Traffic prediction is of great significance for intelligent capacity planning and task scheduling. However, large-scale e-commerce cluster traffics have various uncertain emergencies, such as online promotion activities and user aggregation requests. These uncertain events may cause many bursts in the time series, which poses a huge challenge to traffic prediction. At the same time, capacity prediction should be robust to uncertainty. That is, it should cope well with possible future situations and ensure cluster stability, rather than shrink the capacity strictly based on the prediction. For the traffic scenarios of large-scale distributed e-commerce clusters and the requirements of dynamic capacity planning, this paper proposes a real-time load forecasting framework with uncertainty estimates. The framework is based on multivariate long short-term memory auto-encoder and Bayesian theory, which can provide accurate uncertainty interval estimation while performing flow deterministic prediction.