Northwest Pacific Ocean Typhoon precipitation correction method based on deep learning model

Abstract

The invention relates to the technical field of precipitation scales, in particular to a northwest pacific ocean typhoon precipitation correction method based on a deep learning model, and the method comprises the steps: obtaining ERA5 data, GPM data and a tropical cyclone optimal path data set of the northwest pacific ocean, obtaining an active time period and an influence range of historical typhoons, and obtaining a prediction result; extracting rainfall data corresponding to each typhoon from the ERA5 data and the GPM data, and preprocessing the rainfall data to obtain a data set of a training model; constructing a northwest Pacific Ocean typhoon precipitation correction model, and training the model by using the data set; and obtaining ERA5 reanalysis rainfall data of the target typhoon, and inputting the ERA5 reanalysis rainfall data into the trained northwest pacific ocean typhoon rainfall correction model to obtain a rainfall correction result of the target typhoon. According to the invention, through the non-linear feature extraction capability of the deep learning model, the prediction accuracy of the reanalysis data on typhoon precipitation is improved; the model can retain a fine space structure of a precipitation field, capture an evolution rule of typhoon precipitation, and provide a scientific basis for disaster prevention decision.

Description

Technical Field

[0001] The present invention relates to the field of precipitation scale technology, and in particular to a northwest Pacific typhoon precipitation correction method based on a deep learning model. Background Art

[0002] Approximately one-third of the world's typhoons form annually in the northwest Pacific, with my country's southeastern coastal areas being among the hardest hit. In addition to strong winds, heavy rain is also a significant factor in typhoon destructiveness. Typhoon precipitation often leads to secondary disasters such as urban flooding, flash floods, landslides, and mudslides, causing significant economic losses and disasters. Inaccurate precipitation forecasts can trigger natural disasters, impacting urban management and social security. Therefore, obtaining accurate precipitation information and issuing timely weather warnings are crucial for preventing natural disasters and protecting social production, as well as the safety of people's lives and property.

[0003] Traditional typhoon precipitation forecasts primarily rely on numerical weather prediction models, whose forecast products are derived from calculations based on numerous physics and dynamics equations. Despite continuous technological advancements and improvements in numerical weather prediction models, these methods still have limitations in their application to nowcasting. These methods require complex physics equations, and errors in the forecast process can be continuously propagated into the final results, leading to biased results. This makes it difficult to meet the accuracy and real-time requirements for precise forecasts.

[0004] It's worth noting that while ERA5, a widely used reanalysis dataset, boasts high accuracy in assimilating meteorological elements, its precipitation data still exhibit significant systematic biases. Studies have shown that due to the complex microphysical parameterization and convective-scale processes involved in precipitation, ERA5's ability to capture the spatiotemporal distribution of typhoon precipitation is limited. This leads to frequent underestimates of extreme precipitation events, with magnitude forecasts for heavy rainfall areas within typhoon centers generally underestimating them by 20%-40%. Furthermore, due to the limited grid resolution (0.25°×0.25°), ERA5's spatial distribution is limited, making it difficult to accurately capture the spatial details of orographically forced precipitation and small-scale convective systems. Consequently, phase deviations of up to 3-6 hours are common in the time series evolution of precipitation peaks. These biases are transmitted through the data assimilation system to the initial numerical forecast field, further impacting forecast accuracy.

[0005] With the rapid development of deep learning technology, using deep learning to correct numerical forecast data has become a common method for improving forecast accuracy. In particular, the nonlinear mapping capabilities of deep learning can effectively compensate for the inherent biases in reanalysis data such as ERA5. The volume of meteorological observation data and numerical model forecast data is enormous, and deep learning has the ability to extract complex spatiotemporal features from large amounts of high-dimensional, spatiotemporally distributed meteorological data. Deep learning models possess powerful feature extraction, feature learning, and nonlinear modeling capabilities. Correcting the bias of numerical model forecast data based on deep learning models can effectively establish a nonlinear relationship between precipitation observations and forecast data, addressing the shortcomings of traditional numerical forecast models, eliminating biases and systematic errors in forecast data, and thereby improving the accuracy and reliability of typhoon precipitation forecasts.

