The invention discloses a short-term-wind-speed multi-step prediction method and device. The short-term-wind-speed multi-step prediction method includes the steps that an original wind speed time sequence is segmented into a plurality of stable sub-sequences with the heuristic segmentation algorithm; each stable sub-sequence is decomposed into a series of limited-bandwidth sub-patterns with the self-adaptive variable pattern decomposition technology; a basic prediction model for each sub-pattern is built through a robust online limitation learning machine of a forgetting factor based on the COOK distance, and the prediction models are subjected to parameter optimization with the crossing optimization algorithm with the self-adaptive variation mechanism; according to the online ensemble learning method and the orderly aggregation technology, a final prediction value is obtained by carrying out weighted aggregation on the multiple basic prediction models. In this way, the models are updated according to changes of the wind speed through the robust online limitation learning machine of the forgetting factor based on the COOK distance, the multiple basic prediction models are subjected to weighted aggregation with the online ensemble and orderly aggregation technology (OEOA) to obtain the final prediction value, and the accuracy of multi-step prediction is greatly improved.