Predicting Solar-Harvested Energy for Resource-Constrained IoT Devices Using Machine Learning

Rakhat Khamitov, Daniil Orel, Dimitrios Zorbas

Abstract: In this paper, the problem of forecasting solar-harvested energy on resource-constrained Internet of Things (IoT) devices using Machine Learning (ML) and light intensity sensors is studied. The unique aspect of this study lies in the correlation between time, ambient light sensor measurements taken between sleep-time intervals, and harvested power, allowing for the prediction of energy production without relying on weather or positioning data. The research involves creating a dataset by studying this relationship. Subsequently, various ML models such as Quantile Regression, Random Forest, Gradient Boosting, and Multi-Layer Perceptron are employed to predict harvested energy. The results demonstrate that the Random Forest and Gradient Boosting algorithms deliver the most accurate predictions, as evidenced by their Symmetric Mean Absolute Percentage Error (sMAPE) values. For a 30-minute interval between measurements, both algorithms achieved a sMAPE of 21.49%, and for a 5-hour interval, the values were 32.53% and 36.33%, respectively. These findings have been validated using a real prototype built on ESP32 microcontrollers. When compared to the ground truth data, the algorithms exhibit very acceptable sMAPE scores of 16.67% and 16.56%, respectively.