EP-4735921-A2 - TROPICAL STORM FORECASTING SYSTEM AND METHODS

Abstract

A system for and method of generating storm forecasts is provided. The method utilizes de-storm and re-storm processes to avoid washout associated with location divergence. During the de-storm process, storm information is identified in and extracted from a plurality of forecast models, thereby generating a plurality of discrete storm models and a plurality of associated background models. The background models are then blended, and the discrete storm models are weighted, thereby generating a blended background model and a weighted discrete storm model, respectively. During the re-storm process, the weighted discrete storm model is added to the blended background model, thereby generating a forecast representing amplitudes associated with the tropical storm.

Inventors

• LILLO, Sam
• BIJLSMA, FLORIS

Assignees

• DTN, LLC

Dates

Publication Date

20260506

Application Date

20240701

Claims (20)

1. 1. A method of blending a plurality of forecast models, the method comprising: generating a plurality of discrete storm models, each discrete storm model being generated from a respective forecast model of the plurality of forecast models; extracting each discrete storm model from its respective forecast model, thereby generating a background model for each forecast model; blending the background models, thereby generating a blended background model; and adding a weighted storm model to the blended background model, the weighted storm model being generated using information from at least one discrete storm model of the plurality of discrete storm models.
2. 2. The method of claim 1, wherein each forecast model comprises a modeled wind field, each discrete storm model comprises a discrete wind field, and each background model comprises a background wind field that is determined by subtracting the discrete wind field from the modeled wind field.
3. 3. The method of claim 2, wherein generating each discrete storm model comprises: calculating vorticity for the modeled wind field; finding a maximum vorticity associated with a first tropical storm; applying a cosine-tapered window radially outward from the maximum vorticity, thereby isolating vorticity in the vicinity of the tropical storm from any other potential vorticity associated with the respective modeled wind field; and converting the isolated vorticity into a global spectral space.
4. 4. The method of claim 3, further comprising: identifying a geographic center for each extracted discrete storm, thereby identifying an extraction center for each background model; and adding a respective first amount of vorticity to each background model, the added vorticity being centered around the respective extraction center for each background model, wherein each first amount of vorticity is proportional to vorticity of its respective extracted discrete storm model, and wherein the added vorticity for each background model tapers away from the respective extraction center.
5. 5. The method of claim 3, further comprising generating the weighted storm model using information from each of the plurality of discrete storm models, values for each discrete storm model being organized based on a position of each value relative to a center of the respective discrete storm model, and a weight for each discrete storm model being determined based on a distance of the center of the respective discrete storm model from a center of the weighted storm model.
6. 6. The method of claim 5, wherein the weight of each discrete storm model is determined using a reverse exponential relationship with the distance of the center of the respective discrete storm model from the center of the weighted storm model.
7. 7. The method of claim 6, wherein the weight is one for each discrete storm model that is concentric with the weighted storm model.
8. 8. The method of claim 7, wherein the center of the weighted storm model is determined using information from a confirming forecast storm model.
9. 9. The method of claim 8, further comprising utilizing a neural network to identify storm information for each forecast model, wherein extracting each discrete storm model from its respective forecast model comprises extracting the identified storm information from each forecast model.
10. 10. The method of claim 9, wherein the identified storm information is precipitation information.
11. 11. The method of claim 1, further comprising utilizing a neural network to identify storm information for each forecast model, wherein extracting each discrete storm model from its respective forecast model comprises extracting the identified storm information from each forecast model, and wherein the identified storm information is precipitation information.
12. 12. A method for generating a forecast for a tropical storm, the method comprising: utilizing a neural network to identify storm information for each guidance model of a plurality of guidance models; extracting the storm information from each guidance model, thereby generating a background model for each guidance model; processing the background models to generate a blended background model; and adding weighted storm information to a first region of the blended background model.
13. 13. The method of claim 12, wherein the storm information comprises at least one of precipitation information, storm surge information, and wind information.
14. 14. The method of claim 13, wherein the storm information comprises precipitation information, and wherein precipitation is effectively zero in the first region of the blended background model prior to adding the weighted storm information to the blended background model.
15. 15. The method of claim 14, wherein the storm information comprises storm surge information, and wherein storm surge information is effectively zero in the first region of the blended background model prior to adding the weighted storm information to the blended background model.
16. 16. The method of claim 15, wherein the storm information comprises wind information, and wherein wind information is not equal to zero in the first region of the blended background model prior to adding the weighted storm information to the blended background model.
17. 17. A system for generating a forecast for a tropical storm, the system comprising: a de-storm module for identifying and extracting storm information from a plurality of forecast models, thereby generating a plurality of background models; a blending module for blending the background models, thereby generating a blended background model; and a re-storm module for adding a weighted storm model to the blended background model.
18. 18. The system of claim 17, further comprising a weighting module for generating the weighted storm model using information from each of the plurality of discrete storm models, values for each discrete storm model being organized based on a position of each value relative to a center of the respective discrete storm model, and a weight for each discrete storm model being determined based on a distance of the center of the respective discrete storm model from a center of the weighted storm model.
19. 19. The system of claim 18, wherein the weight of each discrete storm model is determined using a reverse exponential relationship with the distance of the center of the respective discrete storm model from the center of the weighted storm model.
20. 20. The system of claim 19, wherein the weight is one for each discrete storm model that is concentric with the weighted storm model.

