Typhoon track and intensity prediction method based on dynamic bias correction and multi-source data fusion

Abstract

The application relates to a typhoon path and intensity prediction method based on dynamic deviation correction and multi-source data fusion, which comprises the following steps: acquiring multi-source typhoon prediction product data and real-time meteorological observation data, and constructing a typhoon path and intensity sample data set; constructing a typhoon path and intensity dynamic correction prediction model, training the prediction model based on the sample data set and a loss function, and acquiring a trained typhoon path and intensity dynamic correction prediction model; acquiring real-time multi-source typhoon prediction products and real-time meteorological observation data, inputting the trained typhoon path and intensity dynamic correction prediction model, and generating typhoon path and intensity prediction products in a time period. The application can significantly improve the accuracy and reliability of typhoon path and intensity prediction, and provides high-precision technical support for typhoon disaster prevention and reduction.

Description

Technical Field

[0001] This invention relates to the field of meteorological forecasting technology, and in particular to a method for forecasting typhoon tracks and intensities based on dynamic bias correction and multi-source data fusion. Background Technology

[0002] Typhoons are among the most severe natural disasters in the world, with their accompanying strong winds, torrential rains, and storm surges posing a significant threat to people's lives and property, as well as to socio-economic development. Accurate typhoon forecasts, especially precise predictions of typhoon paths and intensity, are crucial for disaster prevention and mitigation decision-making.

[0003] Currently, typhoon forecasting mainly relies on numerical weather prediction models, such as the global models of the European Centre for Medium-Range Weather Prediction (ECMWF) and the U.S. National Center for Environmental Prediction (NCEP). While these models can provide general trends in typhoon movement, significant forecast biases remain regarding sudden changes in typhoon tracks and rapid changes in intensity (such as sharp intensification or weakening). Furthermore, systematic differences exist between different numerical models. Effectively integrating the advantages of multiple models and dynamically correcting their biases remains a challenge in current operational forecasting.

[0004] Traditional bias correction methods often rely on statistical post-processing, such as ensemble averages or regression models. These methods are typically based on linear or weakly nonlinear assumptions, limiting their ability to capture complex nonlinear processes and lacking adaptive adjustment mechanisms for forecast errors over time. With the development of artificial intelligence, deep learning models have been attempted for weather forecasting, but current applications mostly focus on improving single elements (such as typhoon track or intensity), and an integrated forecasting system capable of synergistically optimizing typhoon track and intensity with dynamic error feedback capabilities has not yet been formed.

[0005] Therefore, developing a novel neural network architecture that can deeply integrate multi-source observation and forecast data, introduce a dynamic bias correction mechanism, and utilize stronger interpretability and nonlinear fitting capabilities is of great significance for breaking through the current bottlenecks in typhoon path and intensity forecasting and improving the forecasting capability for extreme typhoon events. Summary of the Invention

[0006] The purpose of this invention is to provide a method for forecasting typhoon tracks and intensities based on dynamic bias correction and multi-source data fusion, which solves the problems of insufficient accuracy, lack of dynamic adaptive bias correction capability and physical consistency in the forecasting of typhoon tracks and intensities in existing technologies, so as to achieve high-precision and physically reasonable integrated forecasts of typhoon tracks and intensities.

[0007] To achieve the above objectives, the present invention provides the following solution: Typhoon track and intensity forecasting methods based on dynamic bias correction and multi-source data fusion include: Acquire multi-source typhoon forecast product data and real-time meteorological observation data to construct a sample dataset of typhoon paths and intensities; A dynamic correction forecast model for typhoon track and intensity is constructed. The forecast model is trained based on the sample dataset and loss function to obtain the trained dynamic correction forecast model for typhoon track and intensity. The dynamic correction forecast model for typhoon track and intensity adopts a multi-layer KAN progressive structure and dynamically corrects the future typhoon track and intensity forecast through an error feedback mechanism. Acquire real-time multi-source typhoon forecast products and actual meteorological observation data, input the trained typhoon track and intensity dynamic correction forecast model, and generate typhoon track and intensity forecast products for the time period.

[0008] Optionally, the multi-source typhoon forecast product data includes typhoon track and intensity forecasts from the EC Typhoon Ensemble Model and the NCEP Model, and the EC Typhoon Ensemble Forecast Products are sorted by error size, and forecast factors are constructed using the mean values ​​of different percentiles. The real-time meteorological observation data includes real-time typhoon center location data and wind speed data within the preset range of the typhoon center.

