Temporal Post-Processing of Wind Forecast using Ensemble of Machine Learning Methods

Diandra Akemi Alves Kubo¹; Rafael Toshio Inouye²; Cesar Augustus Assis Beneti²; Alana Renata RIbeiro¹

¹ Universidade Federal do Paraná; ² Sistema Meteorológico do Paraná

doi:10.20906/CPS/CILAMCE2017-0871

Resumo

Wind forecast accuracy is of extreme importance to several applications, ranging from wind energy generation monitoring to the aviation decision making process. The Weather Research and Forecasting Model (WRF) provides a 72-hour forecast for every thirty minutes of a large number of variables, including wind components. Although this model performs fairly well on a mesoscale, it can be biased and too general for specific areas, and therefore the post-processing of this kind of model is a well-established task to improve forecast accuracy. A comparison of which wind variables to use for correction, U and V components versus wind speed and direction is made in this work. For each variable, several tests were made using five different models plus the average bias correction, chosen for their known use for regression in the literature. They were: Support Vector Regression, Random Forest, Multi Layer Perceptron, Linear Regression and M5 tree. Every model was trained using the forecast and the observed values for each variable alone at first, then with a feature that represented the forecast distance from the initial forecast, and then with another feature of the unbiased forecast, calculated with the average bias for each forecast time on the training data. The second set of tests consisted in investigating whether a model trained on only one wind component was better than one trained on two, e.g. if the U wind component is better corrected when the V component data is also used on the model training. Finally, a performance assessment was calculated by training models for each variable using features from the other three variables, e.g., U wind component correction using the V component, speed and direction of wind to train the models. All models were trained with WRF data and observed values from five stations on the northeast region of Brazil from 30 November 2016 to 28 February 2017 with 33% for validation and the remaining for training. Separating the variables in two groups, group A with the U and V componen

Palavras-chave: Forecast Post-Processing; Machine Learning; Wind Forecast