Multi-Horizon Data-Driven Wind Power Forecast: From Nowcast to 2 Days-Ahead

Generation uncertainty is an obvious challenge posed by renewable energy sources such as wind power with effects spawning from stability threats, to economic losses. Datadriven forecasting methods draw increasing attention due to the amount of data available, flexibility and cost-effectiveness among other factors.

However, there are concerns regarding effective feature selection and tuning of these models since common naive approaches focus on Pearson or Shapley. This papers uses the development of an active power forecaster for a wind turbine to conduct a thorough sensitivity analysis addressing how different sampling rates, machine learning (ML) methods, features and hyperparameters influence accuracy.

Which is computed with the Root-Mean Squared Error and compared against Persistence. The selected ML-methods are Random Forest and Long-Short Term Memory Artificial Neural Networks. The forecasters are multi-horizon & multi-output model targeting 1 minute, 1 hour, 5 hours and 2 days ahead by using sampling rates of 1 second, 1 minute, 5 minutes and 1 hour respectively.

The results show which method is more suitable for which horizon and provides insight into which features reduce RMSE of the best performers, whose average is 10, 13, 17 and 25 % for each horizon respectively. The conclusions of the sensitivity analysis can be applied for regions with highly volatile weather, such as coastal areas.