A smart short-term solar power output prediction by artificial neural network


ERDUMAN A.

Electrical Engineering, vol.102, no.3, pp.1441-1449, 2020 (SCI-Expanded) identifier

  • Publication Type: Article / Article
  • Volume: 102 Issue: 3
  • Publication Date: 2020
  • Doi Number: 10.1007/s00202-020-00971-2
  • Journal Name: Electrical Engineering
  • Journal Indexes: Science Citation Index Expanded (SCI-EXPANDED), Scopus, Academic Search Premier, Compendex, INSPEC, DIALNET
  • Page Numbers: pp.1441-1449
  • Keywords: Artificial neural network, Modeling photovoltaic module, PV power prediction, Site matching
  • Hakkari University Affiliated: Yes

Abstract

Regional analyses in solar energy applications take a long time, and the complexity of production modeling of solar panels is a problem for small roof applications. This is because the people who will implement these applications do not have sufficient budgets to establish pyrometers and similar expensive measurement systems, or they do not have enough technical information to analyze them. However, other meteorological data, such as wind, temperature, and humidity, have been recorded for a long time by various institutions around the world, and these data can be accessed. Within the scope of this article, an artificial neural network-based model has been developed, which requires cheaper and less information for the users who will implement the roof-top application. In an area where temperature, wind speed, pressure, and humidity information is given, using the developed neural network learning model, the output power of the most commonly used mono-crystalline, poly-crystalline, and thin-film solar panels was calculated. In order to be used in the modeling of the study, the production values of the three different panels and wind, temperature, and humidity measurements of the region where the panels were installed were recorded in periods of 1 min. The relationship between these data was modeled by artificial neural networks. To increase the reliability of the data, training and test data were used, obtained as a result of the tenfold cross-validation method. Using the developed model, production values were estimated to have 97% accuracy. As a result, a simpler and more reliable analysis model has been developed without the use of radiation and panel information, without the need for PV and system modeling, which requires long mathematical processes.