Evaluating the health status of wind turbines and reasonably allocating their output power is of great significance in reducing the operational and maintenance costs of wind farms. First, an output power prediction model of wind turbines based on a Long Short-Term Memory (LSTM) network is trained using the historical operational data in the Supervisory Control and Data Acquisition (SCADA) system. Then, according to the deviation amplitude between measured output power of wind turbines and its predicted output power, the health status of wind turbines can be divided into three categories: good, average, and bad. Finally, for the health status of each wind turbine, the maximum generating capacity of each wind turbine and power generation directives from power grid dispatching departments to wind farms are established, along with the objective function and constraint formula. A genetic algorithm is used to solve the problem. Thus the output power allocated to each wind turbine is obtained. The simulation results show that the method proposed can not only distribute the power reasonably according to the health status of the wind turbine, but also meet the total power of the wind farm given by the dispatching center. This work is supported by Shaanxi Provincial Key Research and Development Plan of China (No. 2017GY-061).