The power curve can reflect the overall generation performance of wind turbines. To avoid the difficulty in input feature selection in power curve modeling, based on power extraction of air streams, the blade tip speed ratio and pitch angle are taken as the input variables of wind energy utilization coefficient modeling. First, by analyzing the characteristic curve of a wind turbine, its working state is divided into three phases: constant power, constant speed, and maximum power point tracking phases. Then, aiming at the limitations of neural network in training time, hyperparameter selection, and model interpretation, a multivariate polynomial segmented power prediction method is proposed. Based on the supervisory control and data acquisition (SCADA) data of a 2 MW wind turbine, the expressions of the wind energy utilization coefficient in three phases are given. In the constant power phase and constant speed phase, the wind energy utilization coefficient is quadratic and quintic polynomial about the blade tip speed ratio and pitch angle, respectively. In the maximum power point tracking phase, the wind energy utilization coefficient is a cubic polynomial of the blade tip speed ratio. The results show that, in the four methods (multivariate polynomial regression and neural network power curve modeling with and without segmentation), the segmented polynomial regression method can not only improve the interpretation ability of the model, but also has high accuracy, and the mean absolute percentage error is 6.29%.