Short-Term Wind Power Forecasting Method Based on Implicit Scenario Discovery and Dynamic Weighted Fusion

Abstract

Wind power intermittency and uncertainty significantly increase the complexity of power grid dispatch and impose more stringent requirements on the secure and stable operation of power systems. To address this challenge, this paper proposes a scenario-oriented dynamic selection and fusion decision framework based on dynamic quantile thresholds and a performance matrix. First, a forecasting model pool consisting of physical mechanism models, statistical regression models, and individual deep learning models is constructed, and the applicable boundaries of different model categories are clarified. Second, a three-dimensional orthogonal feature space is established, and dynamic quantile thresholds are introduced to enable implicit scenario identification for wind power time series. Finally, a pyramidal evaluation index system is developed, together with a dynamic selection and fusion triggering mechanism. On the basis of the scenario–model performance matrix, the proposed framework adaptively switches between the two modes of single-model locking and multi-model weighted fusion. Experimental results using measured data from a wind farm show that the proposed method effectively mitigates the degradation in forecasting performance of individual models under extreme scenarios, such as ramping events and cut-in/cut-out moments. Consequently, the method significantly improves forecasting accuracy while demonstrating strong scenario adaptability and robustness.