Impact of Data Quality on Day-ahead Photovoltaic Power Production Forecasting

The main challenge of integrating large shares of variable generation remains the accuracy improvement of day-ahead forecasts and the establishment of applicable robust performing practices based on acquired data. To that end, the quality of the data utilized for constructing and verifying the performance of forecasting models is an important influential factor. This work aims to present the impact of data quality and different training regimes on the performance of day-ahead photovoltaic (PV) power production forecasting models. Specifically, a comparative performance evaluation was performed by training an optimally constructed forecasting model with high- and low-quality data acquired from a test-bench PV system. The forecasting model trained with high-quality data exhibited an absolute error reduction of 3.25% compared to the same model trained with low-quality data. Finally, the application of the proposed data correction methodology demonstrated that training with high-quality data provides higher performance accuracy improvements of up to 2.98% for overcast daily patterns.