Large power ramps are a phenomenon which affect wind power plants (WPPs) across the world. They put strain on electricity networks and can be hard to predict. Some ramps can be linked to specific phases of the WPP’s operation such as cut-in and cut-out, and some can be correlated with specific weather regimes. There is a financial incentive to produce an accurate very short-term forecast for WPPs, however the time series models used on these timescales often perform poorly during periods of high variability. The objective here is to use external information to classify power time series into regimes which can be incorporated into statistical models to boost their overall performance. A case study is made of the Horns Rev 1 offshore wind farm in Denmark, where ramps of over 5MW/min are reported, and a very short-term power forecast is required by the system operator. The external information is in the form of Numerical Weather Prediction (NWP) variables from European Centre for Medium-Range Weather Forecasting (ECMWF) as well as the time series derived behaviour statistic standard deviation of power (SDP).
Ramps correlated to those NWP variables which pertain to the stability of the atmosphere have been investigated, as large ramps at Horns Rev 1 are suspected to be more common in the presence of convection. Autoregressive models are used to predict 10-minute mean power 10-minutes ahead. A simple autoregressive model of order 3 is used as the benchmark. Regimes are defined by threshold values of the external information and multiple autoregressive models of the same order as the benchmark are trained and tested on their respective regimes. K fold cross validation was employed to identify threshold values of a range of NWP variables and SDP to minimise forecast error in terms of mean absolute error (MAE).
The performance of the autoregressive model is shown to be improved by regime switching in the Horns Rev 1 case study. Using SDP an improvement of 1.5% in MAE was possible. For the NWP variables, an improvement of 1% in MAE was possible using a combination of boundary layer height and forecast wind speed, focusing on the cut-in region of wind turbine operation. These promising preliminary results are leading us to investigate the use of this external information further.
It is reasonable to conclude that a basic regime switching model using information already available to WPP energy dispatchers can improve very short-term power forecasts. Variation of power output in time from a WPP has different characteristics at different parts of the power curve which can be learned by training on same-type data sets. Finally, the variable nature of the power output of the WPP in the test case has a persistent quality which can be exploited to good effect in choosing the appropriate model for the following prediction.