What drives the accuracy of PV output forecasts?

• Combining methodologies (hybrid models) achieve the lowest errors. • ML models do not show a robust performance but have the fastest improvement. • Using data processing techniques reduces forecast errors. • Forecast horizon and test set length positively correlate with forecast errors. • The possibility of "cherry picking" in reporting errors is observed. In this paper, 180 papers on photovoltaic (PV) output forecasting were reviewed and a database of forecast errors was extracted for statistical analysis. The paper shows that among the forecast models, hybrid models are most likely to become the primary form of PV output forecasting in the future. The use of data processing techniques is positively correlated with the forecast quality, while the lengths of the forecast horizons and out-of-sample test sets have negative effects on the forecast accuracy. The paper also found that the use of data normalization, the wavelet transform, and the inclusion of clear sky index and numerical weather prediction variables are the most effective data processing techniques. Furthermore, the paper found some evidence of "cherry picking" in the reporting of errors and we recommend that the test sets be at least one year long to avoid any distortion in the performance of the models. [ABSTRACT FROM AUTHOR]