Back to the programme printer.gif Print

Delegates are invited to meet and discuss with the poster presenters in this topic directly after the session 'Forecasting: Maximizing grid deliverability and leading your business processes to profitability' taking place on Thursday, 13 March 2014 at 11:15-12:45. The meet-the-authors will take place in the poster area.

Jose Dorronsoro Universidad Autonoma de Madrid, Spain
Jose Dorronsoro (1) F P Alberto Torres (1) Jesus Prada (1)
(1) Universidad Autonoma de Madrid, Madrid, Spain

Printer friendly version: printer.gif Print


Nowcasting meteorological readings for wind energy production


Meteorology based wind energy forecasting is usually done at a short-term (up to 6-12 hours) and medium term (up to 72-96 hours) horizons. NWP-based predictions are usually the basis for the medium term but their update frequency (6 hours) and delay of first forecast imply the need of other input sources for short term forecasts, such as real time meteorology measures. In Spain, AEMET maintains a network of 240 stations that consistently record wind speed and direction and other 6 variables with a 1-hour delay. We will consider the use of such readings to nowcast wind energy production.


In this work we will consider such nowcasting for the aggregated wind energy production of peninsular Spain and that of the Sotavento wind farm in Northwestern Spain. We use as references a day-ahead model DM that predicts wind energy from the ECMWF’s NWP forecasts and an hourly refresh model HM of the DM forecasts that at each hour uses past energy readings to update the coming hour prediction. Our proposal relies on a Lasso model, LM, that uses previous readings from the AEMET network to nowcast wind energy in the following hours.

Main body of abstract

Lasso is a linear model built minimizing the standard square error plus a L1 regularization term with the sum of the absolute weight values. Besides correcting the possible overfitting of a standard least squares linear model, the L1 regularization forces model sparsity, limiting the number of nonzero linear weights. This sparsity allows to detect which weather stations yield non zero linear weights and are, thus, most relevant, something that can be exploited for either a more precise station relevance analysis or to select a reduced number of important variables that could be used to feed a more powerful nonlinear model such as neural networks or kernel based support vector regression. This sparsity can also be exploited in the Sotavento case to avoid a costlier station relevance analysis.
Our preliminary results show the nowcasting LM model yielding moderately but clearly better results than HM in the first horizons for aggregated wind energy prediction. For latter horizons HM and LM models tend to revert to the day-ahead predictions of model DM as the correlations against future energy productions of current meteorology (or energy production measures) decrease. On the other hand, when we apply the same approach to energy forecasts at the Sotavento wind farm, the advantage of the nowcasting LM model over the alternative hourly HM model is smaller. We point out that we apply the LM model in a straightforward way without taking into consideration important issues such as relative positions of AEMET’s station and wind farms.


We show the feasibility of nowcasting wind energy, which appears more effective for large areas than for individual farms. In a large area there seems to be a space-time trade off, with the nowcasting model compensating spatially the correlation time decay between current meteorology and future wind energy production. This time-space compensation seems to be smaller for individual wind farms.
Lasso’s sparsity enforcing can also be used to analyze this, as it progressively “turns on” the most relevant variables. Once detected, their time evolution may further reveal the connection between current meteorology readings and future wind energy production.

Learning objectives
We illustrate a way of nowcasting wind energy that relies on readings of Spain’s weather station and that can be applied elsewhere.
The main tool is Lasso, a widely used modern sparsity enforcing linear model, particularly useful in large dimensionality problems such as wide area nowcasting.
Another advantage is Lasso’s ability to detect the most relevant variables, something useful to analyze nowcasting results or as a first step towards more powerful non linear models.