Description

TROPICAL STORM FORECASTING SYSTEM AND METHODS Cross-Reference to Related Applications [0001] This application claims priority pursuant to U.S. Provisional Patent Application Serial No. 63/524,509, filed June 30, 2023, the entire disclosure of which is incorporated herein by reference. Field of the Invention [0002] The present invention relates generally to tropical storm forecasting. More specifically, the present invention is concerned with an objective method for applying deep learning models for deterministic prediction of full spatiotemporal fields of wind, pressure, and rainfall associated with tropical cyclones. Background [0003] Many existing weather forecast engines are built on a multi-model ensemble blend, with the blend itself involving sophisticated weighting and sophisticated bias correction. Unfortunately, existing blending techniques result in large perturbations in the forecasts being washed out. While presumably satisfactory for certain forecasting purposes, such as a temperature forecast that minimizes root mean square errors, the amplitude of the perturbations is important when forecasts involve extreme events, such as hurricanes, cyclones, typhoons, and the like (each a “tropical storm”). Accordingly, it would be beneficial to have a blending technique that does not wash out large perturbations. [0004] For various existing forecast models, certain important features of each individual tropical storm forecast model diverge spatially over time such that blending leads to blurring, de- amplification, or other distortion (each a “distortion”) of such important features in the final blended tropical storm forecast model. This is especially problematic for situations in which a forecast bifurcates, such as the example multi -model forecast track for Hurricane Sandy shown in Figs. 1A-1H. Accordingly, it would be beneficial to have a modeling technique that adjusts for spatial divergence to avoid distortion of the modeled tropical storm. [0005] Referring to Fig. 2, an example of a distorted forecast is provided. In particular, Fig. 2 shows a Global Forecast System (“GFS”) forecast model that shows two distinct eyes, one being shifted 200 miles westward from the other. As will be understood, the duplicate eyes are a result of blending two models that differ in their anticipated location and/or track for a particular tropical storm. In addition to showing two distinct eyes, the effect of blending these two modeled storms into a single forecast model is a broadened, weakened, and distorted wind field that does not accurately reflect the modeled wind fields associated with either model. This significant distortion of the wind field in this blend has direct ramifications on forcing in wave models that employ the blended forecast. Accordingly, it would be beneficial to have a modeling technique that eliminates or otherwise reduces distortions to the associated wind fields. [0006] Because spatial divergence of forecast models tends to increase as a forecast period increases, existing blending techniques have limited value for forecasts beyond a few days or hours. But preparations for tropical storms often require notice of the severity and anticipated track of the storm several days in advance. Accordingly, it would be beneficial to have a blending technique that facilitates long-term forecasting for tropical storms. [0007] Existing forecasting systems rely on certain critical information provided by existing software, such as the Automated Tropical Cyclone Forecasting System or the like (each being referred to herein as “ATCF”). Unfortunately, precipitation is not a parameter that is described in any manner in the ATCF. For instance, there is no condensed summary or reduced dimensionality of this parameter assigned to specific tropical storms. As such, precipitation must be extracted from the full grids of forecast model guidance. The challenge is then to identify the tropical storm within those grids. To the extent that existing systems may be capable of extracting precipitation from full grids of forecast model guidance, such extraction is time consuming and unreliable. Accordingly, it would be beneficial to have a system for and method of quickly and reliably obtaining precipitation information for tropical storms. Summary [0008] The present invention comprises systems for and methods of blending traditional forecast models. The systems and methods utilize specialized modeling techniques that, when compared with modeling techniques of existing systems, are better suited for forecasting tropical storms. In particular, the blending technique of the present invention does not wash out large perturbations associated with tropical storms. [0009] In some embodiments, the present invention adjusts for spatial divergence, thereby avoiding associated distortions of the modeled tropical storm. In some such embodiments, storm elements are removed from a model prior to a blend, and the storm elements are a