawings, wherein:

[0034] Figure 1 The figure shows a flow chart of a method for predicting long-distance extreme precipitation of typhoons in inland areas based on deep learning and integrating the IDII index in Example 1 of the present invention.

[0035] Figure 2 Shown are vertical cross-sectional views of the dry intrusion index along 40°N at 14:00 (a) and 20:00 (b) on July 21, 2012 in Example 3 of the present invention.

[0036] Figure 3 Shown are the distribution diagrams of dry intrusion index on the isentropic surfaces at 315K (a) and 320K (c) at 14:00 on July 21, 2012 and at 315K (b) and 320K (d) at 20:00 in Example 3 of the present invention.

[0037] Figure 4 Shown is a superimposed graph of the hourly average regional precipitation (a) in Beijing from 08:00 on July 19 to 20:00 on July 23, 2012 and the hourly precipitation (b) at Fangshan Station and the corresponding whole layer (950-200hPa) average dry intrusion index changes in Example 3 of the present invention. DETAILED DESCRIPTION

[0038] The present invention relates to a method for predicting typhoon long-range extreme precipitation (TRP) in inland areas based on deep learning and integration of the whole-layer dry intrusion index (IDII). This method combines deep learning technology with multi-source meteorological data to construct a high-precision precipitation prediction model, aiming to improve the forecast accuracy of extreme precipitation events in inland areas. The specific implementation methods of the present invention will be described in detail below in conjunction with the accompanying drawings. The described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments in the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0039] Example 1

[0040] As a specific example, Figure 1 As shown, the inland typhoon long-distance extreme precipitation prediction method based on deep learning and integration of IDII index of the present invention is used to improve the timing, fixed point and quantitative refined forecasting capabilities of extreme precipitation in inland areas. The prediction method includes the following steps when implemented:

[0041] SS1. Construction of the whole-layer dry intrusion index IDII

[0042] By integrating the dry cold air movement at different height levels on the isobaric surface and the isentropic surface, the whole layer dry intrusion index is obtained. IDI p and IDI θ , to characterize the vertical penetration strength of dry cold air under the long-range effect of typhoon and its impact on extreme precipitation in inland areas:

[0043]

[0044] in, p is the air pressure, p blev and p tlev denote the bottom and top pressures of the isobaric surface integrals respectively; i is the potential temperature, i blev and i tlev denote the bottom potential temperature and top potential temperature of the isentropic surface integral, respectively; V p and V θ represent the horizontal wind speed vector on the isobaric surface and the isentropic surface respectively; T is the temperature, and denote the horizontal temperature gradients on the isobaric and isentropic surfaces, respectively; PV p and PV θ represent the potential vorticity on the isobaric surface and isentropic surface respectively; i se is the pseudo-equivalent potential temperature and is used to characterize the temperature at which a moist air mass reaches saturation in an adiabatic process. ( i se ) P and( i se ) θ represent the pseudo equivalent potential temperature on the isobaric surface and isentropic surface respectively;

[0045] Preferably, in the above step SS1, the bottom pressure and top pressure of the isobaric surface integral are p blev and p tlev The values ​​are set to 1000 hPa and 300 hPa respectively to cover the main atmospheric movement layers in the troposphere and fully capture the penetration characteristics of dry cold air at different altitudes; the bottom potential temperature and top potential temperature of the isentropic surface integral are i blev and itlev They are set to 310K and 342K respectively to reflect the vertical transport effect of dry and cold air on the isentropic surface and to improve the quantitative accuracy of the impact on extreme precipitation. At the same time, the central difference format is used to calculate the temperature gradient in each isobaric layer and isentropic surface, and the fourth-order Runge-Kutta method is used for numerical integration to improve the accuracy and numerical stability of the IDII index calculation.

[0046] In this step, in order to further improve the characterization accuracy of dry cold air, the whole layer dry intrusion index on the isobaric surface and isentropic surface is calculated. IDI p and IDI θ In addition to considering temperature, potential vorticity and pseudo equivalent potential temperature, relative humidity RH is also added as an influencing factor to more comprehensively characterize the characteristics of dry cold air. After adding relative humidity RH, the dry intrusion index of the entire layer on the isobaric surface and isentropic surface is IDI p and IDI θ The calculation formula is as follows:

[0047]

[0048] In the formula, RH p is the relative humidity on the isobaric surface, RH θ is the relative humidity on the isentropic surface.

