Configuring artificial neural networks for the prediction of available energy in solar-powered sensor nodes

The behavior prediction of solar energy harvesting systems requires accurate system models in order to dimension the system with respect to its application and location constraints. In contrast to commonly used equivalent circuit models, artificial neural networks (ANN) allow for the behavior of the entire system to be captured in an efficient manner. In this work, we have investigated the influences of the underlying ANN configuration on the model's prediction performance. It was found that the performance variation between training runs increases with an rising number of neurons, which can lead to a higher model performance, but makes the performance outcome more sensitive to initial training conditions. Moreover, the results demonstrate that even simple ANN configurations capture the system behavior accurately and result in low prediction errors for the harvesting architecture under test.