Enhancing power distribution system performance through machine learning-based forecasting of solar power generation

Abstract

The expansion of photovoltaic (PV) energy has accelerated in recent years, and many large-scale solar power plants are being installed. Due to the intermittent and uncertain nature of PV power plants and all other renewable energy sources, it is essential to find methods that can accurately predict their output power. Machine learning (ML) has become a valuable tool for predictive maintenance, system optimization, power forecasting, and fault detection in renewable energy power plants. More specifically, parametric models for PV power forecasting depend on mathematical relationships that need extensive input data and parameter determination. In contrast, nonparametric models can capture complex nonlinear relationships directly from historical data without the need for parameters or a fixed functional form. Nonparametric ML models have been used in literature for power forecasting but never for large-scale PV farms. This work compares the implementation of ML and deep learning nonparametric ML models on the power forecasting of large-scale PV farm in Egypt, taking the Kom Ombo 26 MW power plant as a case study. The linear regression model resulted in the lowest root mean square error value and was further improved after removing the night hours from the dataset. However, the long short-term memory (LSTM) model proved to be the best model when dealing with historical records of the power plant. Finally, the LSTM model was used to predict the PV output power to choose the maintenance day of the plant. The LSTM model results was implemented on the standard IEEE 69 bus system utilizing the PV power output as a form of distributed generation (DG). The integration of the DG was carried out using particle swarm optimization algorithm for the optimal allocation of the PV power plant on the different busses in the network which resulted in substantial power loss minimization and profit savings.