Enhancing Wind Power Forecasting Accuracy: A Parallel BiLSTM-CNN Hybrid Architecture with Temporal-Spatial Feature Fusion

The stochastic and fluctuating characteristics of wind power generation pose a major challenge to the stable operation of power systems, and accurate prediction techniques need to be studied urgently. Existing methods have limitations in dealing with time-dependent and spatial features, and especially perform poorly under complex meteorological conditions. In this study, we propose a parallel hybrid architecture based on BiLSTM and CNN, which synchronously captures temporal dependence and local features through a two-way feature extraction mechanism, and integrates multi-dimensional information using a feature fusion strategy. Empirical analysis based on the Kaggle public dataset shows that thehybrid model significantly outperforms traditional machine learning methods and single deep learning models, with the root mean square error reduced by 40.2% and 36.6% compared to LSTM and CNN, respectively. Although the training complexity is higher, the inference time (0.23 seconds) still meets the real-time prediction demand. The results confirm that the architecture can effectively solve the prediction challenges of typical complex scenarios such as sudden wind changes and low wind speed fluctuations, providing a new theoretical framework for wind power prediction, which is of great value in promoting renewable energy grid integration and economic operation of power systems.