Multi-Step Prediction Method of Short-Term Wind Power based on CTAR Model

Short-term wind power prediction provides an essential foundation for intelligent scheduling of the power system, which is vital for reducing the impact of wind power grid integration and enhancing the stability of the power system. However, existing short-term wind power prediction methods suffer from insufficient feature extraction capability and low prediction accuracy when dealing with stochastic non-stationary wind power sequences. Therefore, we propose a deep hybrid model for short-term wind power prediction, the CNN-Transformer-Autoregressive (CTAR) model. The CTAR model achieves the prediction task by learning the mapping between historical data and wind power at future moments. We use convolutional neural networks (CNNs) and Transformer to model short-term and long-term temporal patterns of wind power series, respectively. Influences on wind power such as temperature and barometric pressure are also used to enhance the mode's expressive power. Furthermore, the CTAR model uses a traditional autoregressive (AR) model to solve the neural network scale insensitivity problem. By evaluating wind power data from Valencia, Spain, we demonstrate the significant performance improvement of the CTAR model over several other commonly used baseline methods for wind power prediction.