With the advent of 5G networks, cities will become more connected and intelligent, and the Internet of Things (IoT) will become an increasingly present reality in society's daily life. Meeting the diverse service requirements with very specific business profiles and SLAs in a shared infrastructure will be a major challenge for telecom operators. Network slicing is a key technology for 5G networks and beyond, as it enables the logical partitioning of the network to meet different QoS requirements. This work proposes an autonomous system for dynamic resource allocation between network slices based on machine learning. The Random Forest model was used to classify IP flows, obtained from a real network dataset, and achieved an accuracy of 98.6%, making it possible to map the most used applications in each Network Slice. Furthermore, LSTM and GRU recurrent neural networks were applied to predict the memory and CPU resources required to serve each group of applications, using a dataset containing real data from a cloud provider. The system was implemented by simulating three network slices and successfully allocated the resources needed to serve each of them, proving to be a promising solution for efficient resource allocation in 5G environments.