[0049] SS2. Acquisition of TRP-related multi-source meteorological data

[0050] Obtain multi-source meteorological data related to typhoon long-distance extreme precipitation in inland areas, including at least historical TRP precipitation event data in inland areas, corresponding typhoon data, and reanalysis meteorological data of the same period, where TRP precipitation event data is precipitation event data with 24-hour cumulative precipitation reaching the extreme precipitation threshold obtained based on the objective identification algorithm of typhoon long-distance precipitation;

[0051] Preferably, in this step, the TRP precipitation event data at least includes the longitude and latitude of the precipitation site, the precipitation occurrence time, the 24-hour cumulative precipitation and / or the process cumulative precipitation, and the extreme precipitation threshold is set to the 24-hour cumulative precipitation exceeding 300 mm or the process cumulative precipitation exceeding 500 mm according to the actual situation of the prediction area, and the TRP precipitation event at least includes all samples that reach the extreme precipitation threshold in the TRP since 1979; the corresponding typhoon data at least includes the typhoon's path, intensity, structure and life cycle data, wherein the typhoon path data at least includes the typhoon's center position and its moving speed data, the typhoon intensity data at least includes the typhoon's central air pressure and maximum wind speed data, the typhoon structure data at least includes the typhoon's outer wind circle radius and spiral rain belt distribution data, and the typhoon The wind life cycle data at least include the typhoon's generation, development, extinction time and its evolution process data; the corresponding reanalysis meteorological data of the same period at least include the atmospheric motion, dynamic characteristics, thermodynamic parameters and water vapor transport characteristic data of each isobaric layer from 300 to 1000 hPa and each isentropic surface from 310 to 342 K, among which the atmospheric motion data at least include three-dimensional wind field, potential height field, pressure field, temperature field, relative humidity field and potential vorticity field data; the dynamic characteristic data at least include potential vorticity, vorticity and horizontal temperature advection data; the thermodynamic parameter data at least include pseudo equivalent potential temperature, relative humidity and convective effective potential energy data; the water vapor transport characteristic data are used to characterize the interaction between the typhoon's outer water vapor transport belt and the mid-latitude system and at least include the whole-layer water vapor flux and the whole-layer water vapor flux divergence.

[0052] In addition, in this step, TRP precipitation events are preferably obtained based on the objective identification algorithm of typhoon long-distance precipitation. TRP events are objectively identified through rain belt separation, potential TRP rain belt discrimination, and strong water vapor transport belt identification from typhoon to potential TRP area. First, the potential rain belt area is determined based on the station precipitation rate and spatial distribution. Then, the water vapor transport path of the typhoon is tracked and the contact area is determined through the whole-layer water vapor transport flux IVT. The IVT ≥ 200 kg·m -1 s -1 If the potential TRP area overlaps with the strong water vapor transport belt, it is determined to be a TRP event.

[0053] SS3. Preprocessing of TRP-related multi-source meteorological data

[0054] The multi-source meteorological data obtained are preprocessed including time-space matching and quality control. Then, based on the preprocessed meteorological data and combined with the definition of the whole layer dry intrusion index, the isobaric layers and isentropic surfaces are calculated. IDI p and IDI θData, finally, the preprocessed multi-source meteorological data and the calculated IDII index field data are fused to form multi-source fusion data, and organized into training data sets according to the time series according to the occurrence time of TRP precipitation events;

[0055] Preferably, in this step, the preprocessing operations of multi-source meteorological data may include data cleaning, formatting, missing value filling, outlier removal, data normalization, and time-space matching operations, wherein missing value filling adopts a multivariate linear interpolation method based on physical constraints, outlier removal is based on 3D s The data are normalized using the Min-Max normalization method to scale the data to a range between 0 and 1 in order to avoid the impact of differences in the order of magnitude of different features on model training. The spatiotemporal matching uses a dynamic time warping algorithm to align data of different temporal resolutions, and uses a bilinear interpolation method to unify the spatial grid resolution.

[0056] SS4. Construction of TRP prediction model based on deep learning and integrating IDII index features

[0057] The TRP prediction model is constructed by using deep learning technology. Its network structure includes at least an input layer, a feature extraction layer, an IDII feature fusion layer, a fully connected layer, and a TRP prediction output layer. The input layer is used to receive multi-source fusion data in the training data set. The feature extraction layer is used to extract the spatiotemporal feature information of the typhoon field, the meteorological element field, and the IDII index field from the input data. The IDII feature fusion layer is used to perform multi-dimensional fusion of the extracted spatiotemporal feature information. The fully connected layer is used to perform nonlinear transformation on the fused feature vector and map it to the TRP precipitation prediction output layer. The TRP prediction output layer is used to predict TRP precipitation events in inland areas in the future.

