Accurate Onboard Predictions for Indoor Energy Harvesting using Random Forests

Abstract

Indoor energy harvesting has recently enabled long-term deployments of sustainable IoT sensor nodes. The performance of such systems operating in an energy-neutral man-ner can be optimized by exploiting energy prediction models. Numerous prediction algorithms have been developed, yet they are primarily intended for outdoor (solar) energy harvesting. Indoor environments are much more challenging to predict since the primary energy is very variable. We propose a prediction method based on random forests that is capable of capturing and predicting this variability. It estimates the harvested energy for various locations in different scenarios with high accuracy while only requiring limited resources. We deploy the predictor on a dual processor platform powered by indoor lighting with various sensors including indoor air quality sensors. The predictor executes in 22.2 μs and requires 2.60 μJ to generate a prediction. Furthermore, the predictor continuously learns from the system's local environment. The proposed online learning is resource-efficient and requires only limited data, enabling it to run on the harvesting-based system. Over time, online learning reduces the energy required to generate a prediction by up to 77 % while maintaining its high prediction accuracy.