A method and system for sub-seasonal prediction of tropical cyclones in the northwest pacific

Abstract

The application discloses a northwest Pacific tropical cyclone sub-season prediction method and system, and relates to the technical field of atmospheric science. The method comprises the following steps: extracting the ACE distance average of each type of TC in each time window from climatological data and taking the ACE distance average as a prediction label of each time window; taking a plurality of large-scale environmental fields used for climatological reanalysis as potential prediction factors, and obtaining a plurality of components of each potential prediction factor; constructing a single-variable prediction model for each component of each type of TC, pre-training and evaluating each single-variable prediction model through the components and the prediction label, and selecting the optimal component combination; constructing a multi-variable prediction model for each type of TC, training the multi-variable prediction model through the component combination and the prediction label, and deploying the prediction after the training; and generating an ACE high-resolution spatial probability map of a target prediction period of the northwest Pacific according to a prediction result. The application can realize accurate tropical cyclone sub-season prediction.

Description

Technical Field

[0001] This invention relates to the field of atmospheric science and technology, and in particular to a method and system for subseasonal forecasting of tropical cyclones in the Northwest Pacific. Background Technology

[0002] Tropical cyclones cause strong winds, torrential rains, giant waves, storm surges, and indirect geological disasters such as landslides and mudslides, often resulting in significant casualties and socio-economic losses. The Northwest Pacific Ocean has the highest frequency of tropical cyclone (including typhoon) formation, accounting for approximately 30% of the world's tropical cyclone formations. Currently, the tropical cyclone / typhoon forecasting systems of major operational forecasting agencies worldwide include 1) short-to-medium-term (within 5 days) high-resolution numerical model forecasts and 2) seasonal and interannual climate predictions. Due to atmospheric noise and systematic errors in numerical models, the accuracy of typhoon forecasts exceeding one week is relatively low. Summary of the Invention

[0003] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method and system for subseasonal forecasting of tropical cyclones in the Northwest Pacific, thereby solving the technical problem of low accuracy in existing forecasting systems.

[0004] To achieve the above objectives, the present invention is implemented using the following technical solution:

[0005] In a first aspect, the present invention provides a method for subseasonal forecasting of tropical cyclones in the Northwest Pacific, comprising:

[0006] Acquire historical climatological data, extract the average ACE distance of each type of TC in each time window from the climatological data, and use it as the forecast label for each time window;

[0007] Multiple large-scale environmental fields used for climate reanalysis are used as potential forecasting factors, and multiple components of each potential forecasting factor are obtained, including the contemporaneous ISO component, the previous observation ISO component, and the previous observation LFBS component.

[0008] For each component of each TC, a univariate prediction model is constructed. Each univariate prediction model is pre-trained and evaluated using the component and the prediction label. The component combination with the most nonlinear prediction value corresponding to each TC is selected.

[0009] A multivariate prediction model is constructed for each type of TC. The multivariate prediction model is trained by combining the components and the forecast label. The trained multivariate prediction model is then used to predict the average ACE distance for each time window of the target forecast period.

[0010] A high-resolution spatial probability map of ACE for the target forecast period in the Northwest Pacific is generated based on the average ACE distance prediction results for each type of TC.

[0011] Optionally, the classification of the TC includes:

[0012] Obtain the optimal path dataset for TCs during the target time period, and extract target TC samples from the optimal path dataset; use c-means fuzzy clustering analysis to classify the target TC samples and generate TC categories;

[0013] The target TC sample is the sum of the squares of the maximum 6-hour central wind speed of all TCs with maximum surface wind speeds greater than the wind speed threshold in each time window of the Northwest Pacific.

[0014] Optionally, the large-scale environmental field includes outward longwave radiation (OLR), 700 hPa specific humidity, 500 hPa vertical velocity, 850 hPa vertical wind shear, 200 hPa vertical wind shear, 850 hPa vorticity, 200 hPa divergence, surface latent heat flux, surface sensible heat flux, and sea surface temperature (SST).

[0015] Optionally, the concurrent ISO component is the average anomaly field of each time window obtained from the large-scale environmental field of the target forecast period extracted from the existing forecast system;

[0016] The ISO component of the previous observation is the average anomaly field of each time window obtained from the large-scale environmental field of the target forecast period 1-n time windows extracted from the existing observation reanalysis dataset.

[0017] The LFBS component of the previous observation is the low-frequency background state of each time scale obtained from the large-scale environmental field of 1-m time scales before the target forecast period extracted from the existing observation reanalysis dataset. The time scale is a k-time window, where k is greater than 1.