[0058] Preferably, in this step, the feature extraction layer includes at least three parallel feature extraction branches to extract spatiotemporal feature information from the input multi-source meteorological data, wherein the first branch uses a convolutional neural network or a bidirectional LSTM network to extract the temporal and spatial characteristics of the typhoon path, intensity and structure, the second branch uses a multi-scale residual convolutional neural network to extract the spatial characteristics of the meteorological element field, and the third branch uses a three-dimensional convolutional neural network combined with a spatiotemporal attention mechanism to extract IDI p and IDI θThe dynamic evolution characteristics of the field, the spatiotemporal feature outputs of the three branches are multi-dimensionally fused through the IDII feature fusion layer. The fusion method includes at least feature vector splicing, feature weighting and / or feature fusion based on the attention mechanism, so as to capture the complex interaction between typhoons, meteorological elements and dry and cold air penetration characteristics, improve the prediction accuracy and spatiotemporal resolution of TRP precipitation events, and learn the complex nonlinear relationship between the interaction between dry and cold air and typhoon outer water vapor through the fully connected layer operation.

[0059] In addition, in this step, the TRP prediction output layer preferably adopts a multi-task learning framework to design multiple parallel prediction branches to output the prediction results of TRP precipitation events in inland areas in the future. The prediction results at least include TRP precipitation, occurrence time, location and duration, among which: the precipitation prediction branch adopts a mixed Gaussian model to output probability distribution, the occurrence time prediction branch adopts a time series classifier to determine the time window of precipitation occurrence, the location prediction branch adopts a spatial attention mechanism to locate the precipitation center and the impact range, and the duration prediction branch adopts a regression model to estimate the duration of the precipitation process. Each prediction branch promotes each other between tasks through an adaptive weight mechanism, and introduces a Bayesian inference method to evaluate the uncertainty of the prediction results.

[0060] SS5. Training, validation and testing of TRP prediction model

[0061] The model training data set constructed in step SS3 is divided into a training set, a validation set, and a test set according to a preset ratio. The multi-source fusion data in the training set is used for model training through a deep learning algorithm. The validation set is used for model parameter adjustment and hyperparameter optimization. The test set is used to evaluate the generalization ability of the trained TRP prediction model under different typhoon, meteorological, and IDII index conditions.

[0062] Preferably, in this step, the training set is used for parameter optimization of the model, and the multi-source fusion data in the training set is used for model training through a deep learning algorithm, so that it can accurately learn the complex relationship between the input typhoon, meteorological, IDII index data and TRP precipitation event data, and the loss function of the model is based on a multi-task learning framework to comprehensively consider the errors in precipitation, occurrence time, location and duration, and introduce L1 / L2 regularization terms to suppress overfitting, use back propagation and gradient descent to update the model weights, and use a learning rate adaptive adjustment strategy to improve training efficiency.

[0063] In addition, the validation set in this step is used for model parameter adjustment and hyperparameter optimization. During the model training process, its performance is regularly evaluated on the validation set. According to the calculated spatiotemporal accuracy of precipitation prediction and / or the root mean square error index of precipitation prediction, the model parameters, learning rate, regularization parameters and network structure are adjusted to optimize the model performance. At the same time, the performance of the model on the validation set is used to determine whether overfitting occurs, so as to avoid overfitting of the model on the training set and poor performance on the test set.

[0064] At the same time, the test set in this step is used to evaluate the generalization ability of the trained TRP prediction model. The evaluation indicators include classification prediction indicators, regression prediction indicators and probability prediction indicators. The classification prediction indicators at least include the hit rate, false alarm rate and missed alarm rate of precipitation prediction, the regression prediction indicators at least include the root mean square error RMSE and mean absolute error MAE of precipitation prediction, and the probability prediction indicators at least include the Brier score and reliability diagram based on probability prediction. The prediction performance of the model under different weather systems and environmental conditions is comprehensively evaluated through multi-dimensional evaluation indicators.

[0065] SS6. Prediction and output of TRP events in inland areas

[0066] Acquire typhoon data and meteorological element field data of the target prediction area in real time, pre-process the acquired data and calculate the current moment IDI p and IDI θ The processed multi-source fusion data is input into the trained TRP prediction model to output the TRP prediction results of the target area in the future.

[0067] Specifically, the prediction and output of TRP events in inland areas in this step include:

[0068] SS61. Real-time data acquisition and preprocessing

[0069] Obtain the typhoon data and meteorological element field data of the target area in real time for the next 24 hours, perform quality control on the data and match it with time and space, and calculate the current time according to the method in step SS1 IDI p and IDI θ Field data: the pre-processed typhoon data, meteorological element field data and IDII index field data are combined into a feature vector as the input of the TRP prediction model;

[0070] SS62. TRP prediction model result generation

[0071] The feature vector is input into the trained TRP prediction model. The model predicts TRP events based on current meteorological conditions and historical characteristics, and outputs the precipitation, event time, location and duration in the target area in the future, which can be used for forecasting, warning and disaster prevention and mitigation of extreme precipitation in inland areas.

