Inference of Missing PV Monitoring Data using Neural Networks

Complete photovoltaic monitoring data are required in order to evaluate PV system performance and to ensure confidence in project financing. Monitoring sub-system failures are common occurrences, reducing data availability in meteorological and electrical datasets. A reliable backfilling method can be applied in order to mitigate the impact of long monitoring gaps on system state and performance assessment. This paper introduces a method of inferring in-plane irradiation from remotely obtained global horizontal irradiation, by means of a neural network approach. Generation output is then calculated utilizing a simple electrical model with fitted coefficients. The proposed method is applied to a UK case study for which the mean absolute error in monthly system output was reduced significantly, to as low as 0.9%. This yielded more accurate results in backfilling the missing datasets when compared to standard approaches. The impact of missing data on monthly performance ratio is also investigated. Using backfilling to synthesize lost data increases performance ratio prediction accuracy significantly when compared to simply omitting such periods from the calculation.