Estimation and aggregation of wind power forecasts utilizing master data and zero-shot learning

The aggregation of wind power forecasts to grid areas, such as transmission system operator (TSO) control zones, provides grid operators with valuable information for the planning of grid operations. In particular, as the share of renewable energies continues to increase, improvements in forecast quality are becoming increasingly important. When forecasts at reference wind parks are projected to all wind parks in total and aggregated to grid areas of known installed capacity, a broad distribution of reference parks is advantageous for accounting for local weather situations. This often conflicts with the availability of measurement data for training accurate models.
In this R&D project, we propose an approach that uses publicly available wind park master data as well as local weather forecasts as input into a zero-shot learning model. This model generates additional, locally distributed forecasts that are used in the projection process to the total of all parks. For reference park forecasts, we employ a combined multi-task and transfer learning approach. In this approach, measurement data is leveraged to train a multidimensional park embedding, which characterizes the park's behavior. This park embedding is part of the model’s input feature set. The zero-shot model is based on the same multi-task learning artificial neural network architecture and is trained on a portfolio of reference parks. Instead of a trained embedding, master data such as hub height and rotor diameter are employed to embed each park into a characterizing embedding space. This approach enables the zero-shot model to generate forecasts at new locations without training when provided with local master data. As a result, it provides a much more comprehensive coverage in the forecast projection process to the total of all parks. The projection is achieved through an inverse-distance-weighting of wind parks to grid areas of known installed capacity, followed by an aggregation to TSO control zones.
Experimental results with field data demonstrate that in terms of forecast accuracy at control zone level, our approach outperforms methods that only use reference wind parks in the process. Additionally, we observe that the zero-shot learning approach enhances forecast quality compared to physical models when generating forecasts at additional locations. Furthermore, we illustrate how our approach can be integrated with optimization techniques regarding the projection process to further improve overall results.