Ultra-short-term wind power forecasting: A hybrid parallel architecture integrating informer and SENet-enhanced temporal convolutional networks

In recent years, the increasing share of wind energy in power grids has made wind power forecasting crucial for grid stability and efficiency. Advanced technologies such as deep learning have improved forecasting, particularly in addressing the volatility and complexity of wind power data. However, existing studies often lack effective model integration and optimal hyperparameter tuning. Many complex algorithms fail to achieve synergy, and latent information in wind power data remains underexplored, limiting further accuracy improvements. Therefore, efficiently integrating deep learning methods and fully mining data features are key challenges for advancing ultra-short-term wind power forecasting. This paper proposes an ultra-short-term wind power forecasting model based on the parallel fusion of Informer and TCN-SENet architectures. The model fully leverages the efficiency of Informer in long-sequence modeling and the sensitivity of TCN-SENet to local temporal features. Through feature engineering and a sliding window mechanism, the model achieves high-precision forecasting with multivariate inputs and univariate outputs. This paper conducts comparative experiments using real-world wind farm datasets against Transformer-encoder, Informer, and their improved variants. The results demonstrate that this paper’s model consistently outperforms the benchmark models across all performance metrics, exhibiting superior generalization ability and robustness. Furthermore, this paper’s model shows clear advantages in capturing both the trend and fine-grained variations of wind power output. This paper provides an effective new approach for ultra-short-term, high-precision wind power forecasting, and offers theoretical and methodological support for smart grid scheduling and renewable energy integration.