[0018] Optionally, the selection of the component combination with the most nonlinear pre-quote value corresponding to each type of TC includes:

[0019] After the univariate prediction model corresponding to each component of each class of TC is pre-trained, the prediction skill evaluation index of the univariate prediction model is calculated; the prediction skill evaluation index includes the time correlation coefficient (TCC), the spatial correlation coefficient (PCC), and the root mean square error (RMSE).

[0020] Based on the prediction technique evaluation index, the optimal combination of multiple concurrent ISO components, multiple previous observation ISO components, and multiple previous observation LFBS components is selected.

[0021] Optionally, the multivariate prediction model includes a feature extraction module, a feature concatenation module, and a cross-scale nonlinear coupling module;

[0022] The feature extraction module uses the feature extraction module in the pre-trained univariate prediction model of each component in the component combination to extract features;

[0023] The feature splicing module is used to splice the feature extraction results of each component in the component combination in the channel dimension to generate spliced ​​features.

[0024] The cross-scale nonlinear coupling module includes two nonlinear convolutional layers, which are used to learn the interaction between components based on the spliced ​​features and generate a deep coupling result.

[0025] Secondly, the present invention provides a subseasonal forecasting system for tropical cyclones in the Northwest Pacific, comprising:

[0026] The label creation module is configured to acquire historical climatological data, extract the average ACE distance of each type of TC in each time window from the climatological data, and use it as the forecast label for each time window.

[0027] The component generation module is configured to use multiple large-scale environmental fields for climate reanalysis as potential forecasting factors and obtain multiple components for each potential forecasting factor, including contemporaneous ISO components, previous observation ISO components, and previous observation LFBS components.

[0028] The component combination module is configured to construct a univariate prediction model for each component of each TC class, pre-train and evaluate each univariate prediction model using the component and the prediction label, and select the component combination with the most nonlinear prediction value for each TC class.

[0029] The model preparation module is configured to build a multivariate prediction model for each type of TC, train the multivariate prediction model by combining the components and the forecast label, and use the trained multivariate prediction model to predict the ACE distance average for each time window of the target forecast period.

[0030] The forecast generation module is configured to generate a high-resolution spatial probability map of ACE for the target forecast period in the Northwest Pacific based on the average ACE distance prediction results for each type of TC.

[0031] Thirdly, the present invention provides an electronic device, including a processor and a storage medium;

[0032] The storage medium is used to store instructions;

[0033] The processor is configured to operate according to the instructions to perform the steps according to the method described above.

[0034] Fourthly, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the above-described method.

[0035] Fifthly, the present invention provides a computer program product, including a computer program / instructions that, when executed by a processor, implement the steps of the above-described method.

[0036] Compared with the prior art, the beneficial effects achieved by the present invention are as follows:

[0037] This invention provides a method and system for subseasonal forecasting of tropical cyclones in the Northwest Pacific. Addressing the technical gap in subseasonal (1-6 week) forecasting, where typhoon synoptic and seasonal forecasting are relatively mature, this invention fully utilizes the ability of large-scale signals at different scales in the Asian monsoon region to regulate the advance and concurrency of tropical cyclone (TC) activity, and leverages the excellent large-scale field simulation capabilities of advanced climate dynamic models (such as S2S ECMWF). Through deep learning algorithms, it mines multi-scale air-sea interaction mechanisms and develops a multivariate prediction model capable of forecasting the extended ACE (acetylene arc distance) of tropical cyclones. This improves the accuracy of the ACE distance-average prediction results for each type of TC, providing accurate data support for the subsequent generation of ACE spatial probability distribution maps. By accurately forecasting the number, frequency, and cumulative cyclone energy (ACE) of different trajectory types of TCs in the Northwest Pacific 1-6 weeks in advance, and generating high-resolution ACE spatial probability distribution maps based on this, seamless and refined forecasting of typhoon activity is achieved, providing extended-term early warning information of significant operational value for disaster prevention and mitigation. Attached Figure Description

[0038] Figure 1 This is a flowchart illustrating the subseasonal forecasting method for tropical cyclones in the Northwest Pacific provided in this embodiment of the invention.

[0039] Figure 2 This is a comparison chart of the time correlation coefficients of a set of potential forecasting factors under the same component, provided by an embodiment of the present invention;

[0040] Figure 3 This is a performance comparison chart between the multivariate prediction model provided in this embodiment of the invention and the existing hybrid-statistical model;

[0041] Figure 4 This is a schematic diagram illustrating the real-time rolling forecast results of the multivariate prediction model provided in this embodiment of the invention for 2024 and 2025. Detailed Implementation

[0042] The present invention will be further described below with reference to the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and should not be used to limit the scope of protection of the present invention.