[0006] However, typhoon precipitation exhibits high spatiotemporal variability, is influenced by multiple environmental factors, and is generated by a complex coupling of physical processes. Traditional statistical and dynamic precipitation forecasting methods still suffer from the following challenges: They rely heavily on the integrity and accuracy of historical data, resulting in poor adaptability to diverse terrains; they require the use of parameterized physical equations with significant uncertainty to simulate atmospheric motion; and they primarily seek linear relationships between observed and forecasted data, failing to accurately reflect the nonlinear relationships inherent in complex data. Therefore, how to overcome the limitations of traditional methods through deep learning technology and establish a highly accurate and robust typhoon precipitation forecast model remains a pressing technical challenge in the field of meteorological disaster prevention and control. Summary of the Invention

[0007] In view of this, the purpose of the present invention is to provide a typhoon precipitation correction method for the northwest Pacific based on a deep learning model to solve the problems of systematic deviation and insufficient accuracy in typhoon precipitation forecasting by traditional numerical forecasting models in the prior art.

[0008] According to a first aspect of an embodiment of the present invention, a method for correcting typhoon precipitation in the northwest Pacific is provided, comprising: Obtain ERA5 reanalysis precipitation data, GPM satellite precipitation data, and tropical cyclone best track datasets over the Northwest Pacific; The active period and impact range of each typhoon in history are obtained based on the tropical cyclone best track dataset, and precipitation data corresponding to each typhoon is extracted from the ERA5 reanalysis precipitation data and the GPM satellite precipitation data based on the active period and impact range; Preprocess the precipitation data corresponding to each typhoon to obtain the data set for training the model; A deep learning-based typhoon precipitation correction model for the Northwest Pacific was constructed. The ERA5 reanalysis precipitation data in the dataset was used as training data, and the GPM satellite precipitation data in the dataset was used as label data. The Northwest Pacific typhoon precipitation correction model was trained according to the set parameters and loss function. The ERA5 reanalysis precipitation data of the target typhoon is obtained and input into the trained northwest Pacific typhoon precipitation correction model to obtain the precipitation correction result of the target typhoon.

[0009] Preferably, the constructed Northwest Pacific typhoon precipitation correction model based on deep learning includes: Input layer, spatial encoder module, temporal feature extraction module, spatial decoder module and multi-time-step precipitation prediction module; The input layer receives the ERA5 reanalysis precipitation data and converts it into a tensor format suitable for neural network processing; The spatial encoder module extracts multi-scale spatial features from the data output by the input layer through a four-level downsampling structure; The temporal feature extraction module introduces a bidirectional spatiotemporal module in the deepest layer of the encoder, reorganizes the multi-scale spatial features into a temporal sequence for time dependency modeling, and realizes the fusion of spatiotemporal features through linear projection; The spatial decoder module adopts an upsampling structure symmetrical to the encoder and integrates spatial features of different scales through skip connections to obtain decoding features; The multi-time-step precipitation prediction module performs scale adjustment and channel transformation on the decoded features, and finally outputs a multi-step precipitation correction field.

[0010] Preferably, the four-stage downsampling structure of the spatial encoder module includes convolution operation, batch normalization and nonlinear activation at each stage.

[0011] Preferably, the method further comprises: The mean square error is used as the loss function to measure the difference between the corrected map at each time step in the multi-step precipitation correction field and the true GPM satellite precipitation data.

[0012] Preferably, training the Northwest Pacific typhoon precipitation correction model further includes: Calculating the absolute deviation, root mean square error and anomaly correlation coefficient between the multi-step precipitation correction field and the true GPM satellite precipitation data; Performing a comprehensive evaluation of the performance of the Northwest Pacific typhoon precipitation correction model based on the absolute deviation, the root mean square error, and the anomaly correlation coefficient; Based on the evaluation results, the model parameters are optimized and adjusted.

[0013] Preferably, the impact range of each typhoon in history is obtained based on the tropical cyclone best track dataset, including: The central longitude and latitude coordinates of each typhoon in history are obtained based on the tropical cyclone best track dataset; The typhoon's impact range is defined as a circular area with a radius of 500 km, centered on the typhoon's central latitude and longitude coordinates.

[0014] Preferably, the precipitation data corresponding to each typhoon is preprocessed, including: The sliding window technique is used to construct the correspondence between the 24-hour input sequence and the 24-hour forecast target, with a sliding step of 3 hours. The improved Z-score method is used to detect anomalies. If an outlier is identified, it is replaced by the temporal and spatial neighborhood mean to obtain the ERA5 and GPM datasets corresponding to the typhoon.

[0015] Preferably, the method further comprises: The data in the ERA5 dataset and the GPM dataset are standardized using the following formulas:

[0016] Among them, x is the original value of the data, is the mean of the data set, is the standard deviation of the data set, is the value after data standardization.

