Containerization is increasingly used in cloud service for its convenience and heterogeneous consistency. The container workload is the combination of the coexistence of second-level burst requests and hour-level periodic tasks, and that makes the workload prediction a complex and time-consuming job. However, container instance requires rapid prediction method to ensure the minute-level elastic scaling, thus make a conflict. In this paper, we propose an efficient neural network model for workload prediction. Our approach combines flash-attention mechanism with Long Short-Term Memory neural network to improve the efficiency and performance of our model on the real-world workload data from the Alibaba Cluster Trace dataset. Finally, we compare the training time and performance of our model with the LSTM-based model and the Transformer-based model in the different time scales, and it brings up to 15.18% improvement in performance and an average of nearly 20% less training time. The result of the experiment demonstrates its efficiency and better performance.