DST-CAN: A dynamic spatiotemporal convolutional attention network for short-term wind power forecasting

Abstract

ABSTRACT With the continuous increase in wind power penetration, accurate and reliable short-term wind power forecasting is pivotal for optimizing grid dispatch and ensuring system stability. However, the complex spatiotemporal coupling within wind farms poses severe challenges for forecasting. Most existing Spatiotemporal Graph Neural Networks (STGNNs) are confined to predefined static graph structures, failing to effectively capture the dynamic wake effects induced by wind direction fluctuations and the non-stationarity of spatial distributions. Furthermore, the temporal modules of existing models often struggle to balance the acute capture of local high-frequency abrupt changes with the effective modeling of global long-range dependencies. To address these limitations, this paper proposes a Dynamic Spatiotemporal Convolutional Attention Network (DST-CAN). First, we design a Metric-Learning-based Dynamic Topology Module that reconstructs the time-evolving spatial dependency topology among wind turbines by adaptively fusing physical distance priors with real-time feature similarities. Second, a hierarchical hybrid spatiotemporal architecture is introduced, which synergistically utilizes dilated causal convolutions to extract local trends and instantaneous fluctuations, while incorporating a multi-head self-attention mechanism to aggregate global context information, thereby achieving a complementary fusion of local and global features. Extensive experiments on the real-world datasets demonstrate that DST-CAN significantly outperforms existing methods in multi-step forecasting tasks. Ablation studies further validate the effectiveness of dynamic topology modeling in handling non-stationary wake effects.