Traditionally, servers in high performance clusters are always keeping in active state to process the incoming tasks. However, it may cause energy wastage when there are few requests, or service rejections when there come bursts of requests exceeding the service capacity of the cluster. In this paper, we propose an online scheduling of the tasks and servers through dynamically scaling the computing capacity by leveraging DVFS and On/Off switching of servers in cloud environment to minimize the total energy consumption. The stochastic nature of task arrivals complicates decisions regarding when and how many servers to switch On/Off, while the energy consumption and delays associated with these transitions cannot be ignored. Although DVFS can instantaneously increase processing speed, it may also significantly increase energy consumption. To address these challenges, we first introduce a heuristic algorithm called ETA (Energy-aware Task Allocation, which allocates tasks to the servers with the least incremental energy consumption in priority of their urgency, while minimizing the number of active servers. We then propose a DQN-based algorithm to further optimize On/Off switching by intelligently incorporating predicted task numbers as environmental states using LSTM. Numerical experiments demonstrate that our approach significantly reduces energy consumption in cloud data centers.