To address the capacity allocation issue in frequency modulation auxiliary services provided by shared energy storage for wind farm clusters, a dynamic energy storage capacity allocation method integrating power generation prediction and optimal scheduling is proposed. First, a long short-term memory neural network (LSTM) is used to achieve high-precision prediction of the real-time power generation of wind farm, and then constructs a capacity optimization model with the goal of maximizing the overall revenue of shared energy storage and wind farm clusters. A dynamic balance constraint on the state of charge (SOC) of shared energy storage in wind farm clusters is introduced into the model, and a particle swarm optimization (PSO) algorithm is adopted to globally optimize the capacity allocation scheme, realizing the on-demand dynamic allocation of shared energy storage resources among wind farm clusters. Typical case studies verify that the method has significant advantages in improving the utilization efficiency of energy storage systems and maintaining the stability of the SOC of shared energy storage. Compared with the traditional equal distribution strategy, the proposed strategy can significantly enhance the response capability and economic benefits of wind farm cluster in frequency modulation services.