Improving Immediate Wind Forecast and Wind Power Generation

Wind forecasting is difficult. Increasing the accuracy of short-term wind forecasts is necessary to maximize the operation of wind farms, their production, and promote their development. The aim of this project is to improve existing techniques for predicting wind speed and the power produced by wind turbines. After an in-depth analysis of wind forecasting techniques, the Kalman filter emerged as a promising method. It is an adaptive method for correcting errors associated with physical forecasting models and can be applied immediately without requiring long training times on historical datasets. While most recent methods rely solely on wind speed forecasts, the addition of other variables such as wind direction could improve model accuracy. Although some approaches, such as artificial neural networks, have been shown to reduce the mean error, the inclusion of directionality in Kalman filter models applied to wind speed or power has not been studied. This paper proposes a new method that models Kalman filters based on differences between raw forecasts from numerical weather prediction models and non-linearly modelled observations of wind speed and direction. For energy forecasting, two methods were developed for cases where wind speed observation data is unavailable. Firstly, the Kalman filter is applied directly to power to estimate bias. Secondly, a technique is introduced where observed power is converted to assumed wind speed using an inverse power curve, enabling filters to function based on wind speed. Historical observation and forecast data from 20 sites in Europe and North America were used to analyze model performance. Several indicators were calculated to quantify the accuracy of different models, including bias, the mean absolute error (MAE), and the root mean square error (RMSE). A first-order polynomial yielded the best results for flow-optimized Kalman filters, while higher-order polynomials applied to power reduced the mean error. Adding directionality improved forecast accuracy across all methods studied.