login

Search documents

Browse topics

Document details

 
Title:
 
Recurrent Neural Network for Short-Time Power Forecast for a 540 kWp Grid-Connected PV Plant Installed at the University of São Paulo, Brazil
 
Author(s):
 
W.W. Ferreira Fonseca, F. Ramos Martins, R. Zilles
 
Keywords:
 
Artificial Neural Network, PV Power Forecast, Machine Learning
 
Topic:
 
PV Systems – Modelling, Design, Operation and Performance
Subtopic: Solar Resource and Forecasting
Event: 38th European Photovoltaic Solar Energy Conference and Exhibition
Session: 5BV.4.4
 
Pages:
 
1125 - 1129
ISBN: 3-936338-78-7
Paper DOI: 10.4229/EUPVSEC20212021-5BV.4.4
 
Price:
 
 
0,00 EUR
 
Document(s): paper, poster
 

Abstract/Summary:


A RNN model was developed to predict up to three hours ahead of the production of a 540 kWp gridconnected PV plant installed at the Energy and Environment Institute in São Paulo, Brazil. The RNN's predictors are meteorological and environmental data and a set of power output data of two subset PV plants of 156 kWp each. Two RNN models were set concerning the two typical climate seasons - dry and rainy months. The aerosol optic depth data provided by MERRA-2 reanalysis model was included as a predictor. A brute force algorithm determined the architecture with three hidden layers and thirty neurons in each layer. We evaluated the accuracy of the PV power forecast through Mean Squared Error (MSE), Nominal Root Mean Squared Error (NRMSE), and Skill Score (SS) comparing to a persistence model. Both forecast models resulted in a good fit comparing observed and forecasted data, with NRMSE ranging from 8.01 to 13.88 kWh. The method applied in this work consistently outperformed the persistence model for all evaluated scenarios, especially during rainy months. SS values observed range from 0.72 to 0.91. Finally, the inclusion of aerosol data has slightly improved the forecast model performance during dry months.