Spatio-Temporal Graph Neural Networks for Wind Power Forecasting

Wind Power Forecasting (WPF) represents a significant and challenging aspect of power supply management. However, forecasting wind power production is difficult due to the inherent unpredictability of wind behavior. In the literature, various methodologies exist to address this challenge. This thesis work aligns with the 2022 KDD Cup Baidu challenge, it proposes a method for predicting wind farm active power for a 48-hour horizon using past data and wind turbine spatial positioning. The main method uses a spatio-temporal graph neural network model called Graph WaveNet that captures intricate spatial and temporal relationships. Graph WaveNet shows his superiority in short-term forecasting (2 hours, 6 hours, and 12 hours ) compared to other models like a Lasso regression model and a GRU network. In Long-term forecasting (48-hour) the prediction has been further improved through the use of an ensemble model combining prediction of short-term and long-term settings. The best-performing model in this study achieves results comparable to the top 10 participants out of more than 2000 entrants in the 2022 KDD Cup challenge.