The power capture capability of wind farms is often constrained by various factors. To maximize the power output of wind farms and address the wake effects and random wind speeds, this paper proposes a control scheme for wind farms based on deep reinforcement learning, integrating both model-based and model-free methods within a TD3 network framed by Actor-Critic architecture. This study improves the Jensen wake model by enhancing its accuracy through the consideration of time delays. Delay sensitivity is introduced as a factor in deep reinforcement learning, allowing for the optimization of control strategies while exploring the long-term impacts of upstream turbine changes on downstream turbines, thus adjusting the current control scheme. A simulation model is established, significantly improving the power capture efficiency of the wind farm and achieving maximization of its power output.