Solar power forecasting: Investigating the robustness of feedforward neural networks to adapt to utility-scale PV-system fault conditions

Photovoltaic (PV) power forecast models are an important solution to reduce the uncertainty associated with PV system power supply, which is characteristically intermittent. Traditionally, machine learning forecast models are trained with a perfect-data representation of PV-system behaviour. However, utility-scale power output can change abruptly due to internal PV-system fault conditions (broken modules, failed inverters, etc.) or routine maintenance. Given this context, there is concern regarding the ability of these perfect-data models to adapt forecasts to a sudden change in power output. The popular Feedforward Neural Network (FFNN) model is used to deliver 1 - 6 h ahead forecasts for a 75 MW rated PV system. With power output systematically reduced by introducing fault conditions, the robustness of this widely used ML-based forecasting solution is evaluated. From the overall results, the FFNN model was shown to maintain an impressive degree of robustness for scenarios where less than 10 % of rated PV-system capacity was lost.