Prediction of grid-connected photovoltaic performance using artificial neural networks and experimental dataset considering environmental variation

Photovoltaic (PV) investment requires a feasibility study of the PV system in terms of environmental parameters at the location, which is the implementation time and cost. In this study, a 1.4 PV system was installed in Sohar, Oman and the system recorded data, which was modelled using an artificial neural network (ANN). The contribution of this study is to use three proposed ANN models (MLP, SOFM, and SVM) to predict similar systems in twelve other locations throughout the country based on measured solar irradiance and ambient temperature in these locations. The experimental results of Sohar show feasible values of 6.82 A, 150-160 V, 800-1000 W, and 245.8 kWh, peak current, voltage, power, and energy, respectively. Also, the proposed models show an excellent prediction with less error and high accuracy. Furthermore, statistical and sensitivity analyses are presented with a comparison of results found by researchers in the literature for validation. The lowest RMSE was found for SOFM (0.2514) in the training phase compared with (0.2528) for MLP and (0.2167) for SVM. The same sequence but with a higher accuracy was found for SOFM (95.25%), while (92.55%) and (89.19%) for MLP and SVM, respectively. In conclusion, the sensitivity analysis shows that solar irradiance has more effect on the output compared with ambient temperature. Also, a prediction of PV output for Duqm was forecasted till 2050, where it is found insignificant deviation due to climate change compared with 2020.