DSGCformer: Dual Self-Gated Convolution iTransformer for Short Wind Power Forecasting

Accurate and reliable wind power forecasting models are essential for stable power system operations and large-scale grid integration. Nevertheless, predicting wind power remains challenging due to the inherent unpredictability and stochastic characteristics of wind resources. Hence, this paper proposes a prediction model named DSGCformer. This model comprises Multivariate Attention Block and Dual Self-Gated Convolution Block (DSGCB). Derived from the iTransformer, Multivariate Attention Block preserves the multi-head attention mechanism's ability to capture long-term dependencies in time series data. Additionally, it incorporates a dedicated focus on physical relationships between individual variables. Furthermore, DSGCformer replaces the Feedforward Network (FFN), a key component of iTransformer, with the DSGCB. The DSGCB consists of three convolutional layers. Compared to FFN, it demonstrates an enhanced capacity to interpret complex temporal patterns and capture local abrupt changes. The efficacy of our model has been validated using two public datasets sourced from Kaggle. The findings reveal that the DSGCformer has the highest prediction accuracy and the best forecast effect.