Online machine learning for 1-day-ahead prediction of indoor photovoltaic energy

Enthalten in
IEEE Access 
Volume
11
Start Page
38417
End Page
38425
Citation
IEEE Access 11 (): 38417-38425 (2023)
Publisher DOI
10.1109/ACCESS.2023.3267810
Scopus ID
2-s2.0-85153508910
Publisher
IEEE
We explore the potential for predicting indoor photovoltaic energy on a forecasting horizon of up to 24 hours. The objective is to enable energy management approaches that exploit harvesting opportunities more strategically, for which they require more accurate energy intake predictions. Our study is based on a data set covering over 3 years, for which we simulate online machine learning algorithms with different amounts of training data and input features. Our results show that relatively simple machine learning methods can outperform a persistent predictor considerably, and we observed a reduction of errors of up to 56%. When devices obtain a significant amount of sunlight, adding the weather forecast improves the prediction accuracy. We discuss prediction features, the amount of training data and analyze the sources of errors to understand the potential of indoor photovoltaic energy harvesting predictions.
Schlagworte
energy harvesting
photovoltaic systems
energy predictions
machine learning
DDC Class
004: Informatik
600: Technik
620: Ingenieurwissenschaften