ENERGY-AWARE AGENTIC TASK OFFLOADING IN RESOURCE-CONSTRAINED MEC NETWORKS THROUGH DIGITAL TWINS AND TINYML

In 6G networks, Mobile Edge Computing brings computation close to where data is produced, but deciding which edge node should run each task depends on a state that changes continuously. Reactive strategies intervene only once degradation is evident, while anticipating it requires intelligence whose cost may exceed its benefit on resource-constrained devices. This thesis proposes a framework in which a TinyML classifier running on each edge node turns local metrics into a compact operating state and a confidence value, a Digital Twin Network collects these states with adaptive synchronisation, and an agent is invoked only when needed, selecting nodes through an indicator that balances service quality and energy consumption. The experimental work, carried out on a physical testbed based on a Raspberry Pi 4 with an independent power-measurement chain, covers the node-side part of the framework, while the digital twin, the agent and the node-selection indicator are presented at design level. A 139-parameter multilayer perceptron, trained on 1221 windows collected by replaying realistic IoT traffic, reaches a macro F1 of 0.958 in cross-validation, occupies 556 bytes and runs an inference in about 3 µs. Power measurements show that the intermediate, degraded state is the most expensive, drawing about 25% more than the normal one. On a measured trace, the classifier's confidence signalled degradation 98 s before the rise in power consumption that a reactive, power-based policy would detect. A robustness analysis built on the deployed model indicates that this lead time recurs across simulated repetitions and that the proactive approach remains energetically favourable as long as the agent is not triggered by every isolated signal.