Come meet the poster presenters to ask them questions and discuss their work
Check the programme for our poster viewing moments. For more details on each poster, click on the poster titles to read the abstract.
PO037: Automatic Feature Selection and Forecast Combination to Enhance and Generalize Renewable Energy Forecasting
Georges Kariniotakis, Professor, Head of Renewable Energies & SmartGrids Group, Centre PERSEE, MINES Paris, PSL
It is generally beneficial to spatially aggregate renewable energy sources (RESs) as a virtual power plant (VPP) when participating in electricity markets. Due to the smoothing effect, variability decreases and the time series becomes more straightforward to forecast. A direct consequence of spatially aggregating RESs is an increase in the number features available from potentially different data sources, e.g., numerical weather prediction (NWP), remote sensing or endogenous. The present work developed in the Smart4RES project aims at developing scalable RES forecasting models that efficiently large and heterogeneous input data sets. Feature selection is an important tool in machine learning to relieve the curse of dimensionality and avoid overfitting. However, and similar to when using all available features, there is no guarantee that a single set of features results in calibrated probabilistic forecasts, which is a prerequisite for market participation. To alleviate the uncertainty induced by feature selection, we test several agnostic selection methods and afterwards combine the probabilistic forecasts using linear and nonlinear forecast combination methods. The case study comprises real-world data from a 124 MW VPP that includes photovoltaic systems with a total installed capacity of 4 MW and wind turbines with a total installed capacity of 120 MW, located in mid-west France. The focus is on intra-day forecasts because that is when the heterogeneous input data set is most pertinent. The 831 input features comprise NWP forecasts, satellite-derived irradiance maps and local measurements. We show that the feature selection methods result in a heterogeneous feature set but that the probabilistic forecasts are underdispersed, i.e., the models are too confident based on the selected feature sets. However, the linear and nonlinear forecast combination methods effectively recalibrate the underdispersed probabilistic forecasts. The recalibration is expected to enhance profitability in electricity markets where trading decisions should be based on probabilistic forecasts, e.g., based on the optimal quantile or stochastic optimization in general.