SFINet: Softmax-Guided Feature Fusion with Inception Blocks for Accurate Short-Term Wind Power Forecasting

Against the background of the global transition to new energy, wind power has gained widespread application due to its advantages in resource availability and environmental sustainability. However, influenced by natural meteorological dynamics, wind power output exhibits strong temporal fluctuations and intermittent characteristics, posing potential risks to the power grid. This necessitates short-term forecasting of wind power output to support real-time dispatch and system stability control. To address this, this paper proposes a novel deep learning forecasting model—SoftFusion-Inception Network (SFINet). This model employs a feature fusion architecture guided by scaling and softmax, enhancing feature expression capabilities and model interpretability, while utilizing Inception modules to form branches within the network. The model was evaluated on real-world wind power data from Natal, Brazil, performing 24-hour forecasts. The SFINet was compared against six other models, including InceptionV3 and GRU. Experimental results demonstrate that the SFINet achieves optimal performance across multiple metrics, including MAE, RMSE, and MAPE, while significantly reducing training time. This validates the effectiveness of the SFINet in capturing the complex spatiotemporal dynamics of wind power, providing robust technical support for short-term renewable energy forecasting in practical scenarios.