Dynamic cloud resource scheduling needs real-time adaptation to changing workloads to keep system performance high and stable. Traditional methods like FCFS and RR lack the ability to adjust resources dynamically in complex conditions. This paper presents DynaSched-Net, a dual-network framework that uses a Deep Q-Network (DQN)-based reinforcement learning scheduler and a hybrid LSTM-Transformer predictor. The reinforcement learning module assigns resources based on system states to improve load balance. The predictor learns short-term and long-term workload patterns to guide decisions. A joint loss function helps optimize both parts of the system. Stabilization methods like experience replay and target network updates help keep training stable. Experiments show that DynaSched- Net performs better than traditional methods and provides an efficient way to manage cloud resources.