Evaluation of Two ANN Approaches for the Wind Power Forecast in a Mountainous Site

Abstract

Accurate wind power forecast is very important in order to construct smart electric grids. Nevertheless, this task still constitutes a challenge because wind is a very variable and local phenomenon. It is difficult to downscale information coming from Numerical Weather Prediction (NWP) models down to wind farm level and this is especially true onshore, in complex terrain conditions. Artificial Intelligence often comes at hand, for its power in learning what is hidden inside data: Artificial Neural Networks (ANN) are therefore commonly employed for wind power forecast. In this work, a pure ANN method is compared against a hybrid method, based on the combination of ANN and a numerical method based on physically-consistent assumptions (Computational Fluid Dynamics). Both approaches are validated against the SCADA data of a wind farm sited in Italy in a very complex terrain. It arises that the two methods have overall similar performances on average. However, pure ANN turns out to forecast better at mid-energy levels and during cut-off events at the highest wind speed, whereas the hybrid method forecasts better during low and high wind speed ranges. This makes the two approaches complementary and promising for future applications through an ensemble strategy.