Chasing the Wind: Background Flow Tracing for Wind Speed Forecasting

Junteng Zhao, Xutao Li, Yikai Li, Kaiqi Zhao, Yunming Ye, Guangbo Deng, Chenxunlai, Huiqi You

Keywords: Time series forecasting, wind speed forecasting, physics-guided machine learning, multi-modal learning, context-enriched learning

TL;DR: We propose a novel physics-guided architecture that traces upstream advection in background wind fields to enhance local wind speed forecasting at meteorological stations.

Abstract: Wind energy is inherently intermittent and fluctuating, and the uncertainty in wind speed poses significant challenges for power system stability. Wind speed forecasting, particularly at wind farm sites, is crucial for balancing power generation and scheduling backup energy sources. Most existing studies rely solely on time-series forecasting methods, while ignoring the physical nature of wind as a spatiotemporal phenomenon primarily driven by atmospheric momentum advection. In this paper, we propose a framework that leverages the surrounding wind field as context to capture the wind dynamics. By explicitly modeling wind advection, our method traces atmospheric momentum transport, enabling the forecast of future trends. We further introduce a multi-modal dataset consisting of wind speed observations from multiple geographically distributed sites and their associated background wind field data. Experimental results demonstrate that our method achieves superior performance over traditional time-series forecasting models and state-of-the-art methods that leverage wind field information.

Primary Area: learning on time series and dynamical systems

Submission Number: 10959