Wind Power and Wind Speed Forecasting in (semi-)Complex Terrain Using an Artificial Neural Network (ZiANN)

With the increasing amount of renewable energy production reliable forecasts of the possible intra-day and day-ahead power production are needed. Especially for balancing the load on the electricity network and trading of wind power useful predictions are important. In regions with semi-complex and complex terrain such as the Vienna Basin or (more or less) gentle mountain ridges as the Moschkogel in Styria predictions with good skill scores are challenging. The NWP system and even high resolution nowcasting systems are often not tailored enough for wind farm site. Furthermore, even with a temporal model run of hourly to 3-hourly and an hourly output frequency these predictions are not available early enough for the intra-day predicitions.

Therefore, a non-physical approach using an artificial neural network (ANN) for operational purposes of forecasting wind speed (at observation and wind farm sites) and wind power (at wind farm sites) has been investigated. ANNs are able to digest the information of recent and historical information of observations and NWP forecasts and combine them with any other useful (e.g. topographical indices) or maybe not-so-useful (e.g. trafic jam in Vienna) information. Similarily to statistical methods such as linear regression or ARIMA models they can be used as additional forecasting step in tailored post-processing NWP forecast for specific sites. ANNs such as a convolutional NN (CNN) are able to use gridded input and output fields. Their advantage is, if trained properly and using a large dataset, the computational effort is rather cheap and forecasts are available in just seconds. One of their biggest disadvantage is that their black box-like behavior, thus, how and what the network uses and weights.

A feed forward ANN was developed and modified to our needs to forecast wind power and speed at different Austrian wind power sites, at different hub heights as well as at meteorological observation sites. Different input fields, NWP forecasts of deterministic and probabilisitc models and time-lagged input, observations of meteorlogocial parameters, different "age-types" of observations and topographical indices were investigated using data mining methods.

Different sites were considered for a test period, including high alpine sites such as the Windpark Niedere Tauern as well as low-land sites as the wind farm sites at the Parndorfer Platte, for statistical reference. Different nacelle heights were considered, too. Therefore, different kinds of interpolation techniques of NWP forecasts to nacelle heights were investigated. Also using 10 m NWP wind fields as input to forecast at nacelle height was evaluated.

Also, station/wind farm clustering, different lengths of training periods, etc. were investigated. For the operational version of the ZiANN, a set of different deterministic NWP models, such as the ALARO, the AROME and the ECMWF model, are used. The ZiANN performance is evaluated using different statistical scores and cross-validation. Results show that the developed interval-based ensemble approach, now operational, ZiANN outperforms the available NWP model forecasts, a simple regression model and a MOS-based multiple NWP forecast.