Abstract:The widespread adoption of cloud computing in fields such as power and industry has posed scheduling challenges for measurement data centers, posing diverse tasks and volatile resources. This paper proposes a resource scheduling optimization method that integrates CNN-LSTM and deep reinforcement learning. By constructing a time-series-based resource prediction model, channel attention and self-attention mechanisms are introduced to improve prediction accuracy. A hybrid scheduling strategy (HSS-DQN) combining a deep Q-network (DQN) with multi-objective optimization is designed to achieve dynamic resource allocation and service assurance in a Kubernetes cluster. Experimental results show that the proposed method achieves mean absolute error (MAE), root mean square error (RMSE), and mean absolute percentage error (MAPE) of 0.101, 0.145, and 8.12%, respectively, on the test set, outperforming competing algorithms. In simulations of four business scenarios, the proposed method demonstrates significant advantages in average scheduling delay and task completion rate. In a resource-constrained scenario, the method achieves a minimum scheduling delay of 78ms, a task completion rate of 93.1%, a resource cost of 71.79 yuan, and a resource utilization rate of 76.2%. This method significantly mitigates the performance degradation caused by resource constraints, validating its advantages in scheduling intelligence, responsiveness, and service assurance.