This paper proposes probabilistic wind turbine power curve (WTPC) models to quantify the uncertainties of energy conversion and highly scattered relationships of actual wind speed to power. First, new model inputs (i.e. pitch angle and wind direction) and novel data clearing methods are presented to improve the model accuracy, which is rare in the previous studies. Second, the models are established based on three nonparametric algorithms, i.e. Monte Carlo, neural network, and fuzzy clustering. Third, to fill the research gap on model evaluation, the desirable properties of a probabilistic WTPC model are defined as expected variance ratio (EVR), and this index is formulated by calculating the cumulative gaps between the simulated and actual power distribution in each wind speed segment. Data from two Chinese wind farms are used to validate and compare the proposed methods using the mainstream deterministic index and the proposed EVR. Results show that (i) new model inputs and data clearing methods are able to improve the accuracy for probabilistic models regardless of the afterwards modelling method; (ii) fuzzy outperforms other probabilistic models.