[0017] The technical solutions provided by the embodiments of the present invention may have the following beneficial effects: It is understood that the technical solution shown in the present invention can obtain ERA5 data, GPM data, and tropical cyclone optimal path data sets for the Northwest Pacific, derive the active periods and impact ranges of historical typhoons, and then extract precipitation data corresponding to each typhoon from the ERA5 data and the GPM data, which are then preprocessed to obtain a data set for training the model; construct a Northwest Pacific typhoon precipitation correction model based on deep learning, and train the model using the data set; obtain ERA5 reanalysis precipitation data for the target typhoon, input it into the trained Northwest Pacific typhoon precipitation correction model, and obtain the precipitation correction result for the target typhoon. The present invention improves the forecast accuracy of typhoon precipitation using reanalysis data through the nonlinear feature extraction capability of the deep learning model; the model can preserve the fine spatial structure of the precipitation field, capture the evolution of typhoon precipitation, and provide a scientific basis for disaster prevention decision-making.

[0018] It is to be understood that the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention.

[0020] Figure 1 This is a schematic diagram of a method for correcting typhoon precipitation in the northwest Pacific based on a deep learning model according to an exemplary embodiment; Figure 2 is a schematic diagram of a data processing flow according to an exemplary embodiment; Figure 3 is a structural diagram of a correction model according to an exemplary embodiment. DETAILED DESCRIPTION

[0021] Exemplary embodiments will be described in detail herein, examples of which are illustrated in the accompanying drawings. In the following description, when referring to the drawings, like numbers in different figures represent like or similar elements unless otherwise indicated. The embodiments described in the following exemplary embodiments are not intended to represent all possible embodiments consistent with the present invention. Rather, they are merely examples of apparatus and methods consistent with certain aspects of the present invention, as detailed in the appended claims.

[0022] In one embodiment, Figure 1 This is a schematic diagram of a method for correcting typhoon precipitation in the Northwest Pacific based on a deep learning model according to an exemplary embodiment. Figure 1 , provides a typhoon precipitation correction method for the northwest Pacific based on a deep learning model, including: Step S11: Obtain ERA5 reanalysis precipitation data, GPM satellite precipitation data, and tropical cyclone best track dataset for the northwest Pacific.

[0023] Step S12: derive the active period and impact range of each typhoon in history based on the tropical cyclone best path dataset, and extract precipitation data corresponding to each typhoon from the ERA5 reanalysis precipitation data and the GPM satellite precipitation data based on the active period and impact range.

[0024] Step S13: Preprocess the precipitation data corresponding to each typhoon to obtain a data set for training the model.

[0025] Step S14: Construct a deep learning-based typhoon precipitation correction model for the northwest Pacific Ocean, use the ERA5 reanalysis precipitation data in the dataset as training data, use the GPM satellite precipitation data in the dataset as label data, and train the northwest Pacific Ocean typhoon precipitation correction model according to the set parameters and loss function.

[0026] Step S15: Obtain the ERA5 reanalysis precipitation data of the target typhoon, input it into the trained Northwest Pacific typhoon precipitation correction model, and obtain the precipitation correction result of the target typhoon.

[0027] It is understood that the technical solution shown in the present invention can obtain ERA5 data, GPM data, and tropical cyclone optimal path data sets for the Northwest Pacific, derive the active periods and impact ranges of historical typhoons, and then extract precipitation data corresponding to each typhoon from the ERA5 data and the GPM data, which are then preprocessed to obtain a data set for training the model; construct a Northwest Pacific typhoon precipitation correction model based on deep learning, and train the model using the data set; obtain ERA5 reanalysis precipitation data for the target typhoon, input it into the trained Northwest Pacific typhoon precipitation correction model, and obtain the precipitation correction result for the target typhoon. The present invention improves the forecast accuracy of typhoon precipitation using reanalysis data through the nonlinear feature extraction capability of the deep learning model; the model can preserve the fine spatial structure of the precipitation field, capture the evolution of typhoon precipitation, and provide a scientific basis for disaster prevention decision-making.

[0028] In practice, in step S11, the ERA5 reanalysis precipitation data is ECMWF's fifth-generation atmospheric reanalysis dataset, providing hourly estimates of a large number of atmospheric, land, and oceanic climate variables. This product has a coarse spatial resolution (25 km) and a temporal resolution of 1 hour, covering the period from 1979 to the present.

