A machine learning approach to hybrid renewable system power forecasting

Abstract

The unprecedented growth of renewable energy has introduced the negative effect of variability in the Dutch grid. The most important risk associated with this variability is an unstable grid due to the difficulty of matching production and consumption. A novel way of mitigating this risk is by accurate power forecasting with the aid of machine learning.
This research aimed to create insight into the application of machine learning in the space of renewable power forecasting, by comparing hybrid renewable energy system forecasting models with proposed benchmark models. Improved grid planning and value creation could be two potential benefits of this. In this way, the Dutch TSO, TenneT, and commercial operators can profit from an improved forecasting method.
This study proposes two machine learning techniques for the application of power forecasting. Namely, the tree regressor XGBoost (XGB) and a neural network (NN). Both techniques were used for the models of an individual PV plant and wind farm for the 1-hour and 6- hour ahead forecast horizon. Then, the hybrid models were built with the machine learning techniques and were compared to the individual models and Persistence with respect to the normalized root-mean-squared error (NRMSE). Next to the NRMSE the value creation of the models was studied with respect to Persistence on the 6-hour forecast horizon.
The results show that all proposed models outperform Persistence on the 6-hour forecast horizon. Furthermore, the results show that the hybrid models for both XGB and NN outperform the individual models on the 6-hour forecast horizon. Overall, the XGB models outperform the NN models on both the forecast horizons. Lastly, the proposed models all created positive value with respect to Persistence. However, there was a non-linearity between the NRMSE performance and the cumulative value creation over the test period.