Artificial Neural Network Model for Forecasting Distributed Solar Photovoltaic Generation

Rooftop photovoltaic (PV) generation has been playing an increasingly important role in the Brazilian Interconnected Power System, with significant expansion expected in the coming years. Since only the net load (load–PV generation) is measured, the actual amount of generation is not known by the operator and, as its share increases in the system, its variability must be individualized. Models based on irradiance measured at the closest meteorological station are used to estimate the generation, and PERSC [1] has been used in Brazil, with a relevant level of inaccuracy due to the distance between the station and the generation site.

In this context, this work proposes the use of an artificial neural network (ANN) model for rooftop PV generation forecasting [2] that uses data from four stations and the generation data from PV plants. The meteorological data (including solar irradiance, temperature, relative humidity, and wind speed, among others) are obtained from the Brazilian National Institute of Meteorology.

Initially, individual ANN models were developed using data from utility-scale PV plants to validate the model. Subsequently, a generalized ANN model was created and trained using data from multiple utility-scale PV plants grouped by geographic region, aiming at the forecasting of rooftop PV generation.

Forecasts covering 8,760 hours (one year) were carried out using the individual ANN models for the Pirapora, Bom Jesus da Lapa, and Dracena PV plants. Additionally, rooftop PV generation was estimated for a case with measured generation data using the generalized ANN model.

The performance of the ANN models was compared with the PESRC estimation by means of the statistical metrics R² and RMSE [3] with respect to the measured generation. In general, the results demonstrated superior performance of the ANN models. As future work, the explicit consideration of inverter clipping effects is being investigated.