Posters

Abstract

Accurate forecasts of expected wind power generation and the associated variability are now crucial to operate in the day-ahead power market. Given the recent increase in intermittent renewable capacities, market participants are now more exposed to risks associated with significant forecast-on-forecast changes. The incurring losses or missing generation force market participants to resort to the balancing markets like never before.  Historically, wind power generation forecast models solely use weather variable as inputs, e.g. wind speeds, air surface temperature, air pressure, aiming at forecasting expected wind generation conditioning on the most recent weather forecasts. In this framework, wind power generation variability is forecast by running the model on each trajectory of an ensemble weather forecast (usually 50 or 100), computing confidence intervals ex-post. This – ‘bottom-up’ – technique explains wind power generation variability only by leveraging the variability of its weather inputs, and it is completely exposed to eventual misrepresentations of the inputs’ variability from the ensemble weather forecast it is based off.  We find that confidence intervals built with this technique significantly under-represent wind power generation variability, i.e. they are narrower than they should be, even for relatively short forecast horizons, such as day-ahead, for example. To solve this issue, we propose the use of quantile regression – ‘top-down’ – modelling techniques which yield confidence intervals with a more appropriate coverage. In this case, the regression procedure targets the conditional quantiles of the wind power generation distribution, always using weather variable as inputs. A further advantage of this framework is that all sources of variability are naturally accounted for, weather-driven or not.  We perform our model benchmarking by measuring the empirical coverage rates of forecast confidence intervals coming from the two approaches, supporting our claim for which quantile regression methods are necessary to reach an accurate representation of wind power generation variability.