Using Artificial Intelligence on environmental data from Internet of Things for estimating solar radiation: Comprehensive analysis
| Autor: | Damir Ivankovic, Ivana Nizetic Kosovic, Toni Mastelic |
|---|---|
| Rok vydání: | 2020 |
| Předmět: |
Pyranometer
Artificial neural network Mean squared error Renewable Energy Sustainability and the Environment business.industry Computer science 020209 energy Strategy and Management 05 social sciences Feature selection 02 engineering and technology Solar energy Industrial and Manufacturing Engineering 050501 criminology 0202 electrical engineering electronic engineering information engineering A priori and a posteriori Artificial intelligence business Solar radiationSoft sensorsMachine learningHybrid modelInternet of thingsSustainable environment Wireless sensor network 0505 law General Environmental Science Efficient energy use |
| Zdroj: | Journal of Cleaner Production. 266:121489 |
| ISSN: | 0959-6526 |
| DOI: | 10.1016/j.jclepro.2020.121489 |
| Popis: | Solar radiation measurements are highly important for achieving energy efficiency in smart buildings as well as solar energy production. They are commonly acquired with pyranometer sensor device. However, due to its high initial and maintenance costs it is not densely deployed in the field. Consequently, it provides only limited coverage as a data source for solar radiation. Hence, theoretical, empirical and/or data-driven models are utilized to estimate solar radiation in areas without pyranometers using only data from meteorological sensor stations, which on the other hand are widely available and obtained from sustainable sensor networks. In this paper, end to end process is described for building hybrid models for solar radiation using Artificial Intelligence (AI), or more specifically Machine Learning (ML) methods, after which a detailed analysis is performed on (1) the accuracy of the models regards to their parameters and input features, (2) the sustainability of the models in the real world, and finally (3) their feasibility in (near) real-time monitoring. The results are expressed with relative root mean squared error (RRMSE) and they show that hybrid models outperform model- and data-driven ones, with artificial neural network giving the best results (RRMSE = 0.0393). Additionally, the models can be enhanced by performing an informed feature selection, where a posteriori selection proves to be better than a priori selection (RRMSE = 0.0371). Further investigation shows that randomly selected input data gives faster model convergence as expected. However, sequential input data can match it if model training starts with autumn or spring data when weather exhibits sufficient variety. When applied on different times scales, all models perform best on 3 h scale rather than daily, where random forest (RRMSE = 0.0275) outperforms neural network (RRMSE = 0.0315). However, for (near) real time usage the models perform almost the same as for daily, with RRMSE of 0.0469 for 1 m i n scale with neural network. This demonstrates the feasibility of the hybrid models in Internet of Things (IoT) applications, which commonly require at least hourly intervals for solar radiation. |
| Databáze: | OpenAIRE |
| Externí odkaz: |
|
Full Text from ScienceDirect