Automating day-ahead forecasting of photovoltaic power generation: Model design, monitoring, and adaption

Abstract:
Reliable PhotoVoltaic (PV) power generation forecasting is critical to the efficient operation of future energy systems. Designing PV power generation forecasting models incorporate three major challenges: i) missing meta-information about the tilt and azimuth angles, ii) missing or limited historical data for training the model (cold-start problem), and iii) changing generation capabilities during operation (so-called concept drifts), e.g., due to soiling, aging or technical defects. To tackle these three challenges, a new automated method for day-ahead PV power generation forecasts is introduced, which is an enhancement of AutoPV. AutoPV is based on creating an ensemble of forecasting models trained with historical data from different PV plants with various tilt and azimuth angles to address the first and the second challenge. For a new PV plant, all models in the ensemble are initially weighted equally. When first PV power generation data is available, the ensemble is adapted by re-weighting the models to minimize the forecasting error. In this way, only the PV plant’s peak power rating is required, as tilt and azimuth angles are implicitly reflected in the ensemble weights. To address the third challenge, the enhanced AutoPV also estimates the PV plant’s efficiency, which allows adaption to drifting generation capabilities. The performance of AutoPV is evaluated on real-world data, simulating a cold-start of the forecast. AutoPV achieves comparable performance to a non-cold-start model that uses two years of historical training data. In addition, the online adaptability of AutoPV is evaluated on data with synthetically inserted concept drifts, outperforming an incrementally trained model. The automated design and operation of AutoPV combined with the cold-start capability are vital for real-world applications to keep pace with the expansion of PV power generation capacity required for the energy transition.