Instant solar irradiation forecasting for solar power plants using different ANN algorithms and network models


YÜZER E. Ö., BOZKURT A.

Electrical Engineering, 2023 (SCI-Expanded) identifier

  • Publication Type: Article / Article
  • Publication Date: 2023
  • Doi Number: 10.1007/s00202-023-02067-z
  • Journal Name: Electrical Engineering
  • Journal Indexes: Science Citation Index Expanded (SCI-EXPANDED), Scopus, Academic Search Premier, Compendex, INSPEC, DIALNET
  • Keywords: Algorithms, Artificial neural networks, Forecasting, Solar energy, Solar irradiation
  • Hakkari University Affiliated: Yes

Abstract

Solar irradiation is a crucial parameter in the design and operation of solar energy systems. However, its long-term measurement everywhere is hindered by the maintenance and cost of measurement devices. Therefore, numerous research studies have been conducted to determine solar irradiation, leading to the development of various prediction models. Recently, artificial neural network (ANN) models have been shown to enable researchers to make more accurate predictions. This study aims to identify the most effective algorithms and functions for accurately predicting instantaneous solar irradiation using ANN models with different network structures. Five commonly used training algorithms and two different ANN architectures are examined in this study. These models are tested with various transfer functions, and the impact of the number of neurons in the hidden layer on prediction results is also investigated. Meteorological data collected at 5-s intervals from a meteorology station in Hakkâri Province between 2019 and 2021, totaling one million data points, are used for model training. The ANN model with a network structure consisting of 100 neurons, trained with the Levenberg–Marquardt algorithm and “tansig” transfer function, achieved the best prediction performance with a correlation coefficient (R) of 0.9783 and a mean absolute percentage error of 6.79%. For an 80-10-10 data split, the mean-squared error, normalized root-mean-squared error, and mean bias error were found to be 0.024, 7.206, and 0.800, respectively. The solar irradiation prediction performance varied based on the training algorithm and particularly the transfer functions used. Similar approaches can be employed in regions where measurement devices cannot be installed, enabling successful prediction results even without direct irradiation measurements.