Dynamic Resource Allocation (DRA) has been identified as a paramount concept in Sustainable Cloud Sciences because the performance of modern cloud systems in terms of computation and power consumption is not only challenged but also demands of exponentially growing service demands are being faced. The allocation of resources can become efficient, which ensures the maximum consumption of computing resources, elimination of waste of energy and the quality of the delivered services can be ensured despite the alternating loads. However, stagnant or rule-based allocation mechanisms are unable to maintain the dynamism and uncertainty of clouds environments, and cause performance bottleneck and high latency under high energy costs. In order to overcome these challenges, a Deep Reinforcement Learning-Based Dynamic Resource Allocation (DRL-RA) model is proposed in this paper and it can learn the most optimal allocation policies through smart interaction with the environment. The special agent of deep reinforcement of the proposed model dynamically monitors and allocates computing resources such as CPU, memory bandwidth with the real-time change in the demand. The framework is able to trade between accuracy, energy efficiency, latency, resource utilization, and the SLA compliance with the assistance of deep neural representation and policy optimization. The decision making process in the system is normally adaptive which allows the system to enhance its performance over time as it can adjust to experience without necessarily involving the manual processes. It has been experimentally demonstrated that the proposed DRL-RA model is far better than the existing deep learning models such as CNN, LSTM, DQN, AE, GAN, RNN, and GNN in terms of a number of performance metrics.