[0043] Example 1

[0044] like Figure 1 As shown, this embodiment of the invention provides a method for subseasonal forecasting of tropical cyclones in the Northwest Pacific, comprising the following steps:

[0045] Step S1: Obtain historical climatological data, extract the average ACE distance of each type of TC in each time window from the climatological data, and use it as the forecast label for each time window.

[0046] Specifically, in this embodiment, the classification of TC (i.e., Tropical Cyclone) includes:

[0047] Obtain the best path dataset of TCs from the Joint Typhoon Warning Center (JTWC) from 2003 to 2022, and extract target TC samples from the best path dataset;

[0048] Setting the time window to weekly, the target TC sample uses the sum of the squares of the maximum surface wind speeds (MWS) at the center of all TCs in the Northwest Pacific that have a maximum surface wind speed greater than the wind speed threshold (34 knots) over their lifetime, i.e., the accumulated cyclonic energy (ACE). Using this target TC sample comprehensively considers the number, frequency, and intensity of TCs, providing an objective description of the overall destructive potential.

[0049] The c-means fuzzy clustering method is used to classify the target TC samples and generate the TC categories.

[0050] Step S2: Use multiple large-scale environmental fields for climate reanalysis as potential forecasting factors, and obtain multiple components for each potential forecasting factor, including the contemporaneous ISO component, the previous observation ISO component, and the previous observation LFBS component.

[0051] Specifically, in this embodiment, the large-scale environmental field (resolution 1.5°×1.5°) includes outward longwave radiation (OLR), 700 hPa specific humidity, 500 hPa vertical velocity, 850 hPa vertical wind shear, 200 hPa vertical wind shear, 850 hPa vorticity, 200 hPa divergence, surface latent heat flux, surface sensible heat flux, and sea surface temperature (SST).

[0052] If we set the time window to a week, the time scale to a month, n to 6, and m to 3, then:

[0053] (1) The concurrent ISO component is the weekly mean anomaly field obtained from the large-scale environmental field of the target forecast period extracted from the existing forecast system;

[0054] Existing forecasting systems can utilize S2S dynamical model return data, including ECMWF model returns (2003–2022) and 9 years of independent forecasts (2016–2024). The ECMWF model returns have 11 ensemble members; to increase the sample size for modeling, the ensemble average is used as the 12th ensemble member. The ECMWF model reports twice a week (Monday and Thursday), and then integrates over 46 days. The training set covers 2003–2018, with 13,056 returns (16 years × 68 forecasts × 12 ensembles), and the validation set covers 2019–2022, with 3,264 returns (4 years × 68 forecasts × 12 ensembles). The independent forecast phase covers 2016–2024, with 612 forecasts (9 years × 68 forecasts). Modeling directly with model data and TC ACE can avoid the bias in TC ACE caused by the model's prediction bias of the anomaly field.

[0055] To extract the pure sub-seasonal ISO component, the daily large-scale field is subtracted from the model's corresponding lead time and the climatological average of the calendar day. Then, the resulting anomaly field is subjected to a weekly average operation to strictly align it with the spatiotemporal resolution of the TC ACE anomaly, thus obtaining the weekly average large-scale ISO component of the S2S model forecasts at different lead times. Here, the ISO component obtained from the large-scale field of the S2S model is the weekly average relative to the same period of the forecast time.

[0056] (2) The ISO component of the early observation is the weekly average anomaly field obtained from the large-scale environmental field 1-6 weeks before the target forecast period extracted from the existing observation reanalysis dataset;

[0057] (3) The LFBS component of the previous observation is the monthly low-frequency background state obtained from the large-scale environmental field of the target forecast period 1-3 months before the target forecast period extracted from the existing observation reanalysis dataset.

[0058] The existing observational reanalysis dataset can be the ERA5 reanalysis dataset. The daily large-scale fields in the reanalysis data are subtracted from the calendar-day climatological mean, and then the resulting anomaly field is averaged weekly to obtain the weekly average large-scale ISO component, which is 1–6 weeks ahead of the time resolution of the TC ACE anomaly at each forecast time. Here, the large-scale ISO component in the early observation period is the weekly average component 1–6 weeks ahead of the forecast time. The daily large-scale fields in the reanalysis data are subtracted from the calendar-day climatological mean, and then the resulting anomaly field is averaged monthly to 3 months ahead of the forecast time. Here, the large-scale LFBS component in the early observation period is the monthly average component 1–3 months ahead of the forecast time.