[0072] Example 2

[0073] On the basis of the above-mentioned Example 1, the method for predicting long-distance extreme precipitation of typhoons in inland areas of the present invention also includes SS7 for real-time correction and dynamic optimization of the prediction results, by acquiring the latest meteorological observation data of the prediction area in real time, comparing the prediction results of the TRP prediction model with the real-time observation data and calculating the prediction error, and dynamically optimizing the parameters and / or structure of the TRP prediction model by online learning or transfer learning methods according to the prediction error to reduce the prediction error, and performing real-time correction on the prediction results of the TRP prediction model based on the real-time observation data and the model parameter adjustment results.

[0074] Example 3

[0075] On the basis of the above-mentioned Example 1, this Example 3 combines the extremely heavy rainstorm process that occurred in Beijing on July 21, 2012 to verify the effectiveness of the method of the present invention. The reason for selecting this process is that it is extreme and typical. The average rainfall in the whole city reached 190.3 mm, setting a new daily precipitation extreme since 1951; the rainfall at 11 meteorological stations broke the historical extreme since the establishment of the stations; the center of the rainstorm appeared in Hebei Town, Fangshan District, with a rainfall of up to 460.0 mm. This example uses conventional ground meteorological observation data and NCEP / NCAR reanalysis data for verification and analysis, focusing on the spatiotemporal evolution characteristics of the IDII index in this process and its indicative significance for rainstorm prediction. The main verification results are as follows:

[0076] (1) Vertical structural characteristics of the IDII index

[0077] Figure 2 The vertical profiles of the dry intrusion index along 40°N at 14:00 (a) and 20:00 (b) on July 21, 2012 are shown. Figure 2As shown, the vertical profile characteristics of the representative station 54511 in Beijing (116.47°E, 39.80°N) are analyzed as an example. At the beginning of the rainstorm process (Figure a), the IDII index from the lower layer 1000hPa to the upper layer 100hPa above the station showed a negative distribution, and the negative center was located at an altitude of 650hPa; as the process developed, the positive range of the IDII index extended eastward to the low altitude, and during the heavy rain period (Figure b), the positive IDII index area appeared between 600-400hPa, forming a "wedge-shaped" intrusion feature of dry and cold air in the middle layer. This evolution process from negative to positive means that dry and cold air penetrated into the middle and low-level moist air flow, triggering strong convective instability, resulting in a significant increase in precipitation intensity, verifying the indicative role of the IDII index on precipitation triggering and enhancement. In this embodiment, by extracting and analyzing this feature, the deep learning model can accurately predict the enhancement process of extreme precipitation events.

[0078] (2) Horizontal distribution characteristics of the IDII index

[0079] In addition, based on the distribution of dry intrusion index on the 700hPa isobaric surface at 14:00 and 20:00 on July 21, 2012, an obvious positive and negative dipole distribution appeared near 115°E at the beginning of the heavy rain, and there were paired positive and negative centers of the IDII index on both sides of its zero line. During the heavy rain period, this pair of dipole centers is located in the Beijing area, with the positive anomaly center in the west and the negative anomaly center in the east. The intersection of cold and warm air corresponds to the large value area of ​​the gradient of the IDII index contour line. By analyzing the distribution of the dry intrusion index, it can be seen that heavy rain always occurs on the side of the dense belt of the dry intrusion index contour line close to the negative value area. When the positive and negative centers of the dry intrusion index gradually approach, the intensity of precipitation increases accordingly. This feature verifies the scientific nature of the feature fusion layer design in the present invention, that is, to indicate the heavy precipitation area by capturing the spatial gradient characteristics of the IDII index. In this embodiment, multi-level meteorological data of the typhoon long-distance impact area are selected, and the input features of the deep learning model are constructed by analyzing the dry intrusion index at different height levels.

[0080] (3) Isentropic surface IDII index analysis

[0081] Figure 3 The figure shows the distribution of dry intrusion index on the isentropic surfaces of 315K (a, b) and 320K (c, d) at 14:00 (a, c) and 20:00 (b, d) on July 21, 2012. Figure 3 As shown in the isentropic surface coordinate system, the IDII index on the 315K and 320K isentropic surfaces (corresponding to the middle and lower layers of the troposphere) exhibits a dipole distribution characteristic similar to that of the isobaric surface. Figure 3 In a and c, heavy rain occurs in relatively wet areas; during the peak period of heavy rain ( Figure 3In b and d), the heavy precipitation area appears on the side of the IDII index dense belt close to the negative value area, and in the hard-hit area in southwest Beijing, there is an obvious inflection point in the IDII index on the 320K isentropic surface, and the values ​​of the positive and negative anomaly centers are larger than those on the isobaric surface. This verifies the rationality of the multi-task learning framework adopted in the present invention, that is, the contributions of IDII features at different levels to precipitation prediction are different. In this embodiment, by constructing the IDII index on the isentropic surface, the model's ability to recognize dry intrusion effects at different height levels is enhanced. By calculating the dry intrusion index on the isentropic surface, the vertical and horizontal evolution characteristics of dry and cold air can be more comprehensively captured. After this feature is input into the deep learning model, the prediction accuracy of the model for the moment of precipitation enhancement is greatly improved.

