Analysing and Improving Robustness of Predictive Energy Harvesting Systems

Abstract

Internet of Things (IoT) systems can rely on energy harvesting to extend battery lifetimes or even to render batteries obsolete. Such systems employ an energy scheduler to optimize their behavior and thus performance by adapting the node operation. Predictive models of harvesting sources, which are inherently non-deterministic and consequently challenging to predict, are often necessary for the scheduler to optimize performance. Therefore the accuracy of the predictive model inevitably impacts the scheduler and system performance. This fact has been largely overlooked in the vast amount of available results on energy management systems. We define a novel robustness metric for energy-harvesting systems that describes the effect prediction errors have on the system performance. Furthermore, we show that if a scheduler is optimal when predictions are accurate, it is not very robust. Thus there is a tradeoff between robustness and performance. We propose a prediction scaling method to improve a system's robustness and demonstrate the results using energy harvesting data sets from both outdoor and indoor scenarios. The method improves a non-robust system's performance by up to 75 times in a real-world setting.