Operational day-ahead photovoltaic power forecasting based on transformer variant.

With the increasing penetration of photovoltaic (PV) power into the grid, the impact of PV power on the stable operation of the grid is becoming increasingly significant. Our aim is to design a multi-step PV power forecasting method to reduce the uncertainty associated with grid-connected PV operation based on the time and accuracy requirements of real-world forecasting scenarios. A novel data augmentation method is introduced and a new model PTFNet (Parallel Temporal Feature Information extraction Network) is designed to achieve more accurate forecasting, and the prediction results are evaluated to match the real scenarios. Firstly, noting that past work has rarely considered the particularities of PV data, which are reflected in the physical modeling of PV, we incorporate site-specific and future temporal information by adding several physical modeling intermediate variables to the raw dataset. The input data consists of past measured data and numerical weather prediction (NWP) data. Then, to better establish the relationship between these two types of data and PV power, the proposed model utilizes a parallel structure to extract the temporal dependency and inter-feature dependency of these two types of data and PV power, respectively, and combines these two types of information for forecasting. Our model can produce 24-, 36- and 40-h forecasts with 15-min resolution, which means that these forecasts meet the forecast horizon requirements of most current day-ahead PV power forecasting scenarios. Transformer architecture-based models have been increasingly used in the field of time series forecasting in recent years, and we selected 10 typical long-term forecasting models for comparison, the first time that all of these models have been applied to a PV dataset. We conducted experiments on two public datasets, and the results show that PTFNet achieves the overall most competitive forecasting ability. Finally, to match the actual forecasting scenarios, we intercepted the last 24 h of the 40-h forecasts as the actual forecasts submitted to the grid operator and post-process the forecasts. The proposed model achieved the best results on all evaluation metrics of the post-processed data and the predictions of the model meet the accuracy requirement of the grid operator in most cases. • Data augmentation considering PV physical modeling improves prediction accuracy. • A Transformer-based day-ahead photovoltaic power prediction model is established. • Explicitly extract temporal and inter-feature dependencies of the data for prediction. • Post-process the predictions to obtain output consistent with application scenarios. [ABSTRACT FROM AUTHOR]