Spatiotemporal Graph Neural Networks for Wind Energy Production Forecasting

Over the past few decades, global energy consumption has steadily increased, driven by population growth, industrialization, and technological advancements. In response to the growing energy demand and the shift towards sustainable power generation, wind energy has gained significant attention due to its environmental benefits and economic viability. The main element involved in wind energy production is the wind turbine, a device that converts the kinetic energy of wind into electrical energy relying on the principle of electromagnetic induction. However, due to the inherent variability of wind patterns and environmental conditions, wind energy production is characterized by a dynamic output, which poses operational challenges for its integration into the power grid. In this context, predictive modeling of wind energy output plays a relevant role in supporting the dynamic management of wind operations, enabling strategic demand allocation and optimized use of energy storage. Recent advancements in data acquisition and control technologies have facilitated the collection and storage of large volumes of data from wind farms, which has influenced the development of sophisticated predictive models based on deep learning algorithms. The data acquired from wind farms can be naturally organized as a collection of correlated time series representing environmental and operational records for each wind turbine within a wind farm. Graph-based deep learning methods have become popular tools for processing such collections of correlated time series. Unlike traditional multivariate forecasting techniques, graph-based spatiotemporal learners leverage relational dependencies between sensors by conditioning forecasts on a (possibly dynamic) graph that spans the time series collection. This work investigates the predictive performance of Spatiotemporal Graph Neural Networks for wind energy forecasting across short and long-term predictive horizons, providing an evaluation of their effectiveness through a detailed comparative analysis against alternative approaches and highlighting their strengths and potential applications.