[0029] GPM (Global Precipitation Measurement) is an international satellite observation program. It integrates data from multiple sensors, including visible infrared and passive microwave, from geostationary and polar-orbiting satellites to establish a global, high-precision precipitation monitoring system. The core of the GPM system is the GPM Master Satellite, equipped with advanced instruments such as the Dual-Frequency Precipitation Radar (DPR) and the GPM Microwave Imager (GMI), enabling precise measurements of precipitation intensity, spatial distribution, and vertical structure.

[0030] The GPM program's most important data product is the Integrated Multi-Satellite Precipitation Product (IMERG). This product fuses observations from the GPM core satellites with those from other low-orbiting satellites, supplemented by spatial and temporal interpolation of infrared observations from geostationary satellites. This product ultimately produces a global precipitation dataset with a temporal and spatial resolution of half an hour and 0.1° (approximately 10 kilometers). IMERG utilizes advanced Kalman filtering algorithms to process sparse data and performs regional corrections using ground-based precipitation gauges to ensure data accuracy. Based on data processing procedures and accuracy differences, IMERG is divided into three versions: the Early version provides near-real-time precipitation monitoring, the Late version undergoes preliminary corrections, and the Final version is the most accurate and fully verified and calibrated version.

[0031] GPM precipitation data plays a vital role in climate research, weather forecasting, water resources management, and natural disaster monitoring. Its high temporal and spatial resolution enables it to capture small- and medium-scale precipitation processes, providing crucial data support for the study of the global water cycle and extreme precipitation events. GPM has become one of the world's most authoritative satellite precipitation observation programs, and its data products are widely used in scientific research and operational fields.

[0032] The CMA Tropical Cyclone Best Tracks Dataset is authoritative historical typhoon data, covering tropical cyclone activity in the Northwest Pacific (including the South China Sea). Based on multi-source observations (satellite, radar, ship, buoy, etc.) and reanalysis data, and manually verified and revised, this dataset is a core data source for studying typhoon tracks, intensities, and climatic characteristics.

[0033] It should be noted that in step S12, the impact range of each typhoon in history is obtained based on the tropical cyclone best path dataset, including: The central longitude and latitude coordinates of each typhoon in history are obtained based on the tropical cyclone best track dataset; a circular area with a radius of 500 km is taken as the typhoon's impact range, with the central longitude and latitude coordinates of the typhoon as the center.

[0034] In specific practice, taking the northwest Pacific as an example, the distance of the target area can be calculated using the Haversine formula to accurately determine the impact range of the typhoon.

[0035] Haversine(lon1,lat1,lon2,lat2) is a function that calculates the spherical distance (great-circle distance) between two points on the Earth's surface in kilometers (km). It is widely used in meteorology, geographic information systems (GIS), and navigation, and is particularly suitable for calculating the distance between a typhoon center and surrounding precipitation data points.

[0036] The Haversine formula is based on the principles of spherical trigonometry, and its core formula is:

[0037] Among them: is the spherical distance between the two points (km); is the radius of the Earth (mean radius =6371km); , is the latitude of the two points (unit: radians); , is the longitude of the two points (unit: radians); is the latitude difference; is the longitude difference.

[0038] Based on the calculation results of this formula, a circular impact area with a radius of 500 kilometers was established with the typhoon center as the origin. Within this range, ERA5 reanalysis data and GPM satellite precipitation data were simultaneously extracted, and finally a time-aligned typhoon precipitation field dataset was constructed, providing high-quality input data for subsequent deep learning modeling. Figure 2 This method effectively balances the relationship between typhoon precipitation field coverage and data quality. The 500-kilometer extraction radius not only fully includes the main typhoon precipitation area, but also avoids the introduction of irrelevant meteorological noise due to a large range.

[0039] As can be understood, the data was fused using a multi-source data fusion method, integrating ERA5 reanalysis precipitation data, GPM satellite precipitation data, and the Tropical Cyclone Optimum Track (CMA) dataset. During data processing, the CMA first accurately determined the active period of each typhoon based on the typhoon center's latitude and longitude coordinates and life history information. Subsequently, for each typhoon sample, ERA5 and GPM precipitation data for the corresponding period were simultaneously extracted using a 500km radius of influence, based on the typhoon center's location. This rigorous spatiotemporal matching method not only ensures a precise correspondence between precipitation data and typhoon systems, but also effectively captures the spatial distribution characteristics of typhoon precipitation, providing high-quality spatiotemporal matching training samples for subsequent deep learning modeling.

[0040] Then, step S13 needs to be executed to pre-process the acquired data.