[0009] Optionally, constructing a sample dataset of typhoon paths and intensities includes: Based on the multi-source typhoon forecast product data and the real-time meteorological observation data, for each reporting time, the forecast field data of the numerical model closest to the reporting time is selected, and the model's pre-report data from the model's reporting time to the typhoon's reporting time, the future model forecast data at the typhoon's reporting time, and the real-time information corresponding to the model's pre-report data are obtained. Based on the model's early forecast data, future model forecast data, the actual situation information, and the actual situation labels corresponding to the future model forecast data, the typhoon path and intensity sample dataset is constructed. The actual situation information includes path data and intensity data. The path data is the longitude and latitude sequence of the typhoon center, and the intensity data is the maximum wind speed and minimum air pressure sequence near the typhoon center.

[0010] Optionally, training the prediction model based on the sample dataset and the loss function includes: The model's early forecast data is input into the first-layer KAN network for processing, and the early correction field data is output. The loss function value is calculated based on the early correction field data and the real-world information. Update the parameters of the first-layer KAN network, input the future model prediction data into the parameter-updated first-layer KAN network, and output preliminary correction sequence data; The future model forecast data and the preliminary correction sequence data are input into the second-layer KAN network, which outputs secondary correction sequence data and calculates the loss function value based on the secondary correction sequence data and the real-world label. Calculate the real-time error index based on the early forecast data of the model and the corresponding real-time information; The future model forecast data, the real-time error index, and the secondary correction sequence data are input into the third-layer KAN network to output typhoon track and typhoon intensity correction forecast products. The loss function value is calculated based on the typhoon track and typhoon intensity correction forecast products and the real-time labels.

[0011] Optionally, the loss function includes: a typhoon path error loss function and a typhoon intensity error loss function, wherein the typhoon path error loss function adopts a weighted Haversine distance loss, and the typhoon intensity error loss function adopts a combination of weighted mean square error and absolute error.

[0012] Optionally, the typhoon path error loss function for: ; The typhoon intensity error loss function for: ; Where N is the total number of samples, Let be the weight coefficient of the i-th sample. This is the actual latitude and longitude. The latitude and longitude output by the model. This is the actual intensity of the typhoon. The intensity output by the model.

[0013] Optionally, training the forecast model may also include: training with the Adam optimizer, and enabling an early stopping strategy and saving the optimal model weights when the validation set loss no longer decreases for a consecutive target number of periods.

[0014] The present invention also provides an electronic device, comprising: a memory for storing computer software programs; and a processor for executing the computer programs stored in the memory, thereby realizing a method for forecasting typhoon paths and typhoon intensity based on dynamic deviation correction and multi-source data fusion.

[0015] The beneficial effects of this invention are as follows: Innovative model architecture: For the first time, a three-layer KAN network is introduced into the field of typhoon forecasting. Utilizing its excellent nonlinear fitting ability and interpretability, a progressive dynamic bias correction model is constructed. By feeding back the previous error as a key input to subsequent network layers, adaptive and refined correction of future forecasts is achieved.

[0016] Deeply integrated multi-source data: It not only integrates multi-model forecasts and real-time data, but also innovatively utilizes previous error factors, enabling the model to simultaneously learn the systematic biases of numerical models and the climatological statistical patterns of typhoon activity, significantly improving the stability of forecasts.

[0017] Physics-guided loss function: Integrating prior meteorological knowledge into the model training process ensures that the forecast results are not only statistically accurate, but also conform to the basic dynamic-thermodynamic principles of typhoons.

[0018] Significant improvement in forecast performance: Based on historical typhoon case studies, this method reduces the average error of typhoon track forecast by more than 10% and the error of intensity forecast by more than 15%. It performs particularly well at key points such as typhoon track turning and intensity abrupt changes, with a forecast lead time of 24-72 hours.

