<p>The availability of the 5G technology has boosted the demand for internet services, increasing the need for resource management in cellular networks. In order to achieve the best results in the networks, Internet usage forecasting is of great importance. This paper proposes a new machine learning and optimization-based approach for predicting Internet traffic in cellular networks. Our technique uses the Relief algorithm to identify the important features that are relevant to internet usage, and Particle Swarm Optimization (PSO) for building an optimized Emsemble model made by Deep Belief Networks (DBNs). Using PSO to search the solution space effectively, we find the best configuration and weights for each DBN in the ensemble model. In comparison with the conventional methods, this ensemble model improves the accuracy of the predictions with a mean absolute error of 6.91, which is at least 19.73% better than previous methods. In this way, our method helps to enhance the quality of service for the users of the cellular network by optimizing the resource allocation depending on the predicted usage of the network resources.</p>