[0041] It should be noted that the precipitation data corresponding to each typhoon is preprocessed, including: The sliding window technique is used to construct the correspondence between the 24-hour input sequence (8 time steps) and the 24-hour prediction target (8 time steps). The sliding step is 3 hours (1 time step). The improved Z-score method is used to detect anomalies:

[0042] Among them, z is the number of anomalies, which is used to determine whether it is abnormal. is the observation value of a single data point to be detected; is the median of the data set; For the data set The observed value of the data point.

[0043] If an outlier is identified, for example, |z|>3.5, the outlier is replaced by the spatiotemporal neighborhood mean to obtain the ERA5 and GPM datasets corresponding to the typhoon.

[0044] It should be noted that, considering the different dimensions of the ERA5 reanalysis data and the GPM satellite precipitation dataset, which may have a certain impact on the training process, the two datasets are normalized and converted to a distribution with a mean of 0 and a standard deviation of 1. The details are as follows: The data in the ERA5 dataset and the GPM dataset are standardized using the following formulas:

[0045] Among them, x is the original value of the data, is the mean of the data set, is the standard deviation of the data set, is the value after data standardization.

[0046] It is understandable that, after identifying and removing incomplete, incorrect, inaccurate or irrelevant parts of the data, this embodiment performs standardized preprocessing on the precipitation dataset and adopts a strict data preprocessing process to ensure the reliability of model training.

[0047] Preferably, when executing step S13, the preprocessed data can be divided into a training set, a validation set and a test set to facilitate the training of subsequent models.

[0048] For example, the data was divided into three independent time periods chronologically: a training set (2001-2018), a validation set (2019), and a test set (2020). To minimize the risk of data leakage, the three datasets were stored in physically isolated locations. The validation and test sets were strictly kept "unseen" and used only for model tuning and final evaluation. All preprocessing parameters were calculated solely from the training set, ensuring complete independence between datasets. This approach not only meets the characteristics of time series forecasting problems but also truly reflects the model's performance in real-world applications, laying a solid data foundation for subsequent deep learning modeling.

[0049] Then, step S14 is executed to construct a northwest Pacific typhoon precipitation correction model based on deep learning.

[0050] In order to fully exploit the complex nonlinear spatiotemporal relationship between ERA5 input and GPM precipitation output, this embodiment proposes an improved spatiotemporal-UNet architecture for multi-time-step typhoon precipitation correction. The spatiotemporal-UNet precipitation correction model is a deep learning architecture that combines the spatial feature extraction capabilities of convolutional neural networks and the temporal modeling capabilities of long-short-term memory networks. It is specifically designed for multi-step prediction tasks of high-resolution precipitation fields. The model is designed to process the temporal evolution law and spatial distribution characteristics of precipitation data. Its structure is as follows: Figure 3As shown, it should be noted that it includes an input layer, a spatial encoder module, a temporal feature extraction module, a spatial decoder module and a multi-time step precipitation prediction module.

[0051] The ERA5 reanalysis precipitation data received by the input layer is converted into a tensor format suitable for neural network processing. The spatial encoder module extracts multi-scale spatial features from the data output by the input layer through a four-level downsampling structure; each level includes convolution operations, batch normalization and nonlinear activation, and gradually constructs feature representations from local to global. The temporal feature extraction module introduces a bidirectional spatiotemporal module in the deepest layer of the encoder, reorganizes the multi-scale spatial features into a temporal sequence for time dependency modeling, and realizes the fusion of spatiotemporal features through linear projection. The spatial decoder module adopts an upsampling structure symmetrical to the encoder, integrates spatial features of different scales through jump connections, obtains decoding features, and gradually restores high-resolution details. The multi-time-step precipitation prediction module rescales and transforms the decoded features, and finally outputs a multi-step precipitation correction field.

[0052] The entire model achieves a balance between feature extraction and reconstruction through the UNet encoding and decoding architecture, and combines the advantages of spatiotemporal time series modeling to form an end-to-end spatiotemporal forecasting system.

[0053] The model's core innovations include: a multi-scale spatiotemporal feature separation and extraction mechanism: a dual-path approach is used to separately process the spatial structure of the precipitation field and the temporal evolution of the precipitation pattern; a typhoon center adaptive perception design: a 500km radius + 5° expansion input cropping strategy is used to focus on the typhoon's impact area; and spatiotemporal dynamic fusion: a spatiotemporal convolution and global memory unit are introduced at the deepest level of the encoder.