[0019] Promising prospects for commercial application: The method proposed in this invention forms a complete technical chain from data preparation and model training to product generation. It can be integrated into existing meteorological operational systems to achieve automated operation, providing high-timeliness and high-precision technical support for typhoon disaster prevention and mitigation, and has the potential to be promoted to forecasts of other high-impact weather systems. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 This is a flowchart of a typhoon track and intensity forecasting method based on dynamic deviation correction and multi-source data fusion, according to an embodiment of the present invention. Figure 2 This is a schematic diagram of the structure of the layered KAN network typhoon path and typhoon intensity correction model according to an embodiment of the present invention; Figure 3 This is a schematic diagram comparing the typhoon intensity correction forecast results with the actual situation at different reporting times according to an embodiment of the present invention. Detailed Implementation

[0022] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0024] like Figure 1 As shown in the figure, this embodiment proposes a method for forecasting typhoon tracks and intensities based on dynamic bias correction and multi-source data fusion, including: Acquire multi-source typhoon forecast product data and real-time meteorological observation data to construct a sample dataset of typhoon paths and intensities; A dynamic correction forecast model for typhoon track and intensity is constructed. The forecast model is trained based on the sample dataset and loss function to obtain the trained dynamic correction forecast model for typhoon track and intensity. The dynamic correction forecast model for typhoon track and intensity adopts a multi-layer KAN progressive structure and dynamically corrects the future typhoon track and intensity forecast through an error feedback mechanism. Acquire real-time multi-source typhoon forecast products and actual meteorological observation data, input the trained typhoon track and intensity dynamic correction forecast model, and generate typhoon track and intensity forecast products for the time period.

[0025] Furthermore, the multi-source typhoon forecast product data includes typhoon track and intensity forecasts from the EC Typhoon Ensemble Model and the NCEP Model, and the EC Typhoon Ensemble Forecast Products are sorted by error size, and forecast factors are constructed using the mean values ​​of different percentiles. The real-time meteorological observation data includes real-time typhoon center location data and wind speed data near the typhoon center.

[0026] Specifically, historical typhoon location and intensity data were obtained from the optimal tropical cyclone track dataset acquired by the Shanghai Typhoon Institute (CMA-STI) of the China Meteorological Administration, while real-time location data was obtained from the China Meteorological Administration. Multi-source forecast product data were acquired from mainstream numerical models such as ECMWF and NCEP-GFS. All data cover the period from January 2016 to December 2024. Error analysis was performed on the EC typhoon ensemble forecast products: forecasts were sorted by error magnitude, and forecast factors were constructed using the means of different percentiles to quantify model uncertainty and enhance data representativeness.

[0027] Furthermore, the construction of a sample dataset of typhoon paths and intensities includes: Based on the multi-source typhoon forecast product data and the real-time meteorological observation data, for each reporting time, the forecast field data of the numerical model closest to the reporting time is selected, and the model's pre-report data from the model's reporting time to the typhoon's reporting time, the future model forecast data at the typhoon's reporting time, and the real-time information corresponding to the model's pre-report data are obtained. Based on the model's early forecast data, future model forecast data, the actual situation information, and the actual situation labels corresponding to the future model forecast data, the typhoon path and intensity sample dataset is constructed. The actual situation information includes path data and intensity data. The path data is the longitude and latitude sequence of the typhoon center, and the intensity data is the maximum wind speed and minimum air pressure sequence near the typhoon center.

[0028] Specifically, the reanalysis data and model forecast data used cover the period from January 2016 to December 2024. Within this period, all numbered typhoons in the Northwest Pacific were selected, totaling over 222 typhoons. Reporting was conducted at 3-hour intervals (02:00, 05:00, 08:00, 11:00, 14:00, 17:00, 20:00, and 23:00 daily). For each reporting time, the specific steps included: S11: Obtain the numerical model forecast field data closest to the current reporting time; S12: Extract the model's early forecast sequence (model early forecast data) from the model's start time to the current typhoon start time, denoted as X_hit, and the future model forecast sequence for a set time period (such as the next 24-72 hours) from the current typhoon start time, denoted as X_fut; S13: Obtain the typhoon observation information corresponding to the early forecast sequence X_hit of the model, denoted as O_T, and the future typhoon observation label corresponding to the future forecast sequence X_fut of the model, denoted as Y; S14: Perform spatiotemporal alignment processing on the above data to construct a training set and a test set containing multiple reporting time samples. Each sample includes input data (X_hit, X_fut, O_T) and the corresponding real-time label Y.

