Stable and reliable wind power forecasting is of great significance for grid dispatching. However, wind power has strong nonlinear characteristics. Its amount of historical data is very large, and traditional machine learning methods cannot fit the nonlinear relationship well. Furthermore, owing to the seasonality and periodicity of the wind power, the performance of the offline model inevitably deteriorates. To address these problems, this study proposes an adaptive ensemble model for ultra-short-term wind power forecasting. First, a long short-term memory (LSTM) network is established with historical wind power data for an entire year to ensure a strong generalization performance. Second, local weighted partial least squares (LWPLS) were used to obtain LSTM network prediction errors in real time by adaptive modeling and further improving its prediction accuracy. In addition, just-in-time learning (JITL) is used to ensure the adaptability of forecasting using LWPLS. Finally, to fully integrate the prediction advantages of different prediction models, a random forest (RF) was used to transform the determination of the ensemble weight into a three-classification problem. The adaptive fusion of the three different forecasting mechanisms is realized by weighting through the posterior probability of classification. Finally, the effectiveness and superiority of the proposed method are verified in a practical case.