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.

Pål Preede Revheim University of Agder, Norway
Pål Preede Revheim (1) F P Hans Georg Beyer (1)
(1) University of Agder, Grimstad, Norway

Printer friendly version: printer.gif Print


Using random forests for wind power ramp forecasting


The increasing use of wind power as an energy source poses new challenges for the management of electrical grids. One of the major challenges is dealing with sudden large changes in wind power production, normally referred to as ramp events. The sooner ramp events can be predicted, the smoother and more efficiently they can be dealt with. This calls for efficient ways of forecasting ramps.


This paper investigates the possibilities for classifying and predicting ramp events using random forests built on historic on-site and off-site data, a technique that has previously been used for e.g. short term wind power forecasting. Random forests are an ensemble learning method for classification that operate by constructing a large collection of de-correlated decision trees, and then averages them to provide stable outputs. Ramps are seen as deterministic incidents that are either occurring or not occurring, and the random forests are used to classify and predict situations into the three groups “up-ramp” “down-ramp” or “no ramp”.

Main body of abstract

The ramp forecasts are performed in three stages. First the ramp events are identified, then the dataset used for the random forest is constructed, and finally the forecast models are fitted. As ramps in wind power production are of interest, the wind measurements are transformed into power production by the use of a logarithmic height transformation and a generic power-curve. The ramps are identified by the definition |P(t+Δt)−P(t)|>P_lim.

For the site of interest it's assumed that the forecast for the next hour together with the measurements from a number of previous hours contain information about the probability of a ramp. For the off-site data it's assumed that there's a high probability that a ramp event occurred at a downwind site at an earlier time, hence that the ramps are subject to spatial propagation from downwind sites to upwind sites. Similar to the site of interest this is included through information about the forecast for the next hour together with measurements from previous hours. From earlier experience the cross- and autocorrelation of wind power time series tends to decline quickly after the first few hours, and the number of previous hours included is therefore limited to three hours. The performance of the random forest classification is compared to a NWP-based and a linear discrimination (LDA) approach. The ramp forecast methods results are summarized in contingency tables as well as in two evaluation metrics (Hanssen & Kuipers skill score (KSS) and odds ratio skill score (ORSS)).


The ramp forecasts of the random forests and the NWP-based approach show a comparable pattern, but the random forest forecasts were consistently more accurate, especially at predicting up-ramps. The LDA-based approach obtained by far the highest number of correctly forecasted ramps, but this however suffered from a very high number of false positives (~10% of the total observations). Evaluation of the forecast methods through evaluation metrics were found difficult because of the high share of non-events.

Learning objectives
The challenges that were encountered when evaluating the methods by evaluation metrics may be addressed by introducing costs related to the different forecast outcomes. This way it will for instance be possible to give different importance to false positives and false negatives, and optimize the forecasts to an operational situation.