[0059] Based on the modulation mechanism of different LFBS to subseasonal fluctuations, instead of performing a blind search across the entire region, the feature candidate region is dynamically locked in the Eurasian continent and the key modulating sea area in the tropics (2.5°S–60.0°N, 40°E–160°W) to eliminate spurious correlations at the physical level.

[0060] Step S3: Construct a univariate prediction model for each component of each TC class. Pre-train and evaluate each univariate prediction model using the components and forecast labels, and select the component combination with the most nonlinear forecast value for each TC class.

[0061] By modeling each type of TC separately, physical signal interference caused by uniform modeling of the entire ocean basin is effectively avoided.

[0062] As described above, the components include one concurrent ISO component, six previous observation ISO components, and three previous observation LFBS components, along with 10 potential forecast factors. A univariate prediction model is constructed for each scenario, totaling 10 × 10 = 100, as shown in Table 1.

[0063] Table 1: Construction of Univariate Prediction Models

[0064]

[0065] The component combinations with the most nonlinear predicted price values ​​for each type of TC include:

[0066] After the univariate prediction model corresponding to each component of each class of TC is pre-trained, the prediction skill evaluation index of the univariate prediction model is calculated. The prediction skill evaluation index includes the time correlation coefficient (TCC), the spatial correlation coefficient (PCC), and the root mean square error (RMSE). The comprehensive prediction skill evaluation index can be obtained by weighting the time correlation coefficient (TCC), the spatial correlation coefficient (PCC), and the root mean square error (RMSE).

[0067] like Figure 2As shown, the time correlation coefficients (TCCs) of a set of potential forecast factors (W500, vws, Q700, Div200, Vort850, LHF, SHF, and OLR) under the same component are provided. Figure 2 In the diagram, C1 to C7 represent the seven categories of TC, and TCall represents all TCs, which are unified into one category.

[0068] Based on the evaluation index of prediction techniques, the optimal combination of multiple concurrent ISO components, multiple previously observed ISO components, and multiple previously observed LFBS components is selected to form a component combination. For example, four concurrent ISO components, three previously observed LFBS components, and three previously observed LFBS components are selected as the component combination of the current category TC.

[0069] Step S4: Construct a multivariate prediction model for each type of TC. Train the multivariate prediction model by combining components and forecast labels. Use the trained multivariate prediction model to predict the average ACE distance for each time window of the target forecast period.

[0070] The multivariate prediction model includes a feature extraction module, a feature concatenation module, and a cross-scale nonlinear coupling module;

[0071] The feature extraction module uses the feature extraction module in the pre-trained univariate prediction model of each component in the component combination to extract features;

[0072] The feature splicing module is used to splice the feature extraction results of each component in the component combination along the channel dimension to generate spliced ​​features;

[0073] The cross-scale nonlinear coupling module consists of two nonlinear convolutional layers, which are used to learn the interaction between components based on the spliced ​​features to generate a deep coupling result.

[0074] Both the univariate and multivariate prediction models are built upon CNN networks. This invention extracts independent high-dimensional hidden features by calling the feature extraction module in the pre-trained univariate prediction model. After concatenating the multi-source extracted high-dimensional hidden features along the channel dimension, two consecutive 3×3 depthwise convolution operations are performed, forcing the network to learn the nonlinear interactions between different forecast factors and different scales (such as the sub-seasonal signal ISO and the low-frequency background state LFBS) in the latent space (e.g., the amplification effect of the background state on the sub-seasonal wave train amplitude). This allows the model to successfully internalize the multi-scale modulation physical processes that are difficult to accurately simulate using simple dynamical models.

[0075] The training process of a multivariate prediction model can incorporate a smoothed L1 loss function (Huber Loss, with a smoothing parameter β = 0.5). This design exhibits L1 norm in the extreme large value region to suppress gradient explosion, and L2 norm near 0 to ensure fine convergence, significantly improving the model's robustness. This is further enhanced by the AdamW optimizer (initial learning rate 1 × 10⁻⁻⁴). 4 A 50% dropout rate and an early stopping strategy are used to ensure model stability and prevent overfitting.

