Deterministic and Probabilistic Wind Power Forecasts by Considering Various Atmospheric Models and Feature Engineering Approaches

This work proposed a model and the procedures of deterministic and probabilistic forecasts, e.g., hour-ahead and day-ahead, for wind power generation. The contents of this research include numerical weather prediction, data pre-processing technique, and forecasting models using artificial intelligence methods. Regarding the inputs of the model, we had considered three kinds of NWP wind speeds, generated by the Central Weather Bureau based on three atmospheric models, namely WRFD, RWRF and WEPS, and historical wind power generation. The measured wind speeds, out of an anemometer tower, were used to compare with the NWP wind speeds to help us select the least error time combination. Regarding data pre-processing, NWP wind-speed correction based on the height of wind turbines and PCA and EMD for exacting wind-speed feature had been tested. As for the forecast model, we used artificial neural network and XGBoost to predict the generation of wind power, and a number of error indexes had been used to evaluate the performance of the forecasts. The empirical data from a wind farm in Taiwan verifies the accuracy of the proposed method. What worth mentioning, the importance of model selection, numerical weather prediction, and data pre-processing is self-evident.