© IMechE 2020. A probabilistic method for modelling empirical site-specific wind turbine power curves is proposed in this paper. The method is based on the Gaussian mixture model machine learning algorithm. Unlike standard wind turbine power curve models, it has a user-selectable number (N) and type of input features. The user can thus develop and test models with a combination of measured, derived or predicted input features relevant to wind turbine power-output performance. The proposed modelling approach is independent of the site location where the measurable input features (i.e. wind speed, wind direction, air density) are collected. However, the specific models are location and turbine dependent. An N-feature wind turbine power curve model developed with the proposed method allows us to accurately estimate or forecast the power output of a wind turbine for site-specific field conditions. All model parameters are selected using a k-fold cross-validation method. In this study, five models with different numbers and types of input features are tested for two different wind farms located in Ireland. The power forecast accuracy of the proposed models is compared against each other and with two benchmarks, parametric wind turbine power curve models. The most accurate models for each of the sites are identified.