An Ultra-Short-Term Wind Power Forecasting Method Based on Convolutional Transformer Network

Wind power generation is a primary energy supply resource in the field of renewable energy sector. However, the accuracy and timeliness of wind power forecasting (WPF) are among the key challenges faced during the integration of wind power into the grid. When dealing with wind power forecasting problems characterized by complex meteorological changes, conventional physical models and statistical analysis often fail to achieve the desired prediction accuracy. However, with the rise of artificial intelligence (AI) large models in recent years, the transformer, as its core architecture, has gradually been applied to time series forecasting problems. Transformer performs well in forecasting such problems and shows significant potential for broader applications. However, due to its high computational requirements for model training, this process typically incurs substantial time costs when using standard hardware. Therefore, to ensure high prediction accuracy while addressing the issue of low training efficiency, this paper proposes a hybrid network structure called convolutional transformer network (ConvTransNet) to tackle the WPF problem. This paper utilizes publicly available, anonymized data from a wind farm in Fujian, China, as the working dataset. The dataset contains wind power data with multiple meteorological features spanning two years and four months. The experiment evaluates the proposed hybrid network structure using four metrics: root mean squared error (RMSE), mean absolute error (MAE), the coefficient of determination (R2), and training and inference time. The results demonstrate that ConvTransNet improves the training and inference efficiency while optimizing the prediction accuracy of the transformer model.