This paper introduces a machine learning (ML) based model predictive control (MPC) with piecewise-affine approximation (PWA) structure for maximizing wind energy capture for an individual wind turbine operating in wind farms with low-quality wind resources. While MPC has the capability to systematically consider the stochasticity of wind speed and the dynamic process of wind turbine, its real-time implementation in a hardware controller of wind turbine has not been successful due to its high online computational burden and stringent execution time requirement in practice. To address this long-standing issue, this paper proposes a two-phase ML-based method consisting of linear regression and clustering to construct a PWA of the optimal law for original MPC scheme. The two-phase ML-based method is tunable with computational complexity, which can be adjusted to meet the hardware limitation of the given controller of wind turbine to enable real-time implementation, while preserving the optimality of linearized full-fidelity MPC as much as possible. We conduct simulations and experiments to demonstrate the effectiveness of the two-phase ML-based method.