Low-cost sensors for indoor PV energy harvesting estimation based on machine learning

Abstract

With the number of communicating sensors linked to the Internet of Things (IoT) ecosystem in-creasing dramatically, well-designed indoor light energy harvesting solutions are needed. The first step towards this development is to determine the harvestable energy in real indoor environ-ments. But the harvestable energy varying over time with nature (spectra) and intensity of the light multi-sources, lighting data must be collected for sufficiently long periods. Besides, for a real implementation on-site, studies must be able to be carried out simultaneously in several places to determine locations with the highest energy harvesting potential. In this context, this manuscript presents a very low-cost prototype based on commercial photodiodes (rather than very expensive spectrometers), which measures only a very rudimentary number of spectral data. Thanks to a classification supervised machine learning from Matlab, in which an algorithm learns to classify new observations, and thanks to a simple principle of the superposition approximation model de-veloped for flexible GaAs solar cells, our harvestable energy estimation error is less than 5 percents after more than 2 weeks of observation. To measure this error, the data collected leading to an estimate of the harvestable energy is compared to what has been experimentally harvested in a real IoT system Li-ion battery and compared to what has been estimated using an expensive spectrometer during the same period. Our prototype should allow the development and the massive deploy-ment of a new generation of low cost indoor light energy harvesting sensors for future reliable indoor energy harvesters.