[0029] Furthermore, training the prediction model based on the aforementioned sample dataset and loss function includes: The model's early forecast data is input into the first-layer KAN network for processing, and the early correction field data is output. The loss function value is calculated based on the early correction field data and the real-world information. Update the parameters of the first-layer KAN network, input the future model prediction data into the parameter-updated first-layer KAN network, and output preliminary correction sequence data; The future model forecast data and the preliminary correction sequence data are input into the second-layer KAN network, which outputs secondary correction sequence data and calculates the loss function value based on the secondary correction sequence data and the real-world label. Calculate the real-time error index based on the early forecast data of the model and the corresponding real-time information; The future model forecast data, the real-time error index, and the secondary correction sequence data are input into the third-layer KAN network to output typhoon track and typhoon intensity correction forecast products. The loss function value is calculated based on the typhoon track and typhoon intensity correction forecast products and the real-time labels.

[0030] Specifically, such as Figure 2 As shown, the training process of the model specifically includes: S21: Training the first layer of the KAN network (Model_KAN_1): The model's early forecast sequence X_hit is input into the first-layer KAN network, which outputs the early correction field O1. The first loss function value loss_1(O1, O_T) between the early correction field O1 and the typhoon observation information O_T is calculated, and the model parameters of the first-layer KAN network are updated based on the loss_1.

[0031] S22: Training the second-layer KAN network (Model_KAN_2): The future model forecast sequence X_fut is input into the first-layer KAN network after parameter updates, and the preliminary correction sequence O1_fut for the future time period is output. The future model forecast sequence X_fut and the preliminary correction sequence O1_fut are used together as input to the second-layer KAN network, and the secondary correction sequence O2_fut is output. The second loss function value loss_2(O2_fut, Y) between the secondary correction sequence O2_fut and the future typhoon real-time label Y is calculated, and the model parameters of the second-layer KAN network are updated based on loss_2.

[0032] S23: Training the third-layer KAN network (Model_KAN_3): Calculate the real-time error index E, which characterizes the deviation between the model's early forecast sequence X_hit and the actual typhoon observation information O_T. The calculation method is as follows: E = |X_hit - O_T| or E = X_{hit} - O_T; The secondary correction sequence O2_fut, the real-time error index E, and the future model forecast sequence X_fut are used as inputs to the third-layer KAN network to output the final typhoon track and intensity correction forecast product O3_fut. The third loss function value loss_3(O3_fut, Y) between the forecast product O3_fut and the future typhoon real-time label Y is calculated, and the model parameters of the third-layer KAN network are updated based on loss_3.

[0033] Real-time forecasting based on the trained progressive typhoon forecast correction model includes: At the actual forecast time T, the latest model pre-forecast sequence, future model forecast sequence, and actual observation information are obtained. According to the forward propagation calculation logic in step S2 above, the data are processed sequentially through the first, second, and third layers of the KAN network. The third layer KAN network outputs the final typhoon path and typhoon intensity correction results within the future set timeframe of the forecast time T.

[0034] Furthermore, the loss function includes: a typhoon path error loss function and a typhoon intensity error loss function, wherein the typhoon path error loss function adopts a weighted Haversine distance loss, the typhoon intensity error loss function adopts a combination of weighted mean square error and absolute error, and the weight coefficient in the typhoon path error loss function increases with the forecast lead time.

[0035] Furthermore, when constructing a path model, loss_1, loss_2, and loss_3 all use the path error loss function. : ; When constructing an intensity model, the intensity error loss function is always used. : ; Where N is the total number of samples, Let be the weight coefficient of the i-th sample. This is the actual latitude and longitude. The latitude and longitude output by the model. This is the actual intensity of the typhoon. The intensity output by the model; Furthermore, training the forecast model also includes: training with the Adam optimizer, and when the validation set loss no longer decreases for a consecutive target number of cycles, enabling an early stopping strategy and saving the optimal model weights.

[0036] Specifically, the trained KAN typhoon track and typhoon intensity correction model are integrated to form a complete typhoon track and intensity forecasting system. Training uses the Adam optimizer with an initial learning rate of 1e-4, a batch size of 8, and 100 training epochs. When the validation set loss no longer decreases for 10 consecutive epochs, an early stopping strategy is implemented, and the optimal model weights are saved.

[0037] By acquiring three data points—X_hit, O_T, and X_fut—in real time and inputting them into the trained ensemble model, corrected forecast products for the typhoon's path (latitude and longitude sequence) and intensity (minimum pressure near the center and maximum wind speed) for the next 24, 48, and 72 hours can be obtained directly.

[0038] This embodiment also provides an electronic device, including: a memory for storing computer software programs; and a processor for executing the computer programs stored in the memory, thereby realizing a typhoon track and intensity forecasting method based on dynamic deviation correction and multi-source data fusion.