[0082] (4) Forecast indication of the average IDII index of the whole layer

[0083] Figure 4 The figure shows the hourly average precipitation in Beijing (a) and the hourly precipitation at Fangshan Station (b) from 08:00 on July 19 to 20:00 on July 23, 2012, and the corresponding average dry intrusion index of the entire layer (950-200hPa). Figure 4 As shown in the figure, the whole-layer average IDII index of the analysis area average (Figure a) and Fangshan single station (Figure b) has obvious forecasting significance for heavy rain: 6 hours before the heavy rain, the whole-layer average IDII index continues to decrease from near zero; within 6 hours after the start of the heavy rain, the whole-layer average IDII index reaches a valley; within the next 6 hours (the heavy rain intensification period), the whole-layer average IDII index slowly increases; when the whole-layer average IDII index continues to increase and approaches zero, the heavy rain tends to end. This evolution feature shows that the whole-layer average IDII index has a 6-12 hour lead response time to the occurrence and intensification of heavy rain, which verifies the timeliness and practicality of the prediction method of the present invention. In this embodiment, by using the whole-layer average dry intrusion index as a time series input feature and combining it with the time series prediction ability of the deep learning model, it is possible to give an early warning of precipitation in the future. By capturing the time evolution law of IDII, the model can make predictions 6 to 12 hours before the heavy rain occurs, effectively improving the warning timeliness of extreme precipitation events.

[0084] (5) Data processing and fusion

[0085] In this embodiment, data processing and fusion are important steps in model training. By combining historical precipitation event data of typhoon remote impact areas, corresponding typhoon data and reanalysis meteorological data, and through data cleaning, formatting, time and space matching and other steps, the quality and consistency of input data are ensured. Through the integration of multi-source meteorological data, especially the construction and analysis of dry intrusion index, the deep learning model can effectively capture meteorological characteristics at different levels, and through a multi-task learning framework, realize the comprehensive prediction of precipitation, occurrence time, location and duration.

[0086] (6) Model training and prediction results

[0087] In this embodiment, the deep learning model can accurately identify the infiltration process of dry cold air and its impact on rainfall by training multi-source data of typhoon long-distance extreme precipitation events. The model uses a convolutional neural network, a time series module, and an IDII feature fusion layer, which can effectively extract the spatiotemporal characteristics of the dry intrusion index, and combine the changing trend of IDII to predict the spatiotemporal distribution and intensity of future precipitation. Through the training and verification of the model, the results show that with the help of the positive and negative value areas of the dry intrusion index IDII and its dipole distribution characteristics, the model can effectively predict the area where heavy rain occurs and the intensity of precipitation. Especially for TRP events, the application of the dry intrusion index in the model significantly improves the accuracy of the prediction.

[0088] In summary, this example demonstrates the potential of deep learning models in the prediction of extreme precipitation events by combining the construction of the dry intrusion index and its application in typhoon long-distance extreme precipitation. Through a comprehensive analysis of the spatiotemporal evolution of dry cold air and the distribution characteristics on isobaric and isentropic surfaces, the model can predict the occurrence, intensification and end of extreme precipitation in advance, providing strong support for extreme weather forecasting in inland areas.

[0089] The above specific implementation method realizes the TRP prediction of inland areas based on deep learning and integration of IDII index, effectively improves the spatiotemporal resolution and accuracy of the prediction model, and can provide reliable support for inland extreme precipitation forecasts. Those skilled in the art can understand that the present invention includes but is not limited to the contents described in the drawings and the above specific implementation methods. Although the present invention has been described with respect to the embodiments currently considered to be the most practical and preferred, it should be understood that the present invention is not limited to the disclosed embodiments, and any modifications that do not deviate from the functional and structural principles of the present invention will be included in the scope of the claims.