[0054] In practice, the input of the spatial encoder module is a cropped sequence of ERA5 data (500km around the typhoon center + 5° extension):

[0055] is the batch size; , indicating that the input is the past 8 steps (24 hours); is the number of variable channels (eigenvalues) , indicating the spatial coverage (500km radius area under longitude and latitude coordinates).

[0056] The spatial encoder module consists of a basic convolutional unit and a hierarchical downsampling structure.

[0057] The basic convolutional unit consists of the following sequential operations: 3×3 convolution layer (Conv2d), batch normalization (BatchNorm2d), LeakyReLU activation function, and Dropout layer (p=0.4).

[0058] Among them, 3×3 convolution layer (Conv2d):

[0059] in: is the input tensor, B is the batch size; is the number of input channels; Learnable convolution kernels; is the bias term.

[0060] Batch normalization (BatchNorm2d) normalizes the feature map of each channel:

[0061] in, and is the batch statistic; , is a learnable parameter; is a numerically stable term.

[0062] The LeakyReLU activation function is used to improve the nonlinear expression ability of small gradient intervals, where the LeakyReLU activation function is:

[0063] in, is the weight matrix, is the input vector, is the bias vector, is the output vector, It represents the activation function (Leaky Rectified Linear Unit), which has the advantage of alleviating the gradient vanishing problem, especially for sparse heavy precipitation events (such as typhoon center) that are common in precipitation forecasting.

[0064] Dropout is introduced to reduce the risk of overfitting, where Dropout , randomly shielding 40% of neurons to prevent overfitting, the dropout layer randomly "discards" the output results of neurons in the fully connected layer according to a certain probability, that is, only the output results of some neurons in the fully connected layer are considered during training. The introduction of the dropout layer helps make the network easier to train and better extract features.

[0065] Hierarchical downsampling structure: There are four levels of skip connections between the spatial encoder and the decoder, corresponding to feature fusion at different scales, see Table 1.

[0066] Table 1 Four-level jump connection table

[0067] Skip connections are a key U-Net technology. By directly connecting feature maps from the encoder to those from the decoder, they help the decoder recover more detailed information. This concatenation operation combines low-level features (such as edges and shapes) with high-level features (such as semantic information), enhancing the details of the segmentation results. Skip connections effectively combine low-level features with high-level features, avoiding the loss of high-level features. They also help alleviate the vanishing gradient problem and promote efficient network training.

[0068] The temporal feature extraction module includes spatiotemporal dynamic convolution and global memory unit.

[0069] Spatiotemporal dynamic convolution refers to the application of temporal convolution and global average pooling on the deepest spatial features to enhance the model's understanding of temporal evolution characteristics.

[0070] Use 3×3 convolution to integrate temporal dynamic features:

[0071] in This is a two-dimensional convolution operation, using a 3×3 convolution kernel to extract temporal features. z is the feature tensor obtained through the 2D convolution (Conv2D) operation; R is the value of the tensor in real space; B is the batch size.

[0072] The global memory unit uses average pooling to summarize the spatiotemporal dynamics into a single global vector:

[0073] Where X is the output feature of spatiotemporal dynamic convolution, is the index of the time step, are the horizontal and vertical coordinates of the spatial position respectively, and V is the compressed global eigenvector, which represents the overall state of the typhoon system.

[0074] UpBlock of the spatial decoder module:

[0075] Used to study the anisotropic expansion pattern of typhoon precipitation.

[0076] The feature fusion strategy of the spatial decoder module uses skip connections to concatenate encoder features and decoder features layer by layer, preserving: 1. Low-level features: details of short-term precipitation mutations (such as heavy rainfall in the eyewall); 2. High-level features: the overall movement trend of the typhoon.

[0077] Through adaptive average pooling, the entire image is compressed into a 1×1 space, providing a global representation of the typhoon's evolution over the past eight hours. This model also mimics the "state memory mechanism" of a typhoon's path, such as the movement of the storm center and the shifting of precipitation areas. It also learns the "temporal dynamics" of precipitation formation and accumulation, such as persistent and explosive precipitation.

[0078] Multi-time-step precipitation prediction module (output layer): Finally, a 1×1 convolution is used to predict the future Precipitation map:

[0079] Each channel predicts the precipitation field for the next 24 hours.

[0080] It should be noted that the mean square error is used as the loss function to measure the difference between the prediction map of each time step in the multi-step precipitation forecast field and the actual GPM satellite precipitation data:

[0081] in, is the mean square error of the model; is the total number of time steps in the output sequence; is the predicted value of the model at time step t.

