WPFormer: A Spatial-Temporal Graph Transformer With Auto-Correlation for Wind Power Forecasting

The development and utilization of clean energy have been an important demand, with wind power standing out as a representative source due to its convenience and well-established technology. Accurate wind power forecasting holds immense significance as it facilitates the development of future generation plans, enhances the economy and reliability of power systems, and promotes the increased utilization of clean energy. However, an abundance of anomalous data stems from harsh turbine environments. The interplay among natural wind patterns, artificial interventions, and turbine states results in poor cyclicality and significant volatility in wind power generation. To address these challenges, we propose the WPFormer framework. Within this framework, we have designed a <underline xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">W</u>ind pow<underline xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">E</u>r <underline xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">D</u>ual-stream <underline xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">S</u>coring model (WEDS), a semi-supervised learning model based on wind power curves, tasked with anomaly detection and data repair. A <underline xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">F</u>eature s<underline xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">E</u>lection metho<underline xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">D</u> based on the <underline xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">S</u>elf-attention mechanism, known as FEDS, is also proposed for identifying valuable features. Furthermore, the spatial information of the wind turbines is introduced, and multiple series decomposition with autocorrelation and multi-headed attention is used to learn the expected prediction behind the randomness, overcome the weak periodicity, and predict the long-term significant volatility trend. Extensive experiments on real-world datasets demonstrate that WPFormer achieves state-of-the-art performance.