Practical Day-Ahead Power Prediction of Solar Energy-Harvesting for IoT Systems

Technological advances in recent years have led to the widespread use of Internet of Things (IoT) devices. The available energy for these devices, often powered by energy harvesters, is limited. Consequently, there is a need for power monitoring and management, which can by facilitated by predicting the power at the input of these devices. In this work, real-time forecasting models, which are suitable for day-ahead power prediction, in battery-operated IoT devices are investigated. The forecasting models are required to have a low computational cost in order to prevent a significant increase in the energy consumption of the device. Both statistical and machine learning (ML) models are investigated for this purpose, and three models are implemented on a microcontroller. These models include a Naive Prediction model, a Moving Average model, and an appropriate neural network in an effort to minimize the overhead of the models in power. Experimental results including runtime, memory and power, indicate that improving the accuracy of predicting the harvested energy through neural networks entails considerable power requirements, eliminating any additional savings in energy enabled by the higher prediction accuracy. This conclusion can not be reached through theoretical explorations and, consequently, this work offers useful insight in the practical implementation of predicting solar energy supplying IoT nodes.