Automatic tuning of Kalman filters by maximum likelihood methods for wind energy forecasting

Wind energy has the advantages of being clean, having a zero-cost primary energy source (wind) and having low operating and maintenance costs. Despite these advantages, it is difficult to manage due to its variable condition. Recently, there has been an explosion of forecasting tools to integrate wind energy into the electrical grid. This paper is devoted to improve the performance of statistical tools based on Kalman filter models. We substitute the traditional way of setting the values of the model parameters by estimating them by quasi maximum likelihood methods for a certain forecast horizon. It produces an automatic self-tuning of the model parameters for each particular wind farm. We show that this brings the models close to an optimum for all the horizons. We also propose new multivariate models to capture the effect of missing inputs on the predicted power. We have applied our methodology in several wind farms and the results show that these two approaches always provide more accurate predictions, with up to 60% of improvement for the RMSE. Finally, we propose a real-time estimation strategy.