[0039] In practical applications, the electronic equipment can be integrated into meteorological operational servers or high-performance computing clusters to receive real-time multi-source observation and model forecast data, automatically run forecast processes, and output high-resolution typhoon integrated forecast products and extreme risk warning information. The system is automatically triggered four times daily (corresponding to the numerical model release cycle) by scheduled tasks (such as cron jobs) to achieve full operational process execution.

[0040] Example 1: The following example uses the forecast for Typhoon No. 13 in 2024 as an illustration: (1) Data collection. The data collection period is from January 2016 to December 2024. The samples from January 2016 to December 2023 are used to train the forecast model; the samples from January to August 2024 are used to test the forecast model; and the samples from September to December are used for validation. The hourly times every 3 hours within this period (02:00, 05:00, 08:00, 11:00, 14:00, 17:00, 20:00, and 23:00 daily) are used as the start time for the 72-hour typhoon forecast. The collected data are: CMA-STI optimal track data every 3 hours (6 hours in earlier years), ECMWF (Typhoon Ensemble Model), and 72-hour forecast products with corresponding start times from the NCEP-GFS model.

[0041] (2) Constructing a typhoon sample dataset. A total of 2039 valid samples were constructed. Strict spatiotemporal matching was performed on multiple data sources for each sample, and error ranking and percentile mean factor of EC typhoon ensemble forecasts were introduced.

[0042] (3) Training path and intensity correction model. A three-layer KAN network structure model was used. The parameters of the three layers of KAN are as follows: the number of nodes in each layer of the first KAN network (the number of the first and last nodes are the number of input factors and the number of predicted elements, respectively) are 10, 20, 40, 20, 10, 5, N1 (the path is N1=2, the intensity is N1=1); the number of nodes in each layer of the second KAN network (the number of the first and last nodes are the number of input factors and the number of predicted elements, respectively) are 12, 24, 48, 24, 12, 6, N2 (the path is N2=2, the intensity is N2=1); the number of nodes in each layer of the second KAN network (the number of the first and last nodes are the number of input factors and the number of predicted elements, respectively) are 16, 36, 72, 36, 18, 9, N3 (the path is N3=2, the intensity is N3=1); The aforementioned weighted mean squared error loss function was used during training. After training, on the 2024 independent test set, the average distance error of the 24-hour track forecast was 64.9 km, a reduction of 18.7% compared to the original ECMWF model (average error 73.7 km); the average error of the maximum wind speed forecast near the center was 4.7 m / s, a reduction of 22.9% compared to the original model (average error 6.1 m / s).

[0043] (4) Calculate the latest forecast sample. For the typhoon forecast for the next 24-72 hours from 08:00 Beijing time on September 14, 2024, take this time as the starting time and denote it as T. First, determine that the numerical model starting time closest to this starting time is 20:00 Beijing time on September 13, 2024. Second, obtain the ECMWF (Typhoon Ensemble Model) and N starting at 20:00 on September 13, 2024. The CEP-GFS model provides 3-hourly forecast values ​​X_hit for all its members from 20:00 on September 13, 2024 to 08:00 on September 14, 2024, along with the actual situation O_T (latitude, longitude, and intensity information of the typhoon). Finally, it obtains 3-hourly forecast values ​​X_fut for all its members from 11:00 on September 14, 2024 to 08:00 on September 17, 2024.

[0044] (5) Model inference generates corrected forecast products: The input data prepared in step (4) is input into the trained KAN typhoon track and intensity correction model. That is, for track forecasting, the input data about track prepared in step (4) is input into the trained KAN typhoon track forecasting model; and for intensity forecasting, the input data about intensity prepared in step (4) is input into the trained KAN typhoon intensity forecasting model. Finally, the corrected typhoon center latitude and longitude forecast information and typhoon intensity correction forecast products are obtained every 3 hours. Figure 3 The diagram shows a comparison between the revised typhoon intensity forecasts and the actual situation for a specific typhoon at different reporting times.

[0045] (6) Generation of comprehensive typhoon forecast products. Based on the forecast of the typhoon path and intensity for the next 72 hours, comprehensive typhoon forecast products are generated.

