There are many wind farms located in different areas. For wind speed prediction, different models can be learned for different farms. This is feasible when the amount of data is sufficient. However, for a newly-built wind farm, insufficient wind speed data is available for the model design. Togelou et al. proposed a solution to this INCB028050 problem by providing two statistical models that are self-constructed and self-adapted online . They did not consider data from any other sites. An alternative strategy is to learn a model by mixed data from the target domain and other domains together. This makes sense only if the data from other farms can be directly used to train the target model. Since wind speed patterns are different across domains, the above solution is usually infeasible. Therefore, another strategy has been proposed . This strategy refers to the learning of a shared model from the set of domains, and then adapting the model to each individual domain. This only works, however, when the model is able to discover intermediate abstractions pH are shared and useful across domains.