[0076] By innovatively introducing the Smooth L1 Loss function, this invention solves the gradient oscillation problem caused by the large number of extreme values ​​and zero values ​​in typhoon data. While ensuring accurate grasp of seasonal trends, the model avoids weight deviations due to isolated extreme typhoon cases, thus ensuring stability and reliability in real-time business operations.

[0077] like Figure 3 As shown, the performance comparison between the multivariate prediction model used in this invention and the existing hybrid-statistical model shows that this invention performs better in all forecast lead times (1-6 weeks), especially in terms of improvement of 1-2 weeks in advance. Figure 3 (a) represents the time correlation coefficient of all ACEs in the Northwest Pacific to the mean, (b) represents the time correlation coefficient of the sum of ACEs in categories C1–C7 to the mean, (c) represents the time correlation coefficient of all ACEs in the Northwest Pacific to the mean, and (d) represents the time correlation coefficient of the sum of ACEs in categories C1–C7.

[0078] like Figure 4 As shown, the multivariate prediction model used in this invention accurately captured the evolution of the active and inactive periods of the typhoon season in the real-time rolling forecasts for 2024 and 2025, demonstrating the model's excellent ability to predict the sub-seasonal typhoon ACE.

[0079] Step S5: Generate a high-resolution spatial probability map of ACE for the target forecast period in the Northwest Pacific based on the average ACE distance prediction results for each type of TC.

[0080] Taking the ACE anomaly values ​​of various TCs output by the multivariate prediction model for the next 1-6 weeks as an example, we add the weekly seasonal cycle of each ACE to obtain the total weekly ACE of each TC. We then multiply the total weekly ACE of each TC by the corresponding TC's weekly ACE climatological spatial distribution probability matrix, and perform superposition and fusion to finally generate a high-resolution spatial probability map of ACE for the entire Northwest Pacific.

[0081] Example 2

[0082] This invention provides a subseasonal forecasting system for tropical cyclones in the Northwest Pacific, comprising:

[0083] The label creation module is configured to acquire historical climatological data, extract the average ACE distance of each type of TC in each time window from the climatological data, and use it as the forecast label for each time window.

[0084] The component generation module is configured to use multiple large-scale environmental fields for climate reanalysis as potential forecast factors and obtain multiple components for each potential forecast factor, including the contemporaneous ISO component, the previous observation ISO component, and the previous observation LFBS component.

[0085] The component combination module is configured to build a univariate prediction model for each component of each TC class. The univariate prediction model is pre-trained and evaluated using the components and forecast labels to select the component combination with the most nonlinear forecast value for each TC class.

[0086] The model preparation module is configured to build a multivariate prediction model for each type of TC. The multivariate prediction model is trained by combining components and forecast labels. The trained multivariate prediction model is then used to predict the average ACE distance for each time window of the target forecast period.

[0087] The forecast generation module is configured to generate a high-resolution spatial probability map of ACE for the target forecast period in the Northwest Pacific based on the average ACE distance prediction results for each type of TC.

[0088] Example 3

[0089] Based on the subseasonal forecasting method for tropical cyclones in the Northwest Pacific provided in Embodiment 1, this embodiment of the invention provides an electronic device, including a processor and a storage medium;

[0090] Storage media are used to store instructions;

[0091] The processor is used to perform operations according to instructions to execute the steps according to the method described above.

[0092] Example 4

[0093] Based on the Northwest Pacific tropical cyclone subseasonal forecasting method provided in Embodiment 1, this embodiment of the invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the above method.

[0094] Example 5

[0095] Based on the Northwest Pacific tropical cyclone subseasonal forecasting method provided in Embodiment 1, this embodiment of the invention provides a computer program product, including a computer program / instruction, which, when executed by a processor, implements the steps of the above-described method.

[0096] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0097] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0098] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0099] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0100] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for subseasonal forecasting of tropical cyclones in the Northwest Pacific, characterized in that, include: Acquire historical climatological data, extract the average ACE distance of each type of TC in each time window from the climatological data, and use it as the forecast label for each time window; Multiple large-scale environmental fields used for climate reanalysis are used as potential forecasting factors, and multiple components of each potential forecasting factor are obtained, including the contemporaneous ISO component, the previous observation ISO component, and the previous observation LFBS component. For each component of each TC, a univariate prediction model is constructed. Each univariate prediction model is pre-trained and evaluated using the component and the prediction label. The component combination with the most nonlinear prediction value corresponding to each TC is selected. A multivariate prediction model is constructed for each type of TC. The multivariate prediction model is trained by combining the components and the forecast label. The trained multivariate prediction model is then used to predict the average ACE distance for each time window of the target forecast period. A high-resolution spatial probability map of ACE for the target forecast period in the Northwest Pacific is generated based on the average ACE distance prediction results for each type of TC.

