Temporalformer: A Temporal Decomposition Causal Transformer Network For Wind Power Forecasting

Yansong Wang, Lili Pei, Yingying Wang, Yuanlin Zhao, Zhenzhen Xing

Abstract: Accurate short-term forecasting of wind power generation is an effective way to ensure grid stability and rational dispatch. Although the transformer structure has made significant progress in the field of time series forecasting, its forecasting performance for wind power data is a concern due to the high variability and stochasticity of short-term wind power data. To address this issue, this study proposes a novel wind power forecasting model, Temporalformer. First, the wind power series are decomposed using the Variable Mode Decomposition (VMD) algorithm, which reduces the volatility of the series while capturing the wind power series at different time scales. Subsequently, other influencing factors are fused with the decomposed sequences at different scales to reduce the computational complexity of the model. The forecasting model is based on a transformer encoder structure that uses Time2Vec instead of the traditional sine and cosine positional encoding to better capture temporal information. Meanwhile, a temporal causal convolutional network (TCN) is added to the multi-head attention mechanism, which helps to enhance the ability of the Transformer network to extract local temporal information. We validate our proposed model(Temporalformer) using data from an onshore wind farm in China. Experiments show that our model has higher accuracy compared to many classical forecasting models.