Abstract
In regions of high solar irradiance concentration such as South Africa, solar
power is considered an accessible form of renewable energy. However,
this source of energy is not constant in supply. Its stochastic nature can lead
to fluctuations in the voltage and frequency in the energy grid. The deep
learning models namely, Long Short-Term Memory (LSTM) and Recurrent
Neural Network (RNN) promise to provide accurate forecasting of the availability
of solar power which is crucial for operational efficiency and distribution
of power. Due to the inaccessibility of accurate and measured solar irradiance
data, we use the novel application of a digital twin (DT) model to generate
a dataset of meteorological attributes. We use the synthetic dataset to
predict the solar output power of the University of Johannesburg (UJ) Auckland
park (APK) campus solar Photovoltaic (PV) plant. The experimental
results indicate that a DT dataset can be regarded as a viable alternative to
measured onsite data. Using Root Mean Square Error (RMSE), Mean Absolute
Error (MAE) and Mean Square Error (MSE) as performance metrics,
the LSTM network outperformed the other models recording 91.36, 54.49,
8347.58 respectively. The second best model is the RNN recording 95.14,
55.67, 9052.22. These findings suggest that both DT’s and the LSTM network
should be utilized for predicting solar power output. This would allow
operators to take proactive measures to mitigate the impact of variability in
solar energy. Thus facilitating a smoother integration of solar energy into the
national power grid.