Wind speed time series exhibit complex patterns, thus integrating wind energy into the electrical system is challenging. This requires specialized skills for operations and planning practices. Training expert predictors on different parts of the time series enables the identification of complex local patterns. However, the partitioning procedure reduces the number of training instances. So, investigating the size of training partitions for an ensemble is desirable for predicting wind speed. Therefore, this paper proposes a homogeneous ensemble for local pattern recognition denoted as LocPart, that varies in partition size. The results of the Diebold-Mariano hypothesis test show promise for the LocPart method applied to three wind speed time series. The comparison was made relative to individual and bagging methods that use a global mapping of the respective base model of the proposal. The LocPart method with LSTM, ARIMA, and ELM base models won in 100%, 83%, and 50% of the cases, respectively.