Trading-Off Performance and Sustainability in Internet of Things: An Uncertainty-Aware Hierarchical Energy Management Approach

Internet of things (IoT) devices have been heralded as a key technology to enable several high-impact applications, such as healthcare, remote environmental monitoring, and digital agriculture. However, they suffer from energy sustainability challenges due to small form-factor and batteries. Recent research has proposed to integrate energy harvesting (EH) technologies into IoT devices. However, ambient sources are highly stochastic in nature, which introduces new challenges in IoT energy management (EM) since the decisions must account for the associated uncertainty. To this end, this paper presents a multi-target conformal prediction algorithm to provide provably correct uncertainty estimation of future energy prediction for multiple future time steps. Prediction of energy alone is not sufficient since we must use the predictions along with uncertainty estimates in an energy management algorithm to make energy consumption decisions. A common approach in IoT devices is to follow energy-neutral operation (ENO), whereby the energy consumed in a given period is equal to the harvested energy. However, this may lead to low performance in cases where the energy harvest is insufficient. To resolve this, we propose to design a suite of policies that can trade-off performance and energy sustainability. Moreover, the proposed approach also includes a charging management policy that aims to introduce additional energy into the system when the energy harvest is low to maintain system performance. Evaluations with two real-world datasets show that the proposed conformal prediction algorithm achieves the desired coverage level of 90% for all users. The energy management policies achieve performance that is within 10% of an optimal Oracle that has the knowledge of future energy harvest. Moreover, evaluation of the charging management policy shows that it provides additional energy only when EH values are low. Finally, measurements on an IoT device show that the proposed algorithms are light-weight and consume less than 10 mJ per decision.