No abstract

No abstract

For sustainable Internet of Things (IoT) systems, solar-power prediction is an essential element to optimize performance, allowing devices to schedule energy-intensive tasks in periods with excess energy. In regions with volatile weather, this requires taking the weather forecast into account. The problem is how to provide such solar energy predictions with high accuracy for large-scale IoT systems with various devices in an autonomous way, without manual adaptation effort. We present a detailed study on machine-learning approaches for the prediction of solar power intake for large scale IoT systems. We examine which machine learning models, feature sets and sampling rates gain the best results for a medium-term forecasting horizon. We also explore an operational setting in which devices are deployed without prior data and machine learning models are re-trained for each sensor continuously as a form of online learning. Our results show that prediction errors can be reduced by 20 % compared to the state of the art, despite strong weather volatility.

Device management can enhance large-scale deployments of IoT nodes in non-stationary environments by supporting prediction and planning of their energy budget. This increases their ability for perpetual operation and is a step towards maintenance-free IoT. In this paper we consider how to accelerate the collection of relevant training data for nodes that are introduced into an existing deployment to increase the accuracy of their predictions. In particular, we investigate how nodes powered by solar energy can learn their energy intake faster and more accurately by using data from selected nodes that are working in similar conditions. We explore an architecture that utilizes different training data selection policies to manage the learning processes. For validation, we perform a case study to explore how nodes with correlated data can contribute to the learning process of other nodes. The obtained results indicate that this approach improves the accuracy of the predictions of a new node by 14 %.

No abstract