Decentralized forecasting of wind energy generation with an adaptive machine learning approach to support ancillary grid services

The increasing penetration of the power grid with renewable energy (RE) sources comes with an increase of distributed generation units (DGUs), which exhibit a very heterogeneous structure. Grid operators often face challenges when employing RE due to its large volatility. Especially in exceptional situations, e.g. power blackouts or frequency events, conventional generation units are often preferred to stabilize the grid. Making use of highly volatile power sources requires very reliable knowledge about the near future of the available active power (AAP) signal, hence giving rise to the necessity of accurate and robust RE generation forecasting in short time ranges (nowcasting).

In order to maximize the availability of forecast data to the grid operators the forecasting process can be deployed in a distributed fashion. The presented method makes use of several machine learning (ML) algorithms and provides a probabilistic short-term forecast with little requirements to computational resources on-site. The local availability of forecasts improves the AAP signal and hence enables ancillary services from wind farms such as the provision of a frequency containment or restoration reserve. Also it provides robustness against communication failure between grid operation and forecast provider.

The suggested forecasting procedure is two-fold: On the one hand a data-driven nowcasting (from 0 to 6 hours) approach is pursued within decentralized forecast units, that is designed to be deployable on-site of the respective DGUs. This approach employs local sensor data, i.e. active power, wind measurements from the nacelle anemometer and temperature, as well as the azimuth angle of the nacelle. The various ML models in use are adaptively trained on the latest sensor data and produce ensemble forecasts, from which both the minimally available power and the forecast uncertainty can be deduced. The individual model outputs are then combined by an adaptive genetic algorithm.

This purely data-based nowcast is then enriched with a physical forecast based on numerical weather prediction (NWP) model runs with a time horizon of up to 10 days. It considers actual production data from the wind farm as well as turbine specific control behaviour, which leads to excellent forecast quality. This NWP-based forecast is generated at a central computing centre and can be distributed to the forecasting units at the DGUs. On-site the data-based nowcast and the physics-based forecast are combined to provide good reliability for both time horizons.

In our paper we present the method in detail and describe the infrastructure, for which it was developed. We further conduct a performance assessment of the forecasting procedure and provide performance indicators of expected forecast errors and reliability.

Parts of the presented research have been carried out in the joint research project SysAnDUk (FKZ 03EI4004A) funded by the German Federal Ministry for Economic Affairs and Energy.