This poster was presented at the online WindEurope Wind Resource Assessment Workshop in June 2019. The accurate prediction of the expected power production of a wind turbine at a particular site is important in both the wind farm planning and operation phases, including for estimating the expected Annual Energy Production (AEP), examining anomalous behaviour of single wind turbines as well as making short-term power predictions for optimising revenues. However, the power curve provided by the manufacturer is not specific to the atmospheric conditions at the site, and therefore only applies if the conditions are the same as those at the test site. Previous studies have shown that atmospheric conditions can affect the turbine output by as much as 10%. The power curve is therefore not always suitable for accurate site-specific power predictions. Recently, the suitability of applying machine learning for increasing the accuracy of site-specific power predictions of multi-megawatt wind turbines compared to the standard power curve method has been investigated by examining a large amount of simulation data (abstract submitted to TORQUE2020). It was found that regression trees can increase the accuracy of site-specific power predictions by a factor of three. Utilisation of simulation data ensured a high quality and full range of wind speed, turbulence intensity and shear input data, which is required to train and test the model. In this present work, the method has been extended for application to real measurement data, based on the Python DecisionTreeRegressor from the module scikit-learn. For this, publicly-available datasets from a micrometeorological experiment conducted in two distinct operating wind farms in a coastal area of the northeast region of Brazil, called Pedra do Sal Wind Farm (UEPS) and Beberibe Wind Farm (UEBB), were used (https://zenodo.org/record/1475197#.XeoLOdVCdhF). These wind farms are located on the northeast coast of Brazil where meteorological conditions are strongly influenced by trade winds and sea breeze. Both datasets represent a full year of measurements from August 2013 to July 2014. In this work, data was used from a fully instrumented IEC-compliant 100 m met mast at each site, with five levels of first-class calibrated cup anemometers and one level (100 m) with 3D sonic anemometer. The measurement data was used to construct 10-minute average shear factors and turbulence intensities, as well as wind speeds. Additionally, power measurements were taken from the 10 minute SCADA data from the operating wind turbines closest to the met masts. The results showed that the regression tree method works well for real measurement data, and that the power prediction accuracy could be increased by up to a factor of three. It could be applied by manufacturers to train a model for the dependency of power production on atmospheric conditions for a given wind turbine type, and then applied by the planner or operator at any given site for which measurements of atmospheric conditions are available. Further investigations on the ability of the method to predict wake effects are ongoing.