Claims

1. A method for predicting long-range extreme precipitation of typhoons in inland areas based on deep learning and integrating the IDII index, characterized in that: The prediction method comprises at least the following steps when implemented: SS1. By integrating the dry cold air movement at different height levels on the isobaric surface and isentropic surface, the whole layer dry intrusion index is obtained. IDII p and IDII θ , to characterize the vertical penetration strength of dry cold air under the long-range effect of typhoon and its impact on extreme precipitation in inland areas: in, p is the air pressure, p blev and p tlev denote the bottom and top pressures of the isobaric surface integrals respectively; θ is the potential temperature, θ blev and θ tlev denote the bottom potential temperature and top potential temperature of the isentropic surface integral, respectively; V p and V θ represent the horizontal wind speed vector on the isobaric surface and the isentropic surface respectively; T is the temperature, and denote the horizontal temperature gradients on the isobaric and isentropic surfaces, respectively; PV p and PV θ represent the potential vorticity on the isobaric surface and isentropic surface respectively; ( θ se ) P and( θ se ) θ represent the pseudo equivalent potential temperature on the isobaric surface and isentropic surface respectively; SS2. Obtain multi-source meteorological data related to typhoon long-distance extreme precipitation in inland areas, including at least historical TRP precipitation event data in inland areas, corresponding typhoon data, and reanalysis meteorological data of the same period. The TRP precipitation event data is the precipitation event data with 24-hour cumulative precipitation reaching the extreme precipitation threshold obtained based on the objective identification algorithm of typhoon long-distance precipitation; SS3. Preprocess the multi-source meteorological data, including time-space matching and quality control. Then, based on the preprocessed meteorological data and the definition of the whole layer dry intrusion index, calculate the isobaric layers and isentropic surfaces. IDII p and IDII θ Data, finally, the pre-processed multi-source meteorological data and the calculated IDII index field data are fused to form multi-source fusion data, and organized into a training data set according to the time series according to the occurrence time of TRP precipitation events; SS4. A TRP prediction model is constructed using deep learning technology. Its network structure includes at least an input layer, a feature extraction layer, an IDII feature fusion layer, a fully connected layer, and a TRP prediction output layer. The input layer is used to receive multi-source fusion data in the training data set. The feature extraction layer is used to extract the spatiotemporal feature information of the typhoon field, the meteorological element field, and the IDII index field from the input data. The IDII feature fusion layer is used to perform multi-dimensional fusion of the extracted spatiotemporal feature information. The fully connected layer is used to perform nonlinear transformation on the fused feature vector and map it to the TRP precipitation prediction output layer. The TRP prediction output layer is used to predict TRP precipitation events in inland areas in the future. SS5. Divide the model training data set constructed in step SS3 into a training set, a validation set, and a test set according to a preset ratio. Perform model training on the multi-source fusion data in the training set through a deep learning algorithm. The validation set is used for model parameter adjustment and hyperparameter optimization. The test set is used to evaluate the generalization ability of the trained TRP prediction model under different typhoon, meteorological, and IDII index conditions. SS6. Obtain typhoon data and meteorological element field data of the target prediction area in real time, pre-process the acquired data and calculate the current time. IDII p and IDII θ The processed multi-source fusion data is input into the TRP prediction model to output the TRP prediction results of the target area in the future.

2. The method for predicting long-range extreme precipitation of typhoons in inland areas based on deep learning and integrating IDII index according to claim 1 is characterized in that: In the above step SS1, the bottom pressure and top pressure of the isobaric surface integral are p blev and p tlev The values ​​are set to 1000 hPa and 300 hPa respectively to cover the main atmospheric movement layers in the troposphere and fully capture the penetration characteristics of dry cold air at different altitudes; the bottom potential temperature and top potential temperature of the isentropic surface integral are θ blev and θ tlev They are set to 310K and 342K respectively to reflect the vertical transport effect of dry and cold air on the isentropic surface and to improve the quantitative accuracy of the impact on extreme precipitation. At the same time, the central difference format is used to calculate the temperature gradient in each isobaric layer and isentropic surface, and the fourth-order Runge-Kutta method is used for numerical integration to improve the accuracy and numerical stability of the IDII index calculation.

3. The method for predicting long-range extreme precipitation of typhoons in inland areas based on deep learning and integrating IDII index according to claim 1 is characterized in that: In the above step SS1, the whole layer dry intrusion index on the isobaric surface and isentropic surface is calculated IDII p and IDII θ In addition to considering temperature, potential vorticity and pseudo equivalent potential temperature, relative humidity RH is also added as an influencing factor to more comprehensively characterize the characteristics of dry cold air. After adding relative humidity RH, the dry intrusion index of the entire layer on the isobaric surface and isentropic surface is IDII p and IDII θ The mathematical calculation formulas are as follows: In the formula, RH p is the relative humidity on the isobaric surface, RH θ is the relative humidity on the isentropic surface.

