Collaborative duty cycling strategies in energy harvesting sensor networks

Energy harvesting wireless sensor networks are a promising solution for low cost, long lasting structural health monitoring applications. Many common applications of these networks are fundamentally concerned with detecting and analysing infrequently occurring events. To conserve energy a subset of nodes can assume active duty, listening for events of interest, while the remaining nodes enter low power sleep mode. Judicious planning of the sequence of active node assignments is needed to ensure i) the active nodes can successfully detect events, ii) as many nodes as possible can be reached upon detection, and iii) the system maintains capability in times of low energy harvesting capabilities. We propose a novel deep reinforcement learning agent which acts as an autonomous centralised power manager for the system. We develop a simulation environment to emulate the behaviour of an energy harvesting sensor network. We then train the proposed agent to learn optimal node selection strategies through interaction with the simulation environment. The performance is tested on unseen historical solar energy data. The agent is shown to outperform baseline approaches on both seen and unseen data. Analysis and visualisation of the system behaviour is provided to illustrate the learned strategies.