Short-Term Photovoltaic Power Forecasting Using a Neural Network Dynamic Time Series

Currently, short-term photovoltaic (PV) power forecasting is crucial for optimizing PV power operation in both off-grid and grid-connected systems. This paper compares different PV power forecasting models: Nonlinear Autoregressive (NAR), Nonlinear Input-Output (NIO), and Nonlinear Autoregressive with External Input (NARX) for a PV rooftop 23.1 kWp. The primary input data consists of solar irradiance and temperatures from 1 to 6 hours before the present time. The training process utilizes one year of data (2019) to develop the learning model. Data from January 2020 to April 2020 were used to test the models. This paper reports the optimized time series and number of neurons, with MAPE and R2 used as accuracy indicators for evaluating the models. The best model, NARX achieved MAPE=6.84% and R2=0.9 with a time series delay of 3 hours and 50 neurons.