This paper presents a methodology to incorporate large-scale atmospheric information into short-term wind speed forecast over a large geographical area of about 435 thousand square kilometres in Alberta, Canada. The analysis was done using two publicly accessible datasets. The ERA5 reanalysis dataset is used for atmospheric clustering by applying the k-means algorithm and the hidden Markov model on atmospheric variables related to wind speeds. It is shown that atmospheric clustering results align with some known wind pattern in Alberta. For short-term wind forecast, we propose time series regime-switching models and mixture models that integrate the clustering results to predict 6-hour ahead wind speed at 23 weather stations in Alberta, Canada. The predictive performance is compared for atmospheric clustering methods and forecasting models. The results show that models that take into account meteorological conditions perform better than those do not. Furthermore, modelling multiple locations simultaneously produces fewer forecasting errors than modelling at a single location.