4. The method for predicting long-range extreme precipitation of typhoons in inland areas based on deep learning and integrating IDII index according to claim 1 is characterized in that: In the above step SS2, the TRP precipitation event data at least includes the longitude and latitude of the precipitation site, the precipitation occurrence time, the 24-hour cumulative precipitation and / or the process cumulative precipitation, and the extreme precipitation threshold is set to 24h cumulative precipitation exceeding 300mm or process cumulative precipitation exceeding 500mm according to the actual situation of the prediction area, and the TRP precipitation event at least includes all samples that reach the extreme precipitation threshold in the TRP since 1979; the corresponding typhoon data at least includes the typhoon's path, intensity, structure and life cycle data, wherein the typhoon path data at least includes the typhoon's center position and its moving speed data, the typhoon intensity data at least includes the typhoon's central air pressure and maximum wind speed data, and the typhoon structure data at least includes the typhoon's outer wind circle radius and spiral rain belt distribution data, The life cycle data of a typhoon at least include the data on the formation, development, extinction time and its evolution process of the typhoon; the corresponding reanalysis meteorological data of the same period at least include the atmospheric motion, dynamic characteristics, thermodynamic parameters and water vapor transport characteristic data of each isobaric layer from 300 to 1000 hPa and each isentropic surface from 310 to 342 K, among which the atmospheric motion data at least include the three-dimensional wind field, potential height field, pressure field, temperature field, relative humidity field and potential vorticity field data; the dynamic characteristic data at least include potential vorticity, vorticity and horizontal temperature advection data; the thermodynamic parameter data at least include pseudo equivalent potential temperature, relative humidity and convective effective potential energy data; the water vapor transport characteristic data are used to characterize the interaction between the typhoon's outer water vapor transport belt and the mid-latitude system and include at least the whole-layer water vapor flux and the whole-layer water vapor flux divergence.

5. The method for predicting long-range extreme precipitation of typhoons in inland areas based on deep learning and integrating IDII index according to claim 1 is characterized in that: In the above step SS2, TRP precipitation events are obtained based on the objective identification algorithm of typhoon long-distance precipitation. TRP events are objectively identified through rain belt separation, potential TRP rain belt discrimination, and strong water vapor transport belt identification from typhoon to potential TRP area. First, the potential rain belt area is determined based on the station precipitation rate and spatial distribution. Then, the water vapor transport path of the typhoon is tracked and the contact area is determined through the whole-layer water vapor transport flux IVT. The IVT ≥ 200 kg·m -1 s -1 If the potential TRP area overlaps with the strong water vapor transport belt, it is determined to be a TRP event.

6. The method for predicting long-distance extreme precipitation of typhoons in inland areas based on deep learning and integrating IDII index according to claim 1 is characterized in that: In the above step SS3, the preprocessing operations on multi-source meteorological data at least include data cleaning, formatting, missing value filling, outlier removal, data normalization and space-time matching operations, wherein missing value filling adopts a multivariate linear interpolation method based on physical constraints, outlier removal is based on the 3σ criterion and combined with the physical threshold of meteorological elements, data normalization adopts the Min-Max normalization method to scale the data to a range between 0 and 1 to avoid the impact of the order of magnitude difference between different features on model training, and space-time matching adopts a dynamic time warping algorithm to align data with different time resolutions, and uses a bilinear interpolation method to unify the spatial grid resolution.

7. The method for predicting long-distance extreme precipitation of typhoons in inland areas based on deep learning and integrating IDII index according to claim 1, characterized in that: In the above step SS4, the feature extraction layer includes at least three parallel feature extraction branches to extract spatiotemporal feature information from the input multi-source meteorological data, wherein the first branch uses a convolutional neural network or a bidirectional LSTM network to extract the temporal and spatial characteristics of the typhoon path, intensity and structure, the second branch uses a multi-scale residual convolutional neural network to extract the spatial characteristics of the meteorological element field, and the third branch uses a three-dimensional convolutional neural network combined with a spatiotemporal attention mechanism to extract IDII p and IDII θ The dynamic evolution characteristics of the field, the spatiotemporal feature outputs of the three branches are multi-dimensionally fused through the IDII feature fusion layer. The fusion method includes at least feature vector splicing, feature weighting and / or feature fusion based on the attention mechanism, so as to capture the complex interaction between typhoons, meteorological elements and dry and cold air penetration characteristics, improve the prediction accuracy and spatiotemporal resolution of TRP precipitation events, and learn the complex nonlinear relationship between the interaction between dry and cold air and typhoon outer water vapor through the fully connected layer operation.

