Using deep-learning based time series forecasting for the prediction of wind behaviour in sail races
Accurate wind forecasts are critical for tactical race planning in sailboat races. Understanding how wind behaves to establish an effective plan can significantly influence race outcomes. Traditionally, wind behavior is interpreted based on weather model forecasts combined with meteorological expert...
| Autor: | |
|---|---|
| Formato: | tesis de maestría |
| Fecha de publicación: | 2025 |
| País: | España |
| Recursos: | Universitat Politècnica de Catalunya (UPC) |
| Repositorio: | UPCommons. Portal del coneixement obert de la UPC |
| Idioma: | inglés |
| OAI Identifier: | oai:upcommons.upc.edu:2117/449600 |
| Acesso em linha: | https://hdl.handle.net/2117/449600 |
| Access Level: | acceso embargado |
| Palavra-chave: | Weather forecasting Deep learning (Machine learning) Linear models (Statistics) Time series Time steps Forecasting models Deep learning Statistical model Linear model VAR Transformer Encoder Patchification process Channel-independence Wind patterns wWnd behaviour U component V component Wind direction Wind speed MSE Lag Forecast horizon Previsió del temps Aprenentatge profund (Aprenentatge automàtic) Models lineals (Estadística) Àrees temàtiques de la UPC::Informàtica::Intel·ligència artificial::Aprenentatge automàtic |
| Resumo: | Accurate wind forecasts are critical for tactical race planning in sailboat races. Understanding how wind behaves to establish an effective plan can significantly influence race outcomes. Traditionally, wind behavior is interpreted based on weather model forecasts combined with meteorological expertise and manual analysis. However, the expansion of the set of high-resolution data from targeted sailing racing areas, combined with advances in state-of-the-art artificial intelligence algorithms, has introduced powerful new tools for time series forecasting that have the potential to improve the accuracy and adaptability of wind predictions. This thesis investigates and compares the performance of traditional statistical models and state-of-the-art AI methods for forecasting wind direction and speed. Unlike general wind forecasting, this study focuses on predicting highly localized wind dynamics at small-scale, race-specific locations. Specifically, it evaluates Vector Autoregression (VAR), Linear, NLinear, and deep learning approaches such as the standard Transformer and PatchTST. These models are trained and tested on wind data collected during training and pre-race sessions off the coast of Marseille, the venue for the Paris 2024 sailing events. The study examines the effect of model type, input lag length, and forecast horizon on forecast accuracy. Results emphasize the merits and limitations of each modeling approach, highlighting their pragmatic applicability to sailing strategy and race planning. The findings demonstrate that deep learning models are well-suited for localized wind forecasting and pave the way for creating practical tools that support tactical decision-making in competitive sailing. |
|---|