Analysis of photovoltaic panels performance and power output forecasting based on optimized deep learning technique / Muhammad Naveed Akhter

• Main Author: Muhammad Naveed, Akhter
• Format: Thesis
• Published: 2021
• Subjects: TK Electrical engineering. Electronics Nuclear engineering
• Online Access: http://studentsrepo.um.edu.my/14798/1/Muhammad_Naveed.pdf; http://studentsrepo.um.edu.my/14798/2/Muhammad_Naveed_Akhter.pdf; http://studentsrepo.um.edu.my/14798/
• Institution: Universiti Malaya

Alternative renewable energy sources have a significant contribution to meet the world’s energy demand due to population climax and reduce global warming. Solar energy is a major alternative energy source to generate electricity through photovoltaic (PV) systems. However, the generated PV power is susceptible to unpredictable climate and seasonal factors, which cause an unfavorable effect on the stability, reliability, and operation of the grid. Therefore, proper monitoring of the PV system and accurate forecasting of PV power output is required to ensure the stability and reliability of the grid. The purpose of monitoring the PV systems is to keep the PV system in continuous functional status with improved performance. In the first part of this work, the performance of three grid-connected photovoltaic systems installed at the rooftop of the engineering tower building, University of Malaya, Kuala Lumpur, Malaysia, is evaluated. The grid-connected PV systems are based on poly-crystalline (p-si), mono-crystalline (m-si), and a-si (amorphous silicon (a-si)) technologies. The performance is evaluated on monthly and annual data monitored from January 2016 to December 2019. A comprehensive analysis is conducted on eleven performance parameters: performance ratio, capacity factor, array yield, final yield, PV array efficiency, PV system efficiency, inverter efficiency, AC energy, array losses, system, and the overall losses. Secondly, an hour ahead forecasting of solar power output is performed on an annual basis for the aforesaid three PV systems over the same period (2016-2019), based on forecasting accuracy measurement parameters such as RMSE, MSE, MAE, r and R2. A deep learning method (RNN-LSTM) is proposed and compared with regression (GPR, GPR (PCA)), machine learning (SVR, SVR (PCA), ANN), and hybrid methods (ANFIS (GP), ANFIS(SC), ANFIS(FCM)) for an hour ahead forecasting of PV power output on an annual basis for the whole period. Moreover, Salp Swarm Algorithm (SSA) is used to tune the hyperparameters of the developed deep learning method on an annual basis over four years to enhance its forecasting accuracy and is compared with RNN-LSTM, GA-RNN-LSTM, and PSO-RNN-LSTM. Performance analysis findings show that p-si PV system performs better with a higher annual average (array yield (1309.7 h), array efficiency (12.17 %), and system efficiency (11.33 %)) accompanied by less degradation in almost all performance parameters compared to a-si and m-si PV systems. Moreover, the composite PV system has the potential to avoid 28143.7 kg of CO2 emissions in four years. The forecasting results show that the proposed deep learning technique (RNN-LSTM) has presented lower (RMSE, MSE) and higher (r and R2) compared to other techniques. Moreover, the proposed hybrid method (SSA-RNN-LSTM) is found (19.14% and 21.57%), (15.4% and10.81%) and (22.9% and 25.2%) better in terms of (RMSE and MAE) than developed (RNN-LSTM) for p-si, m-si and a-si PV systems respectively. Furthermore, the proposed hybrid method (SSA-RNN-LSTM) has shown higher R2 and maximum convergence speed compared to GA-RNN-LSTM and PSO-RNN-LSTM. In addition, the proposed deep learning and hybrid models (SSA-RNN-LSTM) are found to be robust and flexible in the prediction of power output for three different PV systems over four years duration.