Understanding the Contribution of Varying-coefficient Models to Wind Power Forecasting

Wind power forecasting requires advance non-linear time series models. Varying-coefficient models generalise linear autorregresive models by introducing a dependency between the captured dynamics and a conditioning variable, providing non-linear modelling while preserving model simplicity. The understanding of this dependency is key for selecting appropriate conditioning variables. To this end, we introduce the so-called beta-coefficients, which are specifically defined to provide interpretability on how the model integrates different features of the wind power dynamics into the forecast, such as the skill of the model during fast power changes (ramp events and fluctuations). Experimental results were obtained for a multimegawatt wind farm located in the north of Spain. This allowed us to discuss the introduced notions for the particular case of considering the expected wind speed as conditioning variable.