DS-MS-convformer: Ultra-short-term wind power prediction based on dual-stream multi-scale convolutional transformer and hierarchical dynamic fusion

Against the backdrop of the global energy transition and the integration of large-scale wind power, ultra-short-term wind power forecasting is of great significance for ensuring the secure operation of power grids and the efficient dispatch of wind farms. However, wind power sequences are affected by complex meteorological factors and operating conditions, exhibiting pronounced non-stationarity. Their fluctuation characteristics primarily manifest as a continuous coupling between transient disturbances in the time domain and periodic variations in the frequency domain, posing challenges for multi-scale time-frequency structural representation and dynamic feature fusion. Existing methods often struggle to decouple and collaboratively fuse time-domain dynamic features and frequency-domain periodic structures while preserving the original temporal evolution information. To address this issue, this paper proposes an ultra-short-term wind power forecasting model based on dual-stream time-frequency decoupled modelling, a multi-scale convolutional Transformer, and a hierarchical dynamic fusion mechanism, namely DS-MS-Convformer. The proposed model first models the original time-domain dynamic features and the frequency-domain periodic features separately using a dual-stream time-frequency-decoupled framework. Then, a multi-scale convolution-enhanced Transformer encoder is constructed by integrating convolution kernels with different receptive fields and the self-attention mechanism, thereby jointly capturing local fluctuation features and long-range temporal dependencies. Finally, adaptive fusion of multi-source time-frequency features is achieved through low-, medium-, and high-frequency feature reconstruction, cross-attention interaction, and gated weight allocation. Experimental results on real wind farm datasets show that the proposed method achieves lower forecasting errors across single-step, multi-step, cross-seasonal validation, and generalizability testing. Ablation studies further verify the effectiveness of each key module. The results provide a reference for structured time-frequency modelling of non-stationary wind power sequences.