GCN feature fusion and incremental self-attention-based transformer for wind power continual prediction

The prediction of wind power is influenced by various factors, such as climate conditions, geographical location, and the status of adjacent wind turbine generators, which exhibit complex dependencies. In addition, the data distribution of wind power generation may also shift over time. Prediction models trained on previous data may not have good prediction performance on new data. Therefore, a Graph Convolutional Neural Network (GCN) Feature Fusion and Incremental Self-Attention based Transformer (GCN-IncreFormer) is proposed in this paper for wind power continual prediction. The GCN based feature fusion module effectively integrates and refines key features from different channels while retaining temporal information, thereby enhancing prediction accuracy. Additionally, we introduce incremental learning into the self-attention mechanism of the Transformer and propose IncreFormer. Consequently, model can adapt to continuously changing data distributions by continuously updating with newly collected data to accommodate dynamic environmental changes. Experimental results demonstrate that GCN-IncreFormer can effectively explore the deep correlations of different features in wind power data, thereby improving power prediction accuracy. The introduction of continual learning mechanism enables the model to adapt to data distributions shifts over different time periods. Compared to baselines, GCN-Increformer achieves a reduction in Mean Absolute Error (MAE) of over 2.40. Comprehensive comparative experiments and ablation studies validate the reliability and practicality of the proposed method.