Evolutionary Program Ensembles for Probabilistic Wind Power Forecasting
Forecast skill for wind using current methods is not particularly high, with state-of-the-art mean absolute error (MAE) relative to power capacity of approximately 10% (i.e., if the wind forecast for a 150 MW wind farm calls for 30 MW of generation, the actual observed power may fall in the range of 15-45 MW, 10% of capacity but a 50% power error). Additionally, forecasts of transitions associated with the passage of synoptic waves can be problematic owing to timing errors. The operational context in which these forecasts are used requires that each morning, an assessment of the next day (midnight-to-midnight) power demand and generation be obtained. Market bets, which can work in either direction, and especially market penalties incentivizes much higher accuracy than has currently been demonstrated.
However, the importance of larger-scale atmospheric organization to mid-latitude wind systems in which diurnal and local scale adjustments take place suggests an opportunity to better quantify forecast uncertainty and to improve prediction. For example, the position of a front will be sensitive to synoptic observational error, while the resulting spatial dislocations can determine whether and especially when a particular area will be affected. This real-world uncertainty can be assessed and calibrated using both forecast and observed regional measures.
We use direct measures of wind power at the individual turbine level at three geographically distant sites representing 337 MW of capacity, along with standard meteorological observations and forecasts, to train an ensemble of evolutionary programs to predict the expected farm output and associated confidence intervals at hourly intervals for the next day (midnight-to-midnight) period. Preliminary results from this analysis will be presented.