Full-real-life Wind Power Prediction based on Gradient Boosting Decision Tree

Wind energy is highly random, volatile and uncontrollable. Therefore, in order to ensure the reliable and efficient operation of wind farm power generation connected to the grid in a full-real-life environment, accurate wind power prediction based on weather forecasts is crucial. This paper uses a learning algorithm based on a gradient boosting decision tree (XGBoost algorithm and LightGBM algorithm) to develop two models to study the power prediction effect in a full-real-life environment. Both models are trained using ECMWF weather forecast data and 19 months of wind farm power generation. Preliminary experiments show that the XGBoost algorithm is slightly better than the LightGBM algorithm in terms of accuracy, but the LightGBM algorithm is much shorter than the XGBoost algorithm in terms of training time. Since data is updated frequently in a full-real- life scenario and the model training time requirement is short, a model based on the LightGBM algorithm is selected to verify the power prediction effect of a wind farm in a full- real-life environment. In a full-real-life environment, the ECMWF weather forecast is input into the LightGBM model to perform ultra-short-term, short-term and medium-term forecasts on wind farm power. The full-real-life evaluation standard adopts the industry grid-connected rules. The simulation results of the full-real-life environment confirm that the learning algorithm based on the gradient boosting decision tree can not only be used for ultra-short-term, shortterm and medium-term wind farm power forecasting, but also has excellent prediction results in the ultra-short-term and short-term. At the same time, more accurate weather forecasts are more important for power forecasting.