Abstract: Accurate wind power forecasting provides essential guidance for formulating power generation plans and coordinated dispatching. This paper proposes a heterogeneous similarity-selective ensemble just-in-time learning method (HS-ELWPLS) using locally weighted partial least squares (LWPLS) for ultra-short-term forecasting. First, we construct heterogeneous similarity metrics to adapt to varying wind power process states and establish multiple just-in-time learning (JIT) local models. Second, since not all similarity frameworks suit current operational conditions, evolutionary multi-objective optimization is employed to select similarity functions, generating accurate yet diverse sub-models. Finally, a Stacking ensemble strategy integrates these sub-models for final predictions. The effectiveness of the proposed method is verified by the real data of wind farms in Yunnan Province.