Energy Quantification of Machine learning for Dynamic Spectrum Access in LoRaWAN Device

Low-power wide-area networks (LPWANs), such as LoRaWAN, must balance device lifetime with an increasingly congested sub-GHz spectrum. Machine learning (ML) approaches, such as the Upper Confidence Bound (UCB) multi-armed bandit algorithm, allow for collision-aware channel selection without gateway-level coordination. However, these approaches’ computational and memory requirements could adversely affect device autonomy. This study presents an analysis of the power consumption of the UCB approach on an STM32WL55 LoRaWAN system on a chip (SoC). Current traces were captured for payloads ranging from 10 to 50 bytes and transmission powers ranging from 8 to 16 dBm. The standard algorithm was compared to a UCB algorithm. The results show that the UCB algorithm’s power consumption is less than 1% per uplink-ACK cycle: 0.60% during transmission, and 0.70% and 0.30% during the Rx1 and Rx2 windows, respectively. This overhead becomes negligible as payload size or transmission power increases. These results confirm that lightweight learning is a practical solution for dynamic spectrum access in large-scale IoT deployments.