Comparative Study of Short-Term Photovoltaic Power Generation Forecasting Methods

An accurate forecasting scheme is essential to optimize efficiency when planning a distribution grid with a photovoltaic system. This article proposed a comparative study of the efficiency and advantages between supervised photovoltaic power forecasting methods that can be appropriately applied to various systems. For comparative study, Feedforward Neural Network (FNN) was compared with Long Short-Term Memory (LSTM) and Nonlinear Autoregressive with External Input (NARX) that use the previous output as input via MATLAB program. The forecasting dataset consists of time, cell temperature, irradiance, and power output from the PV system in 8,760 samples (for each hour). The dataset was divided into three parts following: (1) training set is 70%, (2) validating set 20% and (3) testing 10 % The simulation results show that, compared to the three methods, LSTM had the highest accuracy, followed by FNN and NARX, respectively. However, NARX is closer to the peak at high temperatures and irradiance. For data set analysis, the difference in power output ranges between the training and test sets affects forecasting accuracy.