[0046] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A method for forecasting typhoon track and intensity based on dynamic bias correction and multi-source data fusion, characterized in that, include: Acquire multi-source typhoon forecast product data and real-time meteorological observation data to construct a sample dataset of typhoon paths and intensities; A dynamic correction forecast model for typhoon track and intensity is constructed. The forecast model is trained based on the sample dataset and loss function to obtain the trained dynamic correction forecast model for typhoon track and intensity. The dynamic correction forecast model for typhoon track and intensity adopts a multi-layer KAN progressive structure and dynamically corrects the future typhoon track and intensity forecast through an error feedback mechanism. Acquire real-time multi-source typhoon forecast products and actual meteorological observation data, input the trained typhoon track and intensity dynamic correction forecast model, and generate typhoon track and intensity forecast products for the time period.

2. The typhoon track and intensity forecasting method based on dynamic deviation correction and multi-source data fusion according to claim 1, characterized in that, The multi-source typhoon forecast product data includes typhoon track and intensity forecasts from the EC Typhoon Ensemble Model and the NCEP Model. The EC Typhoon Ensemble Forecast Products are sorted by error size, and forecast factors are constructed using the mean values ​​of different percentiles. The real-time meteorological observation data includes real-time typhoon center location data and wind speed data within the preset range of the typhoon center.

3. The typhoon track and intensity forecasting method based on dynamic deviation correction and multi-source data fusion according to claim 1, characterized in that, The typhoon track and intensity sample dataset includes: Based on the multi-source typhoon forecast product data and the real-time meteorological observation data, for each reporting time, the forecast field data of the numerical model closest to the reporting time is selected, and the model's pre-report data from the model's reporting time to the typhoon's reporting time, the future model forecast data at the typhoon's reporting time, and the real-time information corresponding to the model's pre-report data are obtained. Based on the model's early forecast data, future model forecast data, the actual situation information, and the actual situation labels corresponding to the future model forecast data, the typhoon path and intensity sample dataset is constructed. The actual situation information includes path data and intensity data. The path data is the longitude and latitude sequence of the typhoon center, and the intensity data is the maximum wind speed and minimum air pressure sequence near the typhoon center.

4. The typhoon track and intensity forecasting method based on dynamic deviation correction and multi-source data fusion according to claim 3, characterized in that, Training the prediction model based on the aforementioned sample dataset and loss function includes: The model's early forecast data is input into the first-layer KAN network for processing, and the early correction field data is output. The loss function value is calculated based on the early correction field data and the real-world information. Update the parameters of the first-layer KAN network, input the future model prediction data into the parameter-updated first-layer KAN network, and output preliminary correction sequence data; The future model forecast data and the preliminary correction sequence data are input into the second-layer KAN network, which outputs secondary correction sequence data and calculates the loss function value based on the secondary correction sequence data and the real-world label. Calculate the real-time error index based on the early forecast data of the model and the corresponding real-time information; The future model forecast data, the real-time error index, and the secondary correction sequence data are input into the third-layer KAN network to output typhoon track and typhoon intensity correction forecast products. The loss function value is calculated based on the typhoon track and typhoon intensity correction forecast products and the real-time labels.

5. The typhoon track and intensity forecasting method based on dynamic deviation correction and multi-source data fusion according to claim 1, characterized in that, The loss functions include: a typhoon path error loss function and a typhoon intensity error loss function, wherein the typhoon path error loss function adopts a weighted Haversine distance loss, and the typhoon intensity error loss function adopts a combination of weighted mean square error and absolute error.

6. The typhoon track and intensity forecasting method based on dynamic deviation correction and multi-source data fusion according to claim 5, characterized in that, The typhoon path error loss function for: ； The typhoon intensity error loss function for: ； Where N is the total number of samples, Let be the weight coefficient of the i-th sample. This is the actual latitude and longitude. The latitude and longitude output by the model. This is the actual intensity of the typhoon. The intensity output by the model.

7. The typhoon track and intensity forecasting method based on dynamic deviation correction and multi-source data fusion according to claim 1, characterized in that, Training the forecast model also includes: training with the Adam optimizer, and when the validation set loss no longer decreases for a consecutive target number of periods, enabling an early stopping strategy and saving the optimal model weights.

8. An electronic device, characterized in that, include: Memory, used to store computer software programs; A processor is configured to execute a computer program stored in the memory, thereby implementing the typhoon track and intensity forecasting method based on dynamic deviation correction and multi-source data fusion as described in any one of claims 1-7.

Images