2. The method for subseasonal forecasting of tropical cyclones in the Northwest Pacific according to claim 1, characterized in that, The classification of TC includes: Obtain the optimal path dataset for TCs during the target time period, and extract target TC samples from the optimal path dataset; use c-means fuzzy clustering analysis to classify the target TC samples and generate TC categories; The target TC sample is the sum of the squares of the maximum 6-hour central wind speed of all TCs with maximum surface wind speeds greater than the wind speed threshold in each time window of the Northwest Pacific.

3. The method for subseasonal forecasting of tropical cyclones in the Northwest Pacific according to claim 1, characterized in that, The large-scale environmental field includes outward longwave radiation (OLR), 700 hPa specific humidity, 500 hPa vertical velocity, 850 hPa vertical wind shear, 200 hPa vertical wind shear, 850 hPa vorticity, 200 hPa divergence, surface latent heat flux, surface sensible heat flux, and sea surface temperature (SST).

4. The method for subseasonal forecasting of tropical cyclones in the Northwest Pacific according to claim 1, characterized in that, The concurrent ISO component is the average anomaly field of each time window obtained from the large-scale environmental field of the target forecast period extracted from the existing forecast system. The ISO component of the previous observation is the average anomaly field of each time window obtained from the large-scale environmental field of the target forecast period 1-n time windows extracted from the existing observation reanalysis dataset. The LFBS component of the previous observation is the low-frequency background state of each time scale obtained from the large-scale environmental field of 1-m time scales before the target forecast period extracted from the existing observation reanalysis dataset. The time scale is a k-time window, where k is greater than 1.

5. The method for subseasonal forecasting of tropical cyclones in the Northwest Pacific according to claim 1, characterized in that, The component combination with the most nonlinear pre-quote value corresponding to each type of TC includes: After the univariate prediction model corresponding to each component of each class of TC is pre-trained, the prediction skill evaluation index of the univariate prediction model is calculated; the prediction skill evaluation index includes the time correlation coefficient (TCC), the spatial correlation coefficient (PCC), and the root mean square error (RMSE). Based on the prediction technique evaluation index, the optimal combination of multiple concurrent ISO components, multiple previous observation ISO components, and multiple previous observation LFBS components is selected.

6. The method for subseasonal forecasting of tropical cyclones in the Northwest Pacific according to claim 1, characterized in that, The multivariate prediction model includes a feature extraction module, a feature concatenation module, and a cross-scale nonlinear coupling module. The feature extraction module uses the feature extraction module in the pre-trained univariate prediction model of each component in the component combination to extract features; The feature splicing module is used to splice the feature extraction results of each component in the component combination in the channel dimension to generate spliced ​​features. The cross-scale nonlinear coupling module includes two nonlinear convolutional layers, which are used to learn the interaction between components based on the spliced ​​features and generate a deep coupling result.

7. A subseasonal forecasting system for tropical cyclones in the Northwest Pacific, characterized in that, include: The label creation module is configured to acquire historical climatological data, extract the average ACE distance of each type of TC in each time window from the climatological data, and use it as the forecast label for each time window. The component generation module is configured to use multiple large-scale environmental fields for climate reanalysis as potential forecasting factors and obtain multiple components for each potential forecasting factor, including contemporaneous ISO components, previous observation ISO components, and previous observation LFBS components. The component combination module is configured to construct a univariate prediction model for each component of each TC class, pre-train and evaluate each univariate prediction model using the component and the prediction label, and select the component combination with the most nonlinear prediction value for each TC class. The model preparation module is configured to build a multivariate prediction model for each type of TC, train the multivariate prediction model by combining the components and the forecast label, and use the trained multivariate prediction model to predict the ACE distance average for each time window of the target forecast period. The forecast generation module is configured to generate a high-resolution spatial probability map of ACE for the target forecast period in the Northwest Pacific based on the average ACE distance prediction results for each type of TC.

8. An electronic device, characterized in that, Including processor and storage media; The storage medium is used to store instructions; The processor is configured to operate according to the instructions to perform the steps of the method according to any one of claims 1-6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the program implements the steps of the method according to any one of claims 1-6.

10. A computer program product, comprising a computer program / instructions, characterized in that, When the computer program / instructions are executed by the processor, they implement the steps of the method described in any one of claims 1-6.

Images