8. The method for predicting long-distance extreme precipitation of typhoons in inland areas based on deep learning and integrating IDII index according to claim 1, characterized in that: In the above step SS4, the TRP prediction output layer adopts a multi-task learning framework to design multiple parallel prediction branches to output the prediction results of TRP precipitation events in inland areas in the future. The prediction results at least include TRP precipitation, occurrence time, location and duration, among which: the precipitation prediction branch adopts a mixed Gaussian model to output probability distribution, the occurrence time prediction branch adopts a time series classifier to determine the time window of precipitation occurrence, the location prediction branch adopts a spatial attention mechanism to locate the precipitation center and the impact range, and the duration prediction branch adopts a regression model to estimate the duration of the precipitation process. Each prediction branch promotes each other between tasks through an adaptive weight mechanism, and introduces a Bayesian inference method to evaluate the uncertainty of the prediction results.

9. The method for predicting long-distance extreme precipitation of typhoons in inland areas based on deep learning and integrating IDII index according to claim 1, characterized in that: In the above step SS5, the training set is used for parameter optimization of the model. The multi-source fusion data in the training set is used for model training through a deep learning algorithm, so that it can accurately learn the complex relationship between the input typhoon, meteorological, IDII index data and TRP precipitation event data. The loss function of the model is based on a multi-task learning framework to comprehensively consider the errors in precipitation, occurrence time, location and duration, and introduces L1 / L2 regularization terms to suppress overfitting. The model weights are updated using back propagation and gradient descent methods, and the learning rate adaptive adjustment strategy is used to improve training efficiency.

10. The method for predicting long-distance extreme precipitation of typhoons in inland areas based on deep learning and integrating IDII index according to claim 1, characterized in that: In the above step SS5, the validation set is used for model parameter adjustment and hyperparameter optimization. During the model training process, its performance is regularly evaluated on the validation set. According to the calculated spatiotemporal accuracy of precipitation prediction and / or the root mean square error index of precipitation prediction, the model parameters, learning rate, regularization parameters and network structure are adjusted to optimize the model performance. At the same time, whether overfitting occurs is judged by the performance of the model on the validation set to avoid overfitting of the model on the training set and poor performance on the test set.

11. The method for predicting long-distance extreme precipitation of typhoons in inland areas based on deep learning and integrating IDII index according to claim 1, characterized in that: In the above step SS5, the test set is used to evaluate the generalization ability of the trained TRP prediction model. The evaluation indicators include classification prediction indicators, regression prediction indicators and probability prediction indicators, wherein the classification prediction indicators at least include the hit rate, false alarm rate and missed alarm rate of precipitation prediction, the regression prediction indicators at least include the root mean square error RMSE and mean absolute error MAE of precipitation prediction, and the probability prediction indicators at least include the Brier score and reliability diagram based on probability prediction. The prediction performance of the model under different weather systems and environmental conditions is comprehensively evaluated through multi-dimensional evaluation indicators.

12. The method for predicting long-distance extreme precipitation of typhoons in inland areas based on deep learning and integrating IDII index according to claim 1, characterized in that: In the above step SS6, the prediction and output of TRP events in the inland area mainly includes the following sub-steps when implemented: SS61. Real-time data acquisition and preprocessing Obtain the typhoon data and meteorological element field data of the target area in real time for the next 24 hours, perform quality control on the data and match it with time and space, and calculate the current time according to the method in step SS1 IDII p and IDII θ Field data: the pre-processed typhoon data, meteorological element field data and IDII index field data are combined into a feature vector as the input of the TRP prediction model; SS62. TRP prediction model result generation The feature vector is input into the trained TRP prediction model. The model predicts TRP events based on current meteorological conditions and historical characteristics, and outputs the precipitation, event time, location and duration in the target area in the future, which can be used for forecasting, warning and disaster prevention and mitigation of extreme precipitation in inland areas.

13. The method for predicting long-distance extreme precipitation of typhoons in inland areas based on deep learning and integrating IDII index according to claim 1, characterized in that: The method for predicting long-distance extreme precipitation of typhoons in inland areas also includes SS7, which performs real-time correction and dynamic optimization of prediction results. By acquiring the latest meteorological observation data of the prediction area in real time, the prediction results of the TRP prediction model are compared with the real-time observation data and the prediction error is calculated. According to the prediction error, the parameters and / or structure of the TRP prediction model are dynamically optimized by online learning or transfer learning methods to reduce the prediction error. Based on the real-time observation data and the model parameter adjustment results, the prediction results of the TRP prediction model are corrected in real time.

14. A computer program product comprising computer instructions, characterized in that The computer instructions are used to execute the inland typhoon long-distance extreme precipitation prediction method based on deep learning and integration of the IDII index as described in any one of claims 1 to 13 above.

15. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, the method for predicting long-distance extreme precipitation of typhoons in inland areas based on deep learning and integration of the IDII index as described in any one of claims 1 to 13 above is implemented.

Images