A hybrid ANN for output power prediction and online monitoring in grid-connected photovoltaic system / Puteri Nor Ashikin Megat Yunus

• Main Author: Megat Yunus, Puteri Nor Ashikin
• Format: Thesis
• Language: English
• Published: 2023
• Subjects: Neural networks (Computer science); Photovoltaic power systems
• Online Access: https://ir.uitm.edu.my/id/eprint/88921/2/88921.pdf; https://ir.uitm.edu.my/id/eprint/88921/

This thesis presents a new algorithm for output prediction and online monitoring in a Grid-Connected Photovoltaic (GCPV) system based on an artificial neural network. Initially, Multi-Layer Feedforward Neural Network (MLFNN) models for the prediction of total AC power output from a grid-connected PV system have been considered. Three models were developed based on different sets of inputs. It utilizes solar irradiance (SI), ambient temperature (AT) and module temperature (MT) as its inputs. However, all three models utilize similar type output which is total AC output power (PAC) produced from the grid-connected PV system. After that, a hybrid of MLFNN with other optimization methods was introduced, i.e. Improved Fast Evolutionary Programming (IFEP), Evolutionary Programming- Dolphin Echolocation Algorithm (EPDEA) and Evolutionary Programming-Firefly Algorithm (EPFA). The comparison between IFEP, EPDEA and EPFA was compared to determine which model performs better for single-objective optimization. The EPDEA model showed the best in terms of fitness solutions. The results showed that EPDEA scheme provides accurate prediction by producing the highest coefficient of determination, R 2 and the lowest Root Mean Square Error (RMSE). Lastly, the output power performance of a GCPV system progressively monitor at specific interval. The actual data of SI, MT, AT and PAC from the server has been called and uploaded every five-minute interval into Matlab by using File Transfer Protocol (FTP) coding. At this stage, the hybrid EPDEA was selected to be used in the system as it was the best optimization in the hybridization method. All data were then compared to the predicted data that have been developed in the training process, leading to the identification of possible fault in the system. Any predicted AC output power less than the threshold set up, indicates an error has been occurred in the system. The obtained results show that the proposed technique can monitor system under-perform as fast as five minutes. Therefore, the proposed of the system for prediction and online monitoring are justified.