Artificial neural network model for solar resource assessment: An application to efficient design of photovoltaic system

• Main Authors: Santiago, Robert Martin C., Bandala, Argel A., Dadios, Elmer P.
• Format: text
• Published: Animo Repository 2017
• Subjects: Solar energy; Solar radiation; Photovoltaic power systems; Renewable energy sources; Neural networks (Computer science); Electrical and Computer Engineering; Electrical and Electronics
• Online Access: https://animorepository.dlsu.edu.ph/faculty_research/1924
• Institution: De La Salle University

The power output of solar energy conversion facilities such as photovoltaic systems is highly dependent and proportional to the amount of solar radiation absorbed on the collecting surface. In order to have an efficient design of these systems, it is essential to perform solar resource assessment on the intended location prior to installation. Advancements in computational intelligence led to applications of artificial neural networks for solar resource assessment which outperforms existing empirical models in terms of speed and accuracy and overcomes the cost of using expensive solar radiation sensors. In this study, a single recurrent or feedback network is developed and assessed for efficacy in estimating the daily sum of solar radiation in the Philippines using meteorological data such as daily sum of sunshine duration, daily mean air temperature, daily mean air pressure, and daily mean air humidity. The collected data used in this study were obtained for the year 2014 from the Bureau of Soils and Water Management (BSWM) Agro-meteorological Station Lufft sensors in three locations: (1) Tanay, Rizal, (2) Barili, Cebu, and (3) Sto. Tomas, Davao del Norte. The developed model responded with mean squared error (MSE) values of 0.1491, 0.1679, and 0.2297 and regression values of 0.9146, 0.9313, and 0.9277 for the training, validation, and testing phases. The error histogram also shows that low values of error exist for each dataset and most errors fall between the ranges of -0.4581 to 0.5646. Results may further be improved by having larger data for training, validation, and testing phases for the neural network which can make the model more robust for larger variations in the weather patterns. © 2017 IEEE.