[0082] After the model is built, it is trained using the training set. During the training process, the model can also be optimized. By dividing the data set, initially building the model, and setting various parameters as well as the loss function and evaluation indicators, the model is continuously trained until the loss function and various indicators are minimized. Among them, the various parameters of the typhoon precipitation correction model include dynamically adjustable learning rate, number of iterations, batch size, and model architecture parameters. In addition, the model evaluation indicators mainly include absolute deviation (MAE), root mean square error (RMSE), and anomaly correlation coefficient (ACC). It should be noted that it also includes: Calculating the absolute deviation, root mean square error and anomaly correlation coefficient between the multi-step precipitation correction field and the true GPM satellite precipitation data; The absolute deviation is calculated as follows:

[0083] in, is the total number of prediction times, For the The predicted typhoon precipitation value of the time period, For the Observed typhoon precipitation values ​​for a certain period of time.

[0084] The formula for calculating the root mean square error is as follows:

[0085] The parameter definition is the same as above.

[0086] The calculation formula of the abnormal correlation coefficient is as follows:

[0087] in, is the time series average of the predicted typhoon precipitation value, It is the time series average of the observed typhoon precipitation values.

[0088] The performance of the northwest Pacific typhoon precipitation correction model is comprehensively evaluated based on the absolute deviation, the root mean square error and the anomaly correlation coefficient; and the model parameters are optimized and adjusted based on the evaluation results.

[0089] The optimal parameters and the optimal model are obtained through continuous cycle training and model parameter adjustment.

[0090] Finally, the trained convolutional neural network space and long short-term memory network temporal precipitation correction models are used to generate precipitation correction results for the target typhoon based on the ERA5 reanalysis precipitation data.

[0091] The technical solution presented in this invention effectively solves the systematic bias problem of traditional numerical forecast models in typhoon precipitation forecasts by fusing ERA5 reanalysis data with GPM high-resolution satellite precipitation data and combining it with the powerful nonlinear feature extraction capabilities of deep learning models. Using an improved UNet and spatiotemporal architecture, the downscaling conversion from 25km ERA5 data to 10km high-resolution precipitation fields is achieved. The model retains the fine spatial structure of the precipitation field through skip connections and multi-scale feature fusion, improving the accuracy of typhoon precipitation forecasts in the northwest Pacific. At the same time, the innovative introduction of a channel attention mechanism and a spatiotemporal feature fusion module enables the model to not only learn statistical features but also capture the physical evolution laws of the typhoon precipitation system.

[0092] It can be understood that the same or similar parts of the above embodiments can be referenced to each other, and the contents not described in detail in some embodiments can refer to the same or similar contents in other embodiments.

[0093] It should be noted that, in the description of the present invention, the terms "first", "second", etc. are used for descriptive purposes only and should not be understood as indicating or implying relative importance. In addition, in the description of the present invention, unless otherwise specified, the meaning of "plurality" is at least two.

[0094] Any process or method description in a flowchart or otherwise described herein may be understood to represent a module, segment or portion of code comprising one or more executable instructions for implementing the steps of a specific logical function or process, and the scope of the preferred embodiments of the present invention includes alternative implementations in which functions may be performed out of the order shown or discussed, including performing functions in a substantially simultaneous manner or in the reverse order depending on the functions involved, which should be understood by those skilled in the art to which the embodiments of the present invention pertain.

[0095] It should be understood that various components of the present invention may be implemented using hardware, software, firmware, or a combination thereof. In the above-described embodiments, multiple steps or methods may be implemented using software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented using hardware, as in another embodiment, any one of the following technologies known in the art or a combination thereof may be used: a discrete logic circuit having logic gate circuits for implementing logic functions on data signals, an application-specific integrated circuit having suitable combinational logic gate circuits, a programmable gate array (PGA), a field-programmable gate array (FPGA), etc.

[0096] Those skilled in the art will understand that all or part of the steps in the method of the above embodiment can be completed by instructing related hardware through a program, and the program can be stored in a computer-readable storage medium. When the program is executed, it includes one or a combination of the steps of the method embodiment.

[0097] In addition, the functional units in the various embodiments of the present invention may be integrated into a single processing module, or each unit may exist physically separately, or two or more units may be integrated into a single module. The aforementioned integrated modules may be implemented in the form of hardware or in the form of software functional modules. If the integrated modules are implemented in the form of software functional modules and sold or used as independent products, they may also be stored in a computer-readable storage medium.

[0098] The storage medium mentioned above can be a read-only memory, a magnetic disk or an optical disk, etc.

[0099] Throughout this specification, reference to terms such as "one embodiment," "some embodiments," "examples," "specific examples," or "some examples" means that a specific feature, structure, material, or characteristic described in conjunction with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, schematic representations of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in any one or more embodiments or examples.

[0100] Although the embodiments of the present invention have been shown and described above, it will be understood that the above embodiments are illustrative and are not to be construed as limitations on the present invention. A person skilled in the art may change, modify, replace and modify the above embodiments within the scope of the present invention.

Claims

1. A method for correcting typhoon precipitation in the Northwest Pacific based on a deep learning model, characterized in that: include: Obtain ERA5 reanalysis precipitation data, GPM satellite precipitation data, and tropical cyclone best track datasets over the Northwest Pacific; The active period and impact range of each typhoon in history are obtained based on the tropical cyclone best track dataset, and precipitation data corresponding to each typhoon is extracted from the ERA5 reanalysis precipitation data and the GPM satellite precipitation data based on the active period and impact range; Preprocess the precipitation data corresponding to each typhoon to obtain the data set for training the model; A deep learning-based typhoon precipitation correction model for the Northwest Pacific was constructed. The ERA5 reanalysis precipitation data in the dataset was used as training data, and the GPM satellite precipitation data in the dataset was used as label data. The Northwest Pacific typhoon precipitation correction model was trained according to the set parameters and loss function. The ERA5 reanalysis precipitation data of the target typhoon is obtained and input into the trained northwest Pacific typhoon precipitation correction model to obtain the precipitation correction result of the target typhoon.

2. The method according to claim 1, characterized in that Construct a deep learning-based typhoon precipitation correction model for the Northwest Pacific, including: Input layer, spatial encoder module, temporal feature extraction module, spatial decoder module and multi-time-step precipitation prediction module; The input layer receives the ERA5 reanalysis precipitation data and converts it into a tensor format suitable for neural network processing; The spatial encoder module extracts multi-scale spatial features from the data output by the input layer through a four-level downsampling structure; The temporal feature extraction module introduces a bidirectional spatiotemporal module in the deepest layer of the encoder, reorganizes the multi-scale spatial features into a temporal sequence for time dependency modeling, and realizes the fusion of spatiotemporal features through linear projection; The spatial decoder module adopts an upsampling structure symmetrical to the encoder and integrates spatial features of different scales through skip connections to obtain decoding features; The multi-time-step precipitation prediction module performs scale adjustment and channel transformation on the decoded features, and finally outputs a multi-step precipitation correction field.

3. The method according to claim 2, characterized in that The spatial encoder module has a four-level downsampling structure, each level of which includes convolution operations, batch normalization, and nonlinear activation.

4. The method according to claim 2, characterized in that Also includes: The mean square error is used as the loss function to measure the difference between the corrected map at each time step in the multi-step precipitation correction field and the true GPM satellite precipitation data.

5. The method according to claim 2, characterized in that Training of the Northwest Pacific typhoon precipitation correction model also includes: Calculating the absolute deviation, root mean square error and anomaly correlation coefficient between the multi-step precipitation correction field and the true GPM satellite precipitation data; Performing a comprehensive evaluation of the performance of the Northwest Pacific typhoon precipitation correction model based on the absolute deviation, the root mean square error, and the anomaly correlation coefficient; Based on the evaluation results, the model parameters are optimized and adjusted.

6. The method according to claim 1, characterized in that The impact range of each typhoon in history is derived from the tropical cyclone best track dataset, including: The central longitude and latitude coordinates of each typhoon in history are obtained based on the tropical cyclone best track dataset; The typhoon's impact range is defined as a circular area with a radius of 500 km, centered on the typhoon's central latitude and longitude coordinates.

7. The method according to claim 1, characterized in that Preprocess the precipitation data corresponding to each typhoon, including: The sliding window technique is used to construct the correspondence between the 24-hour input sequence and the 24-hour forecast target, with a sliding step of 3 hours. The improved Z-score method is used to detect anomalies. If an outlier is identified, it is replaced by the temporal and spatial neighborhood mean to obtain the ERA5 and GPM datasets corresponding to the typhoon.

8. The method according to claim 7, characterized in that Also includes: The data in the ERA5 dataset and the GPM dataset are standardized using the following formulas: Among them, x is the original value of the data, is the mean of the data set, is the standard deviation